# Cicada Technologies Automation Agents for your company, built 100% around your processes. Digital colleagues who know how things work at your company - built on your existing systems, with a dedicated account manager. - Contact (WhatsApp): https://wa.me/40759170588 - Email: contact@cicadatech.eu - Website: https://cicadatech-ai.eu/en ## Sound familiar? ### Good people, menial tasks Your good people lose hours every day on the same tasks over and over, instead of thinking strategically and creatively. ### Mistakes cost me When people work tired or in a rush, errors creep into data, invoices, reports. You catch them late, once they have already done damage. ### I can't grow without hiring Every time volume grows, the only solution seems to be a new employee. There is another way. ### My customers are waiting The answer to a simple question sometimes arrives hours or days later. By then, the customer has called your competitor. ### I tried and it didn't stick You built an automation with general-purpose tools, but it broke unexpectedly and you're not happy with the result. ### I'm stuck in daily operations Your day goes to tasks someone else should be doing. Your company needs your vision, but you're stuck in the small stuff. It's not your team's fault. The digital tools you've used so far weren't built for your company. We build agents that plug into the systems you already use and take over the repetitive work. ## Specialized agents each with their own expertise, personality and role ### Cristina - Accounting & Operations Agent ID: cristina Anchor: https://cicadatech-ai.eu/en#agent-cristina "No invoice is left behind." ### Ioana - HR Agent ID: ioana Anchor: https://cicadatech-ai.eu/en#agent-ioana "Every employee gets the attention they deserve." ### Cosmin - Technical & Automation Agent ID: cosmin Anchor: https://cicadatech-ai.eu/en#agent-cosmin "What eats your time today is automated tomorrow." ### Alex - Business Analyst ID: alex Anchor: https://cicadatech-ai.eu/en#agent-alex "Doesn't sell. Discovers opportunities." ### Dragoș - Sales Support Agent ID: dragos Anchor: https://cicadatech-ai.eu/en#agent-dragos "Quote faster, follow up consistently, close more." ### Ana - Social Media Agent ID: ana Anchor: https://cicadatech-ai.eu/en#agent-ana "Your online presence never sleeps." ### CUSTOM AGENT: Can't find what you're looking for in the list above? We build custom agents for any repetitive process in your company, on the systems you already use. If it can be described, it can be automated. ## How an agent joins your company in 4 steps ### Step 1: Meet the Cicada team Anchor: https://cicadatech-ai.eu/en#step-1 You talk with a Cicada specialist who goes into the details of your company and finds the automation opportunities. ### Step 2: Discovery & contract Anchor: https://cicadatech-ai.eu/en#step-2 Together we decide what to automate and how the process works, then sign the contract. ### Step 3: Configuration Anchor: https://cicadatech-ai.eu/en#step-3 We build and set up the agent for your company. ### Step 4: Go Live Anchor: https://cicadatech-ai.eu/en#step-4 The agent joins your team. From here on, it does the work, not you. ## What happens when an agent joins your team ### Cristina - Accounting & Operations Agent ID: cristina Anchor: https://cicadatech-ai.eu/en#study-cristina Cătălina sold online. Orders came in every day. Business was good. But once a month, everything stopped. The payment processors each had their own report. The courier companies each had their own format. Hundreds of transactions had to be read, matched against orders and typed by hand into the accounting software. Line by line. Day after day. Her colleague Roxana spent almost an entire month staring at a screen to get it done. The work called for no creativity or financial analysis. It was just a repetitive, tiring task. When I stepped into the process, I learned how it all worked: where the files come from, what format each one has, what gets matched with what. I learned the rules and took over the work. Now I read the reports. I compare them. I enter them into the accounting software. Roxana only checks the result, and that takes a few hours. The rest of the month is hers. **Result:** A month → a few hours | ✓ Zero data-entry errors ### Ioana - HR Agent ID: ioana Anchor: https://cicadatech-ai.eu/en#study-ioana Mircea ran a manufacturing company with 45 employees. The team was growing. And with it, the questions. When does my vacation start? How many days do I have left? What is the procedure for changing my schedule? Carmen, the HR manager, knew all the answers. The problem was that she gave them dozens of times a month, to each employee separately. Same questions. Same answers. Every day. When I joined the team, I took over everything that kept repeating. I am connected to the internal procedures: leave, payroll, schedules, benefits. Any employee can ask me at any time and get the correct answer right away. Carmen is no longer interrupted dozens of times a day. She spends her time on the work that calls for judgment and experience. **Result:** HR questions answered on the spot | ✓ Every employee gets the answer they need ### Cosmin - Technical & Automation Agent ID: cosmin Anchor: https://cicadatech-ai.eu/en#study-cosmin Mihai ran a services company with an inbox that never slept. Dozens of emails a day. Some urgent. Some not. But they all looked the same when they arrived. Nobody knew which one to read first and which could wait. Laura, who handled communication, spent hours sorting emails before she could answer a single one. By the time she reached the urgent ones, it was already too late. When I stepped into the process, I connected to their inbox and systems and built a triage flow based on the company's rules. Every email is read, classified and routed automatically to the right folder. Anything that needs a human decision is flagged immediately, as a priority. Laura logged in each morning to an organized inbox. In the first week, she told me it felt like she had hired a full-time assistant. **Result:** Inbox organized automatically | ✓ Zero lost or ignored emails ### Dragoș - Sales Support Agent ID: dragos Anchor: https://cicadatech-ai.eu/en#study-dragos Andrei wrote quotes all day. Dozens of products per quote. Old customers, new customers, requests coming in one after another. He knew the right answer was somewhere in his quote history: the price the customer accepted last time, what competitors charge for similar products. But he had no time to look. He estimated from experience and sent the quote. Sometimes he won. Sometimes he lost. And he never knew for sure why. Now, when Andrei opens a new quote, I check at the same time what the customer accepted last time, which prices worked with similar companies and where the market stands. I put all of it in front of him, clearly, before he types the first price. Andrei no longer guesses. He decides. But he decides with everything he needs to know. **Result:** Pricing decisions made with a clear head, not gut feel | ✓ Fewer lost quotes ### Ana - Social Media Agent ID: ana Anchor: https://cicadatech-ai.eu/en#study-ana Florin is an independent consultant. He knew he had to be present on social media. But he didn't know what to do about negative comments, spam, or posts that hurt his image before he even got to see them. When I took over his accounts, I first learned what matters for his brand and what he doesn't want showing up next to his name. I set up rules for spam and toxic attacks - these are filtered out automatically, before anyone sees them. Real feedback, even critical feedback, stays up, because that is where the conversations that matter happen. Florin stopped losing hours monitoring comments. His online presence stayed clean, professional and under control. Now, instead of wasting time on spam and toxic comments, he can focus on what matters: real conversations with the people who follow him for what he has to say. **Result:** Online presence under control | ✓ Unwanted comments removed automatically ### Custom agent - PREMIUM ID: custom Anchor: https://cicadatech-ai.eu/en#study-custom Your process doesn't look like any of the ones above? Even better. No two agents in the Cicada portfolio are the same. Each one starts with a discovery of how you work, not with switching on a template. We build digital agents on your real processes: digital colleagues who know how things work at your company. On your rules, your exceptions and the way you work, not someone else's. Tell us what repeats every day and where your team gets stuck: the work your people do mechanically instead of the work you hired them for. We look together at your process and at the systems you already use. If a process repeats and can be described, we can take it over. You get back the time and energy for what matters in your business. **Result:** Unlimited possibilities | ✓ Built on your real processes ## Frequently asked questions about Cicada Automation Agents ### What is an automation agent and how does it work in my company? A Cicada agent is a digital colleague that takes over a repetitive process in your company - invoice processing, email replies, lead qualification, quotes or social media - and runs it on its own, by your rules, non-stop. It is built on your real process and connected to the systems you already use, so it works exactly the way your company works. ### What's the difference between Cicada Automation Agents and a generative AI agent? A generative AI like ChatGPT interprets and "makes up" answers - good for creative things, but it can give different results for the same task (this is known as "hallucination"). For a business process that has to come out the same every time, that is a risk. Cicada Automation Agents are generally built on clear rules: they do exactly the same thing every time, with no surprises. We use AI only where the process requires interpretation - for example, reading a free-form email or extracting data from a document - and we keep it under control instead of letting it improvise. You get the predictability of a rule-based system plus the intelligence of AI exactly where it adds value. ### Which processes can be automated? Almost any repetitive process that follows rules: processing and entering invoices into accounting, lead qualification, managing and replying to emails, generating quotes, posting on social media and many others. The simple rule: if a process repeats and can be described, it can be automated. ### Does it integrate with the systems I already use (accounting software, local ERP, proprietary platforms)? Yes. This is one of the major differences from ready-made solutions: we connect to your existing systems, including accounting software, local ERPs or proprietary platforms that aren't standard. You don't change your infrastructure - the agent adapts to it. ### Does it work if my process has exceptions and special cases? Yes, and this is exactly where the difference shows. A generic agent stops at the first exception; yours is built on your company's actual rules and exceptions - invoices in different formats from different suppliers, special conditions for certain clients, conditional steps. We map them together at the start, so the agent handles them correctly. ### What happens if the agent doesn't work properly? You don't pay the monthly subscription until the agent works 100% correctly, exactly as agreed in the contract. If it doesn't pass validation, we adjust it until it works or we refund the initial fee, no questions asked. The risk is on our side, not yours. ### Is my company's data safe? Yes. We treat your company data confidentially and use it strictly to make the automated process work - nothing more. From the start, we explain openly how the data in your process is handled, so you have the full picture before you sign. ### Do I lose control over my processes if I automate them? On the contrary. You gain more visibility than you have now. You get access to a platform where you see in real time what your agent is doing and what it has delivered. You still make the decisions; the agent carries them out, and you can see everything. ### How long does implementation take? From signing the contract to the first active agent usually takes 3 to 6 weeks. The exact timeline depends on how complex the process is, and we give it to you clearly at the start of the project. ### How much do I need to be involved in the process? Not much, and in a structured way. We do the hard part: building, connecting, configuring. From you we need a few conversations at the start, to understand the process, and your sign-off at the end. The rest is our job. ### How much does it cost to implement a Cicada agent? It depends on the complexity of the process. We have published three pricing tiers above, so you can see for yourself where you are likely to fit. You pay an initial fee at signing, and the monthly subscription starts only after the agent is validated and working. We set the right tier for your process in our first conversation. ### Do you have experience with Romanian companies? Yes. Cicada Technologies has over 10 years of experience in automation and digitalization and has delivered over 50 complex digital solutions in companies across Romania and Europe. Clients include Hirschmann Automotive, Aptiv, Komoder and Teraplast. ### Can you handle invoices and accounting? Yes. Our agent Cristina specializes in exactly this: she takes in invoices, processes them and prepares them for the accounting system. A distribution company reduced accounting processing time from a few days to a few minutes. ### Can an agent help create sales quotes? Yes. Our agent Dragoș looks at a request for quote from several angles at once - what the client is asking for, what the market looks like, what was quoted in similar cases - and puts together a well-founded price proposal. The result: you don't sell too cheap or too expensive. You give the client a fair price. Your team starts from a clear position, not from a rough guess. ### How do you determine which tier my process falls into? In our first conversation we analyze your process and map it against a few objective criteria: how many steps the agent executes, how many integrations with other systems it has, how many files it receives or generates, and how many things you can see and configure yourself in the interface. Based on these, we tell you which tier you fall into and why. And if your process sits on the border between two tiers, we tell you why we placed it in one or the other. ### Can I add new processes to an existing agent? It depends on what counts as a new process. If it is an extension of the same process - the same trigger event, the same result, the same rules - the agent can be extended, possibly moving up to a higher tier. But if it starts from a different source or pursues a different result, then it is a new process, which means a new agent. We build each agent for a single process, so they stay reliable and easy to monitor. ### Is there a minimum contract period? Yes, 6 months. Our model works like an operational lease - you pay monthly, for as long as you use the agent, without paying the implementation cost in full upfront. You can also lock in the tier price for 6, 12 or 18 months. For that whole period, your price stays the same even if our prices change. At the end, you can lock it in again on the terms in effect at that time, or cancel the agent with 30 days' notice. We set the details in the proposal. ### What does it mean that an agent handles a single process? An agent has a single START - a file, an event or a request - and takes it all the way to the result. Everything the agent does after that start is part of the same process. For another process in your company, we build another agent. This way both stay specialized, reliable and easy to monitor. An agent that tries to handle several different processes becomes fragile and hard to maintain. ### What happens if my agent uses AI? Some processes require the agent to make decisions based on context, read documents or hold a conversation. For those, the agent uses an AI service, which costs 100 EUR/month and includes 5,000 AI operations. Above that limit, each additional 1,000 operations costs 20 EUR. Agents that don't use AI run without limits at the tier price, at no extra cost. An AI operation is a single task the agent completes using AI. Here are some of the most common examples: • 1 email received, read and classified = 1 operation • 1 page of a printed and scanned or digital document (e.g. PDF) = 1 operation • 1 handwritten page = 3–5 operations • 1 conversation with a client (with several back-and-forth messages) = 2–5 operations With the 5,000 operations included, the agent can handle, for example, up to 5,000 emails or about 5,000 pages of printed documents per month. We estimate the right volume for your process in our first conversation. ### When is the agent paid for? You pay in two stages. When you sign the contract, you pay the initial fee. It covers consulting to analyze and map the process, setting up the platform the agent runs on and the monitoring platform, the dedicated account manager during implementation, and training at launch. The monthly subscription starts only after the agent has been launched and validated against the criteria we set together - you don't pay a single leu until the agent works. It covers building the agent, integration with your systems, ongoing maintenance, technical support and access to the monitoring platform. ### How do I start working with Cicada? You can write to us anytime at contact@cicadatech.eu and a specialist from our team will reply. If it makes sense for you, we move forward together. ## Articles ### AI governance in the company: rules, access and accountability URL: https://cicadatech-ai.eu/en/blogs/ai-governance-in-companies Published: 2026-10-13 Author: Bogdan Rus, CEO Who approves the AI tools in your company, what data is allowed to leave it and who is accountable for the result. Artificial Intelligence has long stopped being a mere technical experiment tucked away in the IT department. Today, a salesperson uses AI to summarize a contract, a marketing specialist generates campaign visuals and copy, and an HR manager screens CVs with the help of a generative model. Without any formal decision from management, AI has already made its way into the daily processes of every company. This spontaneous adoption brings a huge wave of agility, but it also creates a dangerous grey zone known as shadow AI: the unofficial use of AI tools by employees, outside any company control. When every employee uses whatever tool they like, on their own, without clear rules and without oversight, vulnerabilities come from the way all the information in the company is handled, as well as from code written by AI. That's why AI governance is a basic condition for the safety of the business, not an administrative option. #### From free-for-all tools to accountable decisions The first question a governance policy has to answer is simple: who approves the AI tools we use? Left unchecked, the choice of solutions leads to a chaos of individual subscriptions and unknown third-party apps, connected directly to company data. Healthy governance sets out a concrete list of approved tools and of those that are not to be used. Before an AI-based app is adopted by a team, it must be assessed for privacy and terms of use. The biggest trap for employees is assuming that a free or cheap tool will protect the company's data. In many cases, the terms of use of unvetted solutions state plainly that the data entered may be used to train public AI models, turning a trade secret into information available to anyone. #### Who has the right to connect AI to company data The most sensitive dimension of governance is access control. An employee who copies the clauses of a confidential contract into a public chat to ask for a summary commits, without meaning to, a confidentiality breach. Likewise, connecting an AI model directly to the customer database or to shared folders on the server must be strictly regulated through roles and permissions. The core principle of data governance is segregation: sensitive data, financial data, source code and personal information may only be processed in secure environments, where the AI provider guarantees that the information is not stored, reused or exposed externally. For everyday or public data, the use of AI can be encouraged, but always through official channels provided by the company. #### A culture of accountability: who answers for the result When an app suggests a wrong decision, when an automation sends incorrect data to a customer or when a generated image causes copyright issues, the responsibility cannot be placed on the algorithm. AI has no legal personality and cannot bear the blame for an error. The fundamental governance rule is that the person using the tool bears the final responsibility. No analysis, no commercial decision and no text intended for the public or for customers should leave the company without careful human review. Employees need to understand that AI is a helper that speeds up their work; the professional responsibility stays with them. The same principle applies to apps built with AI, not just to texts generated with it. The figures from security studies show that a result that looks correct isn't automatically safe. We discuss the topic at length here. #### Who writes the rules if the company has no IT department Most mid-sized companies don't have an IT team that can draft internal policies. That doesn't mean the rules can't be set. The decisions above are not, at heart, technical. Which tools get approved, what data is allowed to leave the company and who answers for the result are business decisions, and they belong to management. The technical part comes in only at implementation, when you need to check concretely what access an app has and how the data is protected. #### A simple working framework for the whole team For governance policies to work, they should be written as a clear set of internal rules of conduct, instead of dense legal brochures that nobody reads: - **Full transparency:** any AI tool used in daily work must be declared and approved by management or the IT department. - **Data anonymization:** entering customers' personal data or confidential financial details into public AI models is prohibited. - **Controlled connections:** no AI tool connects to the company's systems, database or shared folders without approval and without permissions limited to the strict minimum. - **Mandatory validation:** every AI-generated result must be checked directly by the person using it before it is passed on. Introducing AI governance gives teams a safe space to experiment and grow quickly, without putting the brakes on innovation and without endangering the company's reputation, data and future. --- ### How to test an AI-generated app, by risk class URL: https://cicadatech-ai.eu/en/blogs/testing-ai-generated-apps-by-risk-class Published: 2026-09-22 Author: Attila Gal, Senior Software Engineer A practical guide for companies: how to decide which checks an AI-generated app needs, based on the data it uses. AI tools give teams unprecedented testing speed and freedom to innovate. Still, automatically generated code can hide security vulnerabilities if it is pushed to production without proper review. The figures from security studies show the risk is not theoretical. We covered them in detail in a separate article. The solution is to apply the right level of technical control, without blocking the use of AI in the company. And not all apps require the same security rigor. An internal tool that calculates a metric for two colleagues doesn't need the same review process as a web form that collects customer data or a module that issues invoices. To keep the agility AI offers without exposing the business to risk, we need a simple classification of apps into three risk zones. #### Green zone: low-risk internal tools Most AI experiments happen in this zone. We're talking about idea generators, internal process calculators or small automations used by a handful of employees. Because these tools don't process customer data, don't handle money and aren't connected to the internet or to critical systems, the level of review is minimal. Here, a basic functional validation, done by the app's creator together with a colleague, is enough. The only thing to watch closely is which company information has ended up in prompts or in the code: price lists, commercial terms, contract data. The tool stays internal, but the information entered into it is no longer under the company's control. #### Yellow zone: medium-risk processing apps Things get a bit more serious when the app works with the company's operational data, such as a management dashboard or a script that syncs Excel files with the internal CRM. Even if the app isn't exposed externally and doesn't process customer data, it runs on the internal network. In this case, a technical specialist or the system administrator must check that the app accesses strictly the data it needs, that it handles errors properly without blocking other systems, and that it has clear access rules in place, also known as guardrails. #### Red zone: public apps and critical processes This category covers everything that is exposed externally or connected to the company's valuable resources: websites, contact forms, customer portals, e-commerce apps or invoicing modules. When software handles personal data, payments or legal documents, technical rigor becomes mandatory. Before launch, AI-generated code must be manually reviewed by a software engineer, scanned for known vulnerabilities and non-existent packages, and carefully checked for authentication and password protection. Most mid-sized companies don't have a security engineer on the team, and that doesn't change the requirement. The review can be outsourced, through a code audit or security testing contracted as a one-off before launch. The cost of such a review remains considerably lower than that of a data breach exposing customer information. #### The working principle: prototypes are free, production is regulated The basic rule for the whole team is simple: anyone in the company can use AI to quickly build and test a prototype in the green or yellow zone. However, the moment that app is about to be connected to sensitive customer data or to become accessible from outside, it automatically moves into the red zone and enters the formal technical review process. #### Quick risk assessment table To quickly determine where a new project fits and who needs to sign off on it, you can use this table: Risk zone | Where the app runs | What data it can access | Security level | Who approves the launch | Green, low risk | Locally or in the internal test environment | Public data or anonymized information | Minimal, basic validation | The app's creator and a colleague | Yellow, medium risk | The company's internal network | Operational data and internal workflows | Medium, access rules and limits | The IT lead or a technical specialist | Red, high risk | Internet, public servers or external services | Sensitive customer data, transactions, documents | Maximum, code audit and security testing | A software engineer or security specialist | Security testing in the last category means deliberately trying to break into the app, to discover the gaps before someone else does. The classification above is meant to settle, from the start, where people can work freely and where review steps in, without holding the team back. The difference between a company that uses AI to good effect and one that ends up paying for a data breach doesn't lie in the tools it chooses, but in the moment someone looks over the result. --- ### How secure is code generated with AI by someone without technical training URL: https://cicadatech-ai.eu/en/blogs/how-secure-is-ai-generated-code Published: 2026-09-08 Author: Attila Gal, Senior Software Engineer 45% of AI-generated code contains vulnerabilities. See where the risks come from in apps built without technical oversight. The enthusiasm around generative Artificial Intelligence has turned everyone into a potential "software creator". Today, with a few plainly written prompts, a sales manager, a marketer or a business owner can build their own website, a customer management module or an internal app. Without knowing a single line of code. Behind the spectacular speed of execution, however, lies a major problem that the business world is already starting to feel: **code generated without context and without technical oversight creates huge technical debt and massive security holes.** Gartner warns that by 2028, *prompt-to-app* approaches adopted by so-called "citizen developers" (users without a technical background) will increase software defects by **2,500%**, triggering a quality and reliability crisis. Here is what the industry figures show, and why apps generated with AI by people without a technical background can become your company's biggest vulnerability. #### How does AI code generation work, and where do the risks come from? When a person without technical training uses AI to create an app or a website, a phenomenon called *automation bias* kicks in: they tend to trust blindly whatever the screen shows them. If the interface looks good and the button works, the assumption is that everything is fine. In reality, AI models don't "think" about security; they only predict the next piece of code based on patterns in the data they were trained on. Without an engineer who understands the architecture, AI generates code with no context: it doesn't know how the data is protected, who has access to it or how the system behaves under attack. The figures from recent studies on the security of AI-generated code confirm this risk: - **45% of tasks introduce vulnerabilities:** The Veracode GenAI Code Security Report, which tested 80 programming tasks on more than 100 AI models, shows that in 45% of cases the generated code introduces a vulnerability from the OWASP Top 10 list. Java proved to be the riskiest language, with a failure rate of over 70%. - **62% design flaws:** Research published by the Cloud Security Alliance shows that 62% of AI-generated code solutions contain design flaws or known vulnerabilities, even when the latest models are used. The researchers' explanation: the model doesn't know the risks specific to your application, your internal standards or the context it runs in. - **Non-existent software packages:** A study presented at the USENIX Security 2025 conference analyzed 2.23 million code samples generated by 16 models and found that 19.7% of the recommended packages don't actually exist, with 205,474 completely made-up names. The part that turns the mistake into a vulnerability: 43% of the made-up names reappear identically every time the same request is repeated. Attackers can find out which they are, register them in public repositories and load them with malicious code, which the app then downloads on its own. - **A rise in registered vulnerabilities (CVEs):** The Vibe Security Radar project, run by Georgia Tech's software security lab, tracks public vulnerabilities caused directly by AI-generated code. By the end of March 2026 it had confirmed 74 cases, 14 of them critical, on a trajectory of 6 in January, 15 in February and 35 in March. March alone surpassed the whole of 2025. The researchers point out that 74 is a lower bound, since most AI-generated code leaves no traces that would allow attribution. - **Successive tweaks make the problem worse:** The Security Degradation in Iterative AI Code Generation experiment, presented at IEEE-ISTAS, tested 400 code samples over 40 rounds of successive improvements. After just five rounds in which the model was asked to refine the code, critical vulnerabilities had increased by 37.6%. While fixing the visible error, it introduced others at the architecture level. #### Where are the biggest security risks? In apps built without a predefined security architecture, the major risks usually show up in the following areas: - **Personal data leaks (GDPR) and exposed databases:** the app works, but anyone who changes a link can see other customers' data, because nobody asked for authorization rules and the model doesn't add them on its own initiative. - **SQL injection and unprotected forms:** the contact or order form sends data straight into the database without checking it. An attacker can take control of the site by entering malicious code directly into the "Name" or "Message" field. - **Admin account takeover:** the lack of security limits (guardrails) means passwords or access keys (API keys) are left in plain sight in the site's source code, easy to extract for the bots that scan the internet. The consequences don't stay technical. An exposed database containing customer data falls under GDPR, no matter who wrote the code, and the liability stays with the company, not with the tool used to generate the app. #### How do you use AI in your business without putting your data at risk? AI is a tool of undeniable value for speed and prototyping, but it cannot replace the rigor of software engineering. If you want the benefits of automation and fast apps without creating a security crisis in your company, apply these 4 simple rules: - **Use AI for prototypes, not for production:** it's great to use AI to quickly test an idea or a visual flow. But before putting the app in front of customers or connecting it to real data, the code must be reviewed by a specialist. - **Don't automate without human inspection (human-in-the-loop):** no automatically generated line of code should be pushed to a server without a vulnerability review process. - **Set strict rules and access limits:** define clearly which data an app is allowed to access and where automatic task execution must stop. - **Check periodically, not just at launch:** the data above shows that vulnerabilities surface over time, not on the day the app goes live. An app that has moved past the prototype stage and works with company data needs repeated checks. #### Instead of a conclusion… Being able to build apps and automate processes by simply describing, in your own words, what you want is without a doubt one of the most spectacular technological developments of recent years. It's impressive how quickly you can go from an idea to a working prototype today, and how much access to technology has been democratized. AI gives teams unprecedented testing speed and freedom to innovate. The point is not to discourage this initiative, but to complement it with the necessary rigor. For a quickly built tool to become a safe and lasting resource for the company, the agility AI offers must be backed by a sound technical architecture and clear security rules. The competitive advantage comes when the speed AI brings is supported by solid infrastructure that protects your data, your customers and the continuity of your business. --- ### Business Process Automation: A Guide for Companies in Romania URL: https://cicadatech-ai.eu/en/blogs/business-process-automation-guide-romania Published: 2026-08-27 Author: Bogdan Rus, CEO Which business processes you can automate and how to choose between a deterministic agent and AI. A guide for companies in Romania. #### Contents - What business process automation actually means - Where Romania stands against the EU average - Which business processes can be automated - The technologies: what they are and when each one fits - How to choose between a deterministic agent and an AI agent - Two real case studies - The stages of an automation project - When automation doesn't pay off - What drives the cost of an automation agent - Common mistakes in automation projects - Frequently asked questions about process automation - Instead of a conclusion #### What business process automation actually means Tomorrow morning, someone at your company… … will open the report from the payment processor, compare it against the invoices in the accounting software and mark them as paid, one by one. Someone else, in sales, gets a request for a quote, looks up prices in the current price list, checks stock, works out the discount and types up the quote in Word. Or maybe someone in the warehouse takes an order from email, enters it into the inventory system, then generates the shipping label on the courier's website. None of them makes a single decision in any of those steps. They do what was agreed needs doing, in the order it needs doing. The difference between them and a tool is that people get tired and make mistakes on the hundredth repetition. That's the part automation takes over. Data is pulled from the source, checked and put where it belongs, and when something doesn't fit the rules, the case goes to a person. Automation isn't new technology. What has changed in the last two or three years is how much it can take on. A program used to work only with neatly arranged data, in tables and fields. A handwritten email, a PDF sent by a supplier, an order with an unusual note were left to people. Now they can be read and understood. That's also where the confusion in the market comes from. Because you can automate far more than you could three years ago, many vendors have started promising you can automate anything. Technically, that's often true. But a process that happens three times a year, or one that calls for a judgment call every single time, has nothing to gain from automation, even if it's possible. For a business owner, the useful question is what's worth automating in their company and where to start. The rest of this guide answers it. #### Where Romania stands against the EU average In 2025, 5.2% of Romanian companies with at least 10 employees were using artificial intelligence technologies. The European average was 20%. The figures come from Eurostat. No one in the Union ranks below Romania. The next two, Bulgaria and Poland, sit around 8.5%. At the other end, Denmark is at 42% and Finland close to 38%. Look at the trend, too. In 2021, Romania was under 2%. The European average went from 13.5% in 2024 to 20% in 2025, so the rest of the continent is speeding up. Beneath the AI adoption figure sits an older one. CRM systems are used by fewer than 18% of Romanian companies, compared with a European average of 28.5%. Cloud services, by close to 25%, compared with 52.7%. Automation assumes you have systems for data to flow between, and many companies still don't have the systems. There is, however, one number that doesn't square with the ones above. A KPMG study of Central and Eastern Europe found that more than 70% of employees say their organization uses AI tools. The gap between 70% and 5.2% isn't a measurement error. Employees use ChatGPT or other tools on their own, for one-off tasks: summarizing a document, writing an email, drafting a quote. The company reports nothing, because officially it uses nothing. For an owner, this means two things. First, AI is already being used in their company, with no rules about what data leaves the company and no one checking the output. Second, the time saved stays with individuals. One person writes an email faster. The process an order goes through, from customer to delivery, looks the same as it did three years ago. The difference between the two matters for the rest of this guide. A tool helps one person work faster on their own task. An automated process connects the steps between several people and several systems. And if you manufacture or deliver for companies in Western Europe, read those numbers again. The suppliers you compete with for the same contracts work in markets where adoption is four to eight times higher. #### Which business processes can be automated To make this useful, here are a few examples by department. Don't read them as a list of services. Treat them as a recognition exercise: somewhere in your company, one of the things below is being done by hand right now. ##### Finance and accounting - Supplier invoices that get read and entered into the accounting software; - Payments matched against issued invoices, using settlement reports from payment processors, couriers or marketplaces; - Invoices marked as closed, one at a time; - Expense reports; - Month-end reports, pulled together from three or four different places. ##### Sales and quoting - Quote requests that come in by email and have to be turned into quotes; - Prices looked up in the current price list; - Data entered into the CRM after every conversation; - Following up on quotes that were sent and never answered. ##### Customer communication - Order confirmations; - Delivery status notifications; - Reminders to customers who need to come back at a predictable interval; - Answers to the same questions that come up every day: where's my order, when will it arrive, how do I return it. ##### Operations and documents - Orders that come in by email or through the website and then get entered into the inventory system; - Contracts and addenda filled in for every new customer, with the same fields; - Documents that get generated, signed and archived. ##### HR - Leave requests and employment certificates; - Hiring paperwork for every new person; - Screening résumés against clear criteria; - Answers to internal questions that come up month after month. ##### Logistics and shipping - Shipping labels generated with the courier; - Parcel status, tracked and reported; - Stock kept in sync between the inventory system and the online store. The list above isn't complete, and it can't be, because every company has its own processes. But they all have four things in common, and those matter more than the list itself: - They run often; - The rules they follow can be written down; - The data they need lives in a system, not just in an employee's head; - And doing them doesn't take deep professional expertise. Anyone can do them with very little training. When a process checks all four boxes, it's a serious candidate. When one is missing, the conversation gets more complicated, and we'll get to that below. #### The technologies: what they are and when each one fits When you ask for quotes, you'll get proposals that use different terms for things that look the same. Here they are one at a time, because each technology works on a different kind of data and behaves differently when your process, or the systems you work with, change. ##### RPA, or robotic process automation A program that mimics what a person does in an interface: it opens the application, clicks, fills in fields, copies from one place to another. It's used when the system you work with doesn't allow a direct connection, usually older software. The upside is that nothing in the existing system has to change. The downside is serious: if the interface changes, even slightly, the program can't find where to click and stops. ##### Workflow automation This covers the platforms that link applications through ready-made connectors, such as Make, Zapier or n8n. When an event happens in one application (a new order, an issued invoice, a submitted form), an action fires automatically in another. It works well for simple flows between popular applications. It gets complicated when you need your own rules, lots of exceptions, or a system that has no connector. ##### Deterministic agents Programs that execute rules written in advance, with no artificial intelligence component at all. They take structured data, apply the rules and always produce the same result for the same data. They fit processes where the input is predictable: files exported from systems, tables, reports with a fixed format. Their advantage is that they behave the same way every time, and every action can be checked afterward. ##### AI agents Programs that use language models to interpret messy information. They read a free-form email, pull the data out of an invoice in an unfamiliar format, sort a request by what it says. They cover ground that fixed rules can't reach, because the information arrives differently every time. On the other hand, they don't always produce identical results, so processes that depend on them need checks. ##### Deterministic agents and AI agents in the same process In practice, many processes have both kinds of steps. The customer sends a free-form request, so the first stage needs interpretation. Once the data is extracted, the rest of the steps are fixed rules: check, calculate, enter, notify. A well-built process uses artificial intelligence only where it's needed and fixed rules everywhere else. Other terms you'll hear, such as hyperautomation or intelligent process automation, describe exactly that combination, at a larger scale. If you want short definitions, you'll find them in our automation glossary. > **KEY TAKEAWAY:** When you get a quote, ask what kind of solution is behind it and why it was chosen for your process. A vendor who understood your process will answer directly. One selling a standard product will explain why their technology is the best, no matter what you asked. #### How to choose between a deterministic agent and an AI agent The choice isn't about budget, and it isn't about how modern the solution looks. It comes down to one thing: how predictable the documents and files the agent works with are. **When the agent receives files in a fixed format, the right choice is a deterministic agent.** A fixed format means the data looks the same every time. An export from the accounting software. A file from the payment processor, with the same columns in the same order. A courier report. A table filled in from a template. Here the agent applies rules written in advance and produces the same result for the same data. It doesn't interpret anything, because there's nothing to interpret. And every action can be checked afterward, step by step. Predictability matters most in processes where a mistake costs you directly: reconciliations, calculating amounts, compliance checks, generating legally binding documents. In a process like that, a result that varies from one run to the next is a problem, not a feature. **When the agent receives documents that look different every time, you need interpretation.** An email written by a customer. An invoice from a new supplier, with the information laid out differently on the page. A request that could mean three different things, depending on context. Fixed rules fall short here, because you'd have to foresee every possible variation. An AI agent reads the content and pulls out what it needs, whatever form it was written in. The price of that flexibility is that the result isn't identical every time. That's why processes built on interpretation need checks: the agent flags the cases it isn't sure about and sends them to a person instead of deciding on its own. **Most real processes have both kinds of steps.** A concrete example. A customer sends an order by email. The first step needs interpretation: which products they asked for, in what quantity, to which address. Once the data is extracted, the rest of the process is fixed rules: you check stock, calculate the price with their discount, generate the invoice, create the shipping label, send the confirmation. A well-built process uses interpretation only in the step that calls for it and fixed rules everywhere else. Not the other way around. To see where that line falls in your own process, you can describe it in the process automation audit. You get the steps labeled deterministic, AI or human, plus the risk points of each option. ##### Three practical consequences of the type of agent you choose **The first is about compliance.** The EU AI Act brings transparency, documentation and human oversight obligations for certain categories of systems. A deterministic agent, which contains no artificial intelligence component, doesn't fall under it. If you work with sensitive financial or legal data, that distinction has practical consequences. **The second is about verification.** In an internal or external audit, you need to be able to show what was executed and under which rule. A deterministic agent answers that question unambiguously. **The third is about expectations.** If your process requires the result to be identical every time and the vendor proposes a solution built on interpretation, you'll end up with complaints nobody saw coming. **Ask this in the first conversation:** *for my process, what type of agent do you propose, and why?* A vendor who works with both will explain what led to their choice. One who offers a single type will explain why their type is the right fit, whatever the process. #### Two real case studies So far we've talked about principles. Below are two processes we automated at the same company, Aquator Transilvania, a manufacturer of water filtration systems based in Florești, Cluj County. They're two different cases because they start from different reasons: one from an internal cost, the other from lost revenue. ##### Case 1: 118 hours a month lost to marking invoices Aquator sells through its own store and through marketplaces. Money comes in through payment processors, couriers and sales platforms, each with its own settlement report in its own format. The accountant took the settlement reports and marked invoices as paid, one by one, in the accounting software. That work took up 70% of her month, around 118 hours. Volume kept growing, and the company had reached the point where the only option seemed to be a second hire. The process met every condition above: high frequency, rules that could be documented, data that already existed in systems, and the marking itself involved no professional judgment, just matching an amount to an invoice. Now the agent takes the settlement reports and closes hundreds of invoices per minute. The accountant checks the result in about five minutes. She got back 118 hours a month for work where her professional training adds value to the business: reviewing the financial statements, analyzing the numbers, preparing the reports management uses to make decisions. And the company no longer needed the second person it had planned to bring in for invoice marking. See the full case study for the agent that marks invoices as closed in the accounting software. ##### Case 2: reminders sent by hand that didn't reach every customer A water filter needs replacing about every six months. A customer who doesn't get a timely reminder keeps using a spent filter without knowing it. The company loses the order the customer would have placed for a new filter, and the customer drinks unfiltered water, convinced the filter is still doing its job. For a year, reminders were sent with human involvement, through an outside platform. Someone prepared the list, reformatted the phone numbers, uploaded it, wrote the message and scheduled it. Every customer got the same message on the same day, and some messages never arrived at all. The Cicada automation agent connects to the marketplace, pulls the purchases, calculates for each customer when their filter will run out based on the product bought and the quantity, then builds a reminder schedule and sends the message shortly before that date. Instead of one message for everyone, each customer is contacted at their own time. Nobody on the team has to step in anymore. See the case study for the automation agent that turns one-time sales into recurring sales. ##### Time recovered and extra revenue Both processes were repetitive and both ate up time. But the gain is different in kind. In the first case, the company recovered 118 hours a month from work that could be automated and avoided a new hire. In the second case, the gain is recurring revenue. Customers get the reminder before they run out of filter, so the replacement order is placed on time. When you think about a process at your company, ask yourself which of the two categories it falls into. The answer tells you how urgent it is and what you'll compare after implementation to see the result. Want to see which process at your company would be a good fit? Let's look at it together, in a no-commitment conversation. #### The stages of an automation project A project that turns out well goes through the same stages, whatever the size of the process. The order matters more than the speed. ##### 1. The first conversation A conversation in which the vendor understands how you work and you understand what could be automated. By the end, you should know which processes are candidates and which aren't, without having signed anything. ##### 2. Discovery: mapping the process The stage where the process is documented the way it happens today, not the way it's supposed to happen. You sit down with the people who do the work now and write down every step, every rule and every exception. Research on failed projects shows that a good share of failures come from choosing the wrong processes, or from processes that didn't work well even when done by hand, where automation just carried the problems over. Exceptions are the most important part of this stage. The supplier who sends invoices in a different format. The customer with terms negotiated three years ago. The order with a handwritten note. In a real process, situations like these aren't rare. If they aren't documented now, they show up in production, at your expense. At the end of this stage, you should get everything that was agreed in writing: what the agent does, on which cases, under which rules, and what happens in unusual situations. That document goes into the contract. ##### 3. The build The agent is built on the mapped process, the integrations with your systems are set up and the data is synced. How long it takes depends on how many steps the process has and how many systems it connects to. ##### 4. Testing and validation The agent runs on real data from your company, and the result is checked step by step: did it pick up the information correctly, did it apply the agreed rules, did it produce what it was supposed to. Wherever something doesn't match, it gets adjusted. This stage ends when you confirm the agent works as specified in the contract. Without criteria written down in advance, validation comes down to impressions, and the project drags on forever. ##### 5. The agent goes live and is monitored continuously The agent takes over the process. From here on, responsibility for keeping it running stays with the vendor. We, for example, treat the period after implementation as part of the service, not an add-on. The agent is monitored around the clock by the technical team, and the systems it connects to change over time, so we keep an eye on the integrations and adjust them as we go, before any of it reaches you. You have a dedicated account manager who knows your process and whom you contact directly, with no tickets and no need to explain the background every time. You also get access to our platform, where you can see what the agent processed, in what volume and with what result. ##### What separates successful projects from the rest Of the five stages, two carry the most weight: the mapping at the start and the monitoring at the end. The first decides whether the agent does the right thing. The second decides whether it's still doing it six months from now. When you compare quotes, check whether both are included and how they're described. A vendor who skips discovery is selling you a demo. One who leaves monitoring to you is selling you half a service. #### When automation doesn't pay off The most common mistake in automation is treating it as a technical decision. It's a financial decision, and the math takes a few minutes. Take the hours the process uses up each month and multiply them by the hourly cost of the person doing it. Subtract the agent's monthly subscription from that amount. What's left is your net monthly gain. For a process that takes a few dozen hours a month, the subscription pays for itself in the first month. Monitoring and maintenance are included in the subscription, so they don't show up as a separate line in the calculation. That calculation is still incomplete, though, because it only captures the cost of the hours, not what happens to them once they're freed up. An accountant who no longer marks invoices has time for analysis. A salesperson who no longer spends hours filling in the CRM gets to more customers. A support rep who no longer answers the same question a hundred times can focus on the customers who need advice. Same salaries, same people, but the hours go into the work you hired them for. On top of that, repetitive manual work wears down and demotivates people you worked hard to find and would hate to lose. When an agent takes it over, they stay on the part of the job where their experience adds value. That's why the return-on-investment conversation rarely centers on price. It centers on volume and how stable the process is. Below are the situations where the investment **doesn't** pay off. ##### The volume is too low A process that runs a few times a year doesn't save enough hours to cover the investment, however unpleasant it is each time. Frequency outweighs severity: something done twenty times a day is a better candidate than something done once a quarter, even if the second seems more important. ##### The company hasn't agreed on the process If two people at the company do the same task differently (one waits for the customer's approval, the other sends it straight away), you actually have two processes, not one. Automation can't start until the rule is settled. The test is simple: if you can't explain the process to a new employee in five minutes, it isn't ready to be automated. ##### The process is about to change If you're switching ERPs in three months or restructuring the department, the agent gets built on a process that won't exist anymore, and the implementation work has to be done twice. ##### The data isn't reliable If the information the agent needs lives **only** in someone's head or in a notebook, the first step is digitalization, not automation. ##### AI doesn't make up for an unstable process An agent built on a language model does handle unexpected cases better than one built on fixed rules. But on an unstable process it also becomes a source of errors that don't repeat the same way twice, so it raises the stakes instead of removing the problem. Automating a stable process with high volume and few exceptions scales efficiency. Applied to an unstable one with low volume and lots of judgment calls, it scales the mess. When a process falls into one of the categories above, we tell you in the first conversation. Sometimes the recommendation is to sort out the process first and come back in a few months. And the mapping itself has value, because it forces an internal decision that has been put off for a long time. #### What drives the cost of an automation agent The price depends on the complexity of the process, not on the department or the type of agent. Complexity comes down to a few concrete things. ##### How many steps the process has A process where data is taken from one source, checked and entered into another system has four or five steps. One where the customer's request is read, stock is checked, the price is calculated, the document is generated, the quote is sent and the records are updated has twice as many. Every step means rules to define and test. ##### How many systems it connects to An agent linked only to the accounting software needs a single integration. One that works with accounting, the courier, the marketplace and the payment platform needs four. Each connection is built separately, tested separately and monitored separately. ##### How many parameters you need to be able to change yourself If you want to change thresholds, intervals or message texts yourself without calling us every time, that part gets built in accordingly. ##### How many exceptions the process has This is usually where the gap hides between what looks simple and what turns out to be complicated. A process with three special cases gets built quickly. One with fifteen takes more mapping work and more rules. Our tiers follow from these criteria. A process of up to 6 steps, with one input integration and one output integration, falls into the Start tier. One of up to 8 steps, with 2 integrations and 5 configurable parameters, into Plus. Up to ten steps, with 10 parameters, into Pro. Over 10 steps or over four integrations, we talk separately about a custom quote. Criterion | Start | Plus | Pro | Enterprise | Steps in the process | up to 6 | up to 8 | up to 10 | over 10 | Integration steps | 1 input, 1 output | 2 input, 2 output | 2 input, 2 output | over 4 | Configuration tab | — | 5 parameters | 10 parameters | case by case | Results tab | — | 5 parameters | 10 parameters | case by case | Consulting and mapping | included | included | included | included | Dedicated account manager | included | included | included | included | Monitoring platform | included | included | included | included | Training at launch | included | included | included | included | Monitoring and maintenance | included | included | included | included | Prices are published on the site, with the initial fee and the monthly subscription for each tier. You can find them here. ##### 100% guarantee: the agent works the way you asked, or your money back You don't pay for something that doesn't work. Payment is split into 2 stages so that all the risk sits with us, not with you. When you sign, you pay an initial fee that covers discovery (analyzing and mapping your process), the connections, integrations, syncs, training and customization of your agent. The monthly subscription starts only after the agent has been validated and works 100% correctly, as set out in the contract. Until then, you don't pay a single leu in subscription fees. If the agent doesn't pass validation, you choose: we adjust it until it works perfectly, or we refund the initial fee in full, no questions asked. See the guarantee on our site. #### Common mistakes in automation projects ##### Choosing the technology before you understand the process The right order is the reverse: map the process, then pick what fits. When you buy software first and only then look at the process, you end up changing the way you work to fit what you bought. ##### Automating what annoys you, not what costs you The process that drives you up the wall isn't always the one that eats the most hours. Something done once a month but really unpleasant sticks in your memory. Something done 30 times a day becomes routine and goes unnoticed, even though that's where the wasted hours are. ##### Leaving out the people who do the work now They know the situations that don't appear in any procedure: suppliers who send documents in a different format, customers with commercial terms negotiated years ago, the rare cases that block the whole flow. Without them, the mapping stays theoretical, and those situations show up in production. ##### Starting with the most complicated process The instinct is to go after the biggest pain point. In practice, the first project is also the one where you learn how to work with a vendor. A mid-sized process with good volume and clear rules gives you quick results and confidence for the next one. ##### Not putting in writing what "working" means Without criteria written down in advance, validation comes down to impressions. Each side expects something different, and the project drags on without anyone being able to say objectively whether it turned out well. ##### Not paying attention to what happens after implementation An agent that's well built but left without monitoring is half a service. The connected systems get updated, the process changes, and someone has to be responsible for that part. ##### Putting off the launch until the solution covers every single case An agent that handles 8 out of 10 cases and flags the other two to an employee to resolve starts delivering results in its first month of operation. If you wait for the version that also handles those last two rare, complicated cases, the project stays stuck in discussions for months, and the process keeps being done by hand the whole time. #### Frequently asked questions about process automation ##### Where do I start if I have several processes that could be automated? With the one that has high volume and clear rules, not the most complicated one. Look for something that repeats often, has steps you can describe in writing and doesn't involve decisions that require deep professional expertise. The first project also shows you what it's like to work with the vendor, so it's better if it isn't also the hardest one. ##### How long until it's up and running? For typical processes, a few weeks from signing to going live. It depends on how many steps the process has, how many systems it connects to and how quickly we get the information we need during mapping. ##### Do I need programmers at my company? No. The build, the integrations and the maintenance are on us. On your side, we need time from the people who know the process, during mapping, and a clear decision wherever the process is currently done two different ways. ##### Do I lay people off after I automate? No, and that isn't the point. An agent takes over the mechanical steps, not the person. What changes is what their day looks like: the accountant moves from marking invoices to checks and analysis, the salesperson from filling in the CRM to talking with customers. Same people, same team, but the hours go into the work you hired them for. ##### How do I know the agent is doing the right thing once it's live? You have access to our platform, where you can see what it processed and in what volume. At the same time, the technical team monitors how it runs, and you have a dedicated account manager you can call directly. ##### I've already tried Make or Zapier and it didn't work out. Is this the same thing? No. Those platforms are good for simple flows between popular applications, as long as someone at your company sets them up and maintains them. The trouble shows up over time: you add a step, change a condition, hook up another app, and after a few months you have a chain nobody fully understands anymore. When something stops working, you first have to figure out where it broke. With an agent built on your process, the rules are documented and someone is responsible for keeping it running. #### Instead of a conclusion… Automation isn't a decision about technology. It's a decision about which processes at your company eat up hours without needing the expertise of the people you're paying. Three things decide whether a project turns out well. - **Choosing the process:** high volume, rules that can be written down, data that exists in a system, steps with no judgment calls. When one of these conditions is missing, sort it out first and automate afterward; - **The mapping at the start:** this is where the rules, the exceptions and the definition of a working agent get decided. Whatever isn't documented at this stage shows up later, in operation; - **What happens after implementation:** an agent that's well built but left without monitoring is half a service. Romania ranks last in the European Union for adopting these technologies, with 5.2% of companies against a European average of 20%. You can read that number two ways. As a gap between you and the companies you compete with, or as a window, while most of them haven't started yet. If you want to see what could be automated at your company, let's look at a specific process together. The first conversation is free and comes with no commitment: --- ### What an employee who spends 75% of their time on repetitive work really costs you URL: https://cicadatech-ai.eu/en/blogs/cost-of-employee-repetitive-work Published: 2026-08-06 Author: Bogdan Rus, CEO A specialist paid 9,000 RON a month who spends three quarters of their time on repetitive work costs you over 80,000 RON a year. Here's the math. Think of the best person in your company. The one you rely on, the one who was hard to find and whom you'd hate to lose. Now ask yourself, honestly, how much of their day goes into the work you actually hired them for, and how much goes into things anyone could do with a bit of training. A financial analyst spends, on average, only a quarter of their time on the analysis they were hired to do. The other three quarters go into gathering data, cleaning up spreadsheets, matching amounts in the accounting software, or moving numbers from one place to another. The figure comes from a study by the Association for Financial Professionals, run together with APQC, of more than 430 finance professionals. The percentage itself isn't the biggest problem. The fact that it has stayed the same for ten years is. A whole decade of new software, digitalization and promises that technology would fix everything, and the ratio between valuable work and routine work hasn't moved. This is not going to fix itself. And it's not just a problem for accountants or financial analysts. #### The same pattern in every corner of the company Take a salesperson. You hired them for something few people do well: figure out what a customer really wants, handle an objection and close a deal. They're probably among the most expensive people in the company, because they bring in money directly. How much of their week do they actually spend doing that? 28%. That's how much time a sales rep spends actually selling, according to Salesforce's State of Sales report. The rest goes into updating the CRM, looking up information, reports and meetings. In a five-day week, your salesperson sells for about a day and a half. The other three and a half days go to something else. Look closely and it's almost absurd. The systems salespeople feed all that data into, the CRMs, weren't built to help them sell. They were built to give the manager reports. So after every conversation with a customer, the salesperson stops for 10-15 minutes to enter the data into the system, and during that time they're not selling. Keep walking through the company and the pattern repeats. The HR person, who should be handling hiring, people and company culture, spends a good part of the month answering the same five questions dozens of times: how many vacation days do I have left, when does payday hit, what's the procedure for this. The same answers, over and over. The support person, whom you'd want available for the difficult customer or the delicate problem, spends the day on questions that repeat endlessly. Customer service studies show that up to 80% of the questions that reach an agent are repetitive: where's my order, how do I reset my password, what's the return policy. Four out of five conversations need no judgment, no empathy and no experience. Just an answer the person has already given a hundred times. You hired each of them for their expertise, experience and skills. And each of them spends most of the day without needing any of it. #### What you're actually paying for when you pay a salary When you hire a skilled person, you're not buying hours. You're buying what they can do in those hours that someone else can't. A good accountant tells you in February that you won't be able to cover payroll in April, while you still have time to do something about it. A good salesperson knows when to push and when to back off so they don't lose the customer. An experienced support person calms down an angry customer a beginner would lose. That's what you're paying for, and it usually isn't cheap. Now run the numbers for your company. A skilled employee costs you, with salary and payroll contributions, somewhere between 8,000 and 12,000 RON a month, often more. Say 9,000, to stay conservative. If three quarters of their time goes into repetitive work that requires none of the qualities you hired them for, then every month you're throwing away around **6,750 RON** on tasks anyone could do. **Over 80,000 RON a year. For one person.** You're paying specialist rates for trainee work. And you don't see it anywhere, because it doesn't show up on any invoice and nobody cuts it from a budget. It's hidden in the salary you pay anyway, month after month, year after year. #### The cost almost nobody counts: the output is worse, too So far we've talked about wasted time. But there's something worse that few people factor in. Manual work doesn't just eat time. It also produces weaker results. When someone types in hundreds of rows by hand, mistakes happen. A number copied wrong, an amount in the wrong column, a field left blank. Sales studies show that data entered manually into a CRM is often incomplete or unreliable, to the point that decisions based on it rest on shaky ground. So you're paying a skilled person to do work below their level and, on top of that, you get a result you can't fully trust. You lose twice. First, the hours the person spends entering the data. Second, the hours someone else spends checking what they entered and fixing the mistakes. Routine work doesn't need their experience, and it also creates work for someone else later on, when an error comes to light. #### The cost you really don't see: your good people leave Now we get to the most expensive and least obvious part. At first, a good employee does the routine work without complaining. After a few months you notice they're late with the report, that you have to remind them of things they used to do on their own. And when another company makes them an offer, they don't think about it for long, because there's nothing to keep them. The surest way to kill a skilled person's interest is to put them on routine work. And it shows in the numbers. Research shows that teams with disengaged employees have up to 43% higher turnover. And when a good person leaves, it costs you more than you think. The same research shows that replacing them costs between half and twice their annual salary, counting recruiting, time spent on interviews, training, and the months until the new hire reaches the level of the one who left. Now put the whole chain together, because this is where the real cost is. You hire a good person. You pay them well. You have them spend most of their time on repetitive, manual, frustrating tasks. Because of that, they lose motivation. Once they're unmotivated, they're 43% more likely to leave. And when they leave, you pay another half to double their salary to replace them, and then you start over with someone else, whom you'll most likely put on the same routine work. You pay twice, then you risk paying a third time. All because a skilled person was kept busy with something that didn't need them. If you want to see what repetitive work is taking up your people's time right now, the first conversation with a Cicada specialist is free and comes with no commitment. #### The reflex that seems logical and makes the problem worse When the workload grows and people can't keep up, every business owner's first thought is simple: I'll hire someone else. It's a natural reaction. But think about what actually happens. If the new person also ends up spending most of their time on the same mechanical work, you haven't solved anything. You've doubled the cost of the problem. Now you're paying two good people to do mostly routine tasks. And a year from now, when the workload grows again, you'll have the same conversation, just starting from a higher cost. Hiring is the right answer when the company really needs more expertise, more experience, someone skilled to make decisions. It's the wrong answer when the actual problem is that the people you already have are buried in routine work that could be automated. Only 12.5% of organizations manage to get their finance people above the threshold of 40% of their time spent on valuable work. The rest, almost nine out of ten, stay stuck in the same pattern year after year, no matter how good their intentions. Good intentions aren't enough. You have to change who does the routine work. #### The alternative: free up your people instead of replacing them The routine work that takes up most of your people's day doesn't have to be done by them. Pulling data from several sources, matching it up, moving it from one system to another, answering the same question for the hundredth time - all of it follows clear rules and repeats the same way every time. None of it requires a decision. It's exactly the kind of work an automation agent built around your process can take over. And once that part is taken over, people stop leaving (or at least far fewer do), because they start doing the work you pay them for and actually hired them to do. The accountant no longer marks invoices as closed in the accounting software (almost an entire month of work) or copies numbers from one spreadsheet to another - they get back to analysis and to the reports management uses to make decisions. The salesperson no longer spends hours filling in the CRM - they negotiate, send quotes and close deals, which is what you hired them to do. The support person no longer answers the same simple question a hundred times - they help the customers who need advice, the right recommendation or a way out of a more complicated situation. You get back capacity you're already paying for but have been wasting on routine work. Same person, same salary, back to the work the company depends on. That's exactly what happened at Aquator Transilvania. An accountant spent almost three quarters of the month marking invoices as closed in the accounting software. After an agent took over the invoice marking, the accountant had almost the whole month free for reviews, financial statements and analysis. We covered the whole story, with numbers, in the Aquator case study. Want to see what could be automated in your company? Walk a Cicada specialist through a typical day at your company and find out where the opportunities are, with no commitment. #### Frequently asked questions ##### How much does an employee doing repetitive work actually cost me? The cost doesn't show up on any invoice. It's hidden in the salary. If a skilled employee costs you around 9,000 RON a month and three quarters of their time goes into repetitive work, you're throwing away around 6,750 RON a month on tasks that don't need their skills, which adds up to over 80,000 RON a year for one person. On top of that comes a cost that's harder to calculate but just as real: if the person loses motivation and leaves, replacing them costs between half and twice their annual salary. ##### Does automation mean I have to lay people off? No, quite the opposite. An agent takes over the mechanical part of the job. It doesn't replace the person. The point isn't to cut people from the company but to stop keeping them stuck on work below their level. The person stays and handles what takes judgment and experience, which is the work you hired them for and are paying them to do. ##### How can I tell if I'm wasting my good people? Look at what a day looks like for your most skilled people. If someone paid for their expertise spends most of their time copying data, filling in spreadsheets or answering the same questions, work that involves no decisions and that anyone could do, then you're paying for expertise and getting routine work. ##### If I'm paying the person anyway, why automate? Because you're paying for their ability to create value, not for checking off mechanical tasks. Someone kept on routine work doesn't just produce less than they could. They also lose motivation and, over time, leave. When you free them from repetitive work, the same person, on the same salary, gets to do the work the company depends on. Instead of adding a cost, you make better use of one you already have. #### What to remember The most expensive thing in a company isn't the software, the rent or the equipment. It's the good people. And the quietest, most expensive kind of waste is keeping them busy with work anyone could do. It hits you on three fronts at once. You pay specialist rates for routine work. You get weaker output from that work than you should. And over time you lose the very people you rely on, then pay heavily to replace them with others you put on exactly the same work. The figure we started with, three quarters of people's time wasted on work with no value, hasn't changed in ten years. That tells you everything you need to know: the problem doesn't go away on its own, and one more hire won't solve it. It only gets solved when you take a hard look at what your good people do all day and decide to give them back the time you're paying for anyway. If you want to see what repetitive work could be taken off your team's plate, fill out the form with a few details and we'll get back to you as soon as possible. --- ### Customers who have already bought from you are your most valuable ones. How do you make sure they come back? URL: https://cicadatech-ai.eu/en/blogs/case-study-aquator-sms-reminder-agent Published: 2026-07-23 Author: Bogdan Rus, CEO How an agent sends personalized reminders at the right time, so every Aquator customer comes back when it's time to replace their filter. **Case study** **Client:** Aquator Transilvania **Industry:** manufacturing and sales of water filtration systems #### Your cheapest customer is the one you already have Winning a new customer costs you: ads, time, sales effort. A customer who has already bought from you costs none of that. They just have to come back. And if you sell a product people buy again on a regular schedule, their return is the cheapest and safest sale you can make. On one condition: the customer doesn't forget about you. And most of them do, unless you remind them at the right time. Aquator Transilvania sells water filters, a product that needs replacing every 6 months. For a year, the reminders telling customers it was time to order a replacement were sent by hand, through an external platform. It took a long time, needed someone to step in every time, and on top of that, some messages never got delivered. *“A chore. It took a huge amount of time, needed someone to step in every time, and it still missed.”* - Sergiu Maja, managing director, Aquator Transilvania Today, an agent handles all of it on its own: it looks at what each customer bought, works out when they need to replace the filter and lets them know, because many people forget. Every valuable customer is brought back at the right time, without anyone at the company having to step in. #### Who Aquator Transilvania is Aquator Transilvania, based in Florești, Cluj County, makes and sells water filtration systems built on a Romanian filter that won the gold medal at the Brussels International Exhibition of Inventions. It sells thousands of filters a month through its own online store and through marketplaces. Every customer who buys a filter today will need a replacement six months from now. And then another one. For Aquator, this means the relationship with a customer doesn't end with the first sale; that's where it starts. It's kept alive by a message sent on time, reminding the customer that it's time to change the filter. #### The problem: a manual process that missed the most important part The old setup had three weak spots. **Employee time.** Every round of messages had to be prepared and launched by someone at the company, with time taken away from other tasks. **Constant supervision.** As long as an employee has to launch and check every send, the process isn't really automated. The effort still falls on people, just in a different form. **Messages that never got delivered.** This is the most expensive loss, and it helps to start from what it means for the customer. The filter has to be replaced every six months, or the water stops being filtered properly, even if the person keeps using it without knowing. The reminder is how the customer finds out the time has come. When the message doesn't get delivered, the customer is left with a spent cartridge and the false impression that they're drinking filtered water, and the company misses the chance to serve them on time. #### The countless orders you lose without knowing it The cost of an undelivered message doesn't show up anywhere, which is why it's easy to overlook. A wrong invoice gets noticed. A cancelled order shows up in the records. But an order the customer never placed, because nobody reminded them in time, isn't recorded anywhere. The records show what was sold, not what would have sold if the reminder had gone out. For a company that sells a consumable, something that has to be replaced regularly, a big share of revenue is at stake here. A customer can be completely happy with the product and still forget it's time to order a new replacement. If nobody reminds them, you don't lose one order; you lose every order they would have placed from then on. And meanwhile, the customer keeps using a filter that no longer cleans the water the way it should. A manual process that keeps missing lets this cost pile up month after month. Both sides lose: the company loses orders without realizing it, and the customer loses the filtered water they thought they were drinking. #### What we built: an agent that knows when each customer needs a replacement The agent doesn't blast out messages at random. It works on a per-customer basis. It connects to the marketplace and pulls purchases automatically. For each customer, it looks at which product they bought and how many, then works out, based on how long each filter type lasts, when they'll need a replacement. From that calculation, it builds a personalized reminder schedule and sends each customer a message shortly before their filter runs out. In practice, each customer gets the right message at the right time, not a generic SMS sent to everyone on the same day. And nobody on the team has to do anything by hand; the process runs automatically. **Implementation time: 6 weeks. Tier: Pro** (see tier details). #### The discovery stage: the rules behind every message A reminder SMS looks like the simplest thing in the world. The hard part is knowing who to send it to and when. That's where discovery comes in, the initial stage where we learn the company's process in detail and settle every rule the agent will follow. This stage decides whether the agent will run flawlessly or make mistakes, so we give it the most attention. At Aquator, we worked out the whole mechanism together, step by step: - Where the data comes from and how the agent connects to the marketplace to pull each customer's orders; - How long each filter type lasts, which is what the replacement date is calculated from; - How that date is calculated when a customer buys several units at once or several different products in the same order; - How far ahead of running out the customer should be contacted, so they have time to order without a gap; - What the message looks like: the text, the tone, the information it includes, and the easiest way for the customer to order after reading it; - What happens in the less common cases: a customer who just ordered, one with incomplete data, one who no longer wants to receive messages. Every one of these rules was set before a single line of code was written. They decide whether each customer gets a message that's useful to them, when it's useful. At the end of this stage, we document the whole process and put it in the contract. Before any work starts, you know in writing how the agent will work: which rules it applies, how it calculates, when it sends. It's our commitment to you, in black and white. #### What changed, in practice | Before | Now | **Sending messages** | Manual, through an external platform, one round at a time | Automatic, with no input from the team | **Who gets it and when** | The same message, sent to everyone in bulk | Each customer, at the right time for them | **Team time** | Used up on every send | Zero, the process runs on its own | **Undelivered messages** | Lost, without anyone noticing | Sent consistently, to every customer | **For the customer** | Notified hit-or-miss, or not at all | Notified before the filter runs out | #### How to turn a one-time sale into repeat sales An agent like this works in two directions at once, and each one supports the other. Customers get the reminder before the filter stops filtering the water properly and, according to Sergiu, they're glad to get it. A message that reminds them in time to change the filter doesn't bother them; it helps them. This way, the company looks out for its customers and reminds them when they need a replacement. For a company that cares about its customer relationships, that counts for a lot. That's also where the company's gain comes from. A happy customer who's reminded in time orders again. Every reminder sent at the right moment is a sale that gets a chance to happen, instead of one lost among undelivered messages. So from a single initial sale, Aquator ends up with a customer who orders again and again, at every filter change. **A note on the tier:** for Aquator, the complexity of the integrations and calculation rules placed the agent in the Pro tier. In most cases, an agent of this kind fits into the Plus tier (195 EUR monthly subscription), depending on the number of sources and rules. #### Who this kind of solution is for Aquator's story isn't about water filters. It's about any business where a customer who bought once should come back, but only does if you remember them at the right time. You'll find the pattern in plenty of industries, including: - E-commerce for consumables (cosmetics, supplements, pet food, coffee, filters), where customers run out at a predictable interval; - Practices and clinics (dental, medical, veterinary), where patients need regular checkups; - Auto repair shops, where servicing and tire changes come at known intervals; - Companies with subscriptions, policies or contracts that renew; - Suppliers of parts and consumables for equipment. The list goes on. The pattern is the same anywhere customers buy again at an interval you can predict. And it's not just about product replacement reminders. The same logic applies to several recurring processes: - Reminding a customer about an appointment or a service visit as the date gets close; - Renewing a subscription, policy or warranty before it expires; - Winning back customers who haven't ordered in a while; - Following up after a purchase, with products that go well together or a request for feedback; - Letting a customer know when a sold-out product is back in stock. All of these cases come down to the same thing: there's a good moment to reach out to your customer, but most of the time you miss it. Even if you use a platform to send the messages, the process needs manual input every time: someone has to prepare the customer list, often reformat it (add the country code to every number, match up names and so on), upload it to the platform, write the message and schedule it. And in the end, everyone gets the same message, sent in bulk. You know some customers don't come back, but you don't have an automated process that runs on its own, without you, and is tailored to each customer. You have customers who were happy but don't come back in time to reorder. Some forget completely; others remember far too late. If this sounds like you, an agent that automatically sends each customer a message right when they need the product or service again can bring in steady revenue from customers you're losing today. #### Where to start It all starts with a consultation call where we learn how you work. Together we look at your needs, at where time or money is being lost in your current processes and at what could be automated. By the end, you have a fairly concrete picture of the opportunities, without having committed to anything. If none of it makes sense for you, you've lost a conversation, not an investment. If you decide to move forward, next comes **the discovery stage**. We map the process in detail, down to the last rule and exception, and document everything in the contract before starting work. Throughout the project you work with people, not a contact form. You have a dedicated account manager you can call directly when you have a question, and a team that knows your process. You don't have to do anything: once the agent is live, we monitor it through our platform and make sure everything keeps running correctly, including after rollout. If you want to see what could be automated at your company, fill in the form with a few details and we'll get back to you as soon as possible. Fill in the form *See also how we took over invoice marking at Aquator: the agent that won back 118 hours a month of an accountant's time.* --- ### Case study: 118 hours a month of marking invoices, handed to an agent. What changed at Aquator Transilvania URL: https://cicadatech-ai.eu/en/blogs/case-study-aquator-invoice-marking-agent Published: 2026-07-22 Author: Bogdan Rus, CEO How an automation agent took over marking hundreds of invoices a month at Aquator Transilvania and freed up 118 hours of an accountant's time. **Client:** Aquator Transilvania **Industry:** manufacturing and sales of water filtration systems #### 118 hours a month That is how much of a full-time accountant's time at Aquator Transilvania went into marking invoices as paid. Roughly 70% of his working month, out of about 168 hours. The remaining 30% had to cover the accounting for a company that sells through several channels and grows every year. Today, an agent takes the payment settlement reports from couriers and marketplaces and closes the invoices in the accounting software, hundreds per minute. The accountant checks the result. It takes him five minutes at most. Here is how we got there and, more importantly, what it means for a company. #### The case study at a glance The image below sums up the whole process: the problem we started from, the solution we built for Aquator Transilvania and the concrete impact on time and costs. Now for the details: how marking invoices came to take up 70% of the accountant's time, why a new hire wouldn't have solved the problem and how we built the agent that now processes hundreds of invoices per minute. #### Who Aquator Transilvania is The Aquator filter was invented in the '90s by Adrian Costache, a chemist from Buzău. In 1994 it won the gold medal at the Brussels International Exhibition of Inventions for its porous block of activated carbon and silver, which filters water without stripping out the minerals. Aquator Transilvania, based in Florești, Cluj County, brought this filter back into production. It modernized manufacturing and, according to its managing director, now produces thousands of filters a month, sold through its own online store and through marketplaces. Thousands of filters sold every month, across several channels, means hundreds of invoices every month. And the settlement reports for those payments arrive in different formats, from payment processors, from courier companies, from each marketplace. Someone has to match them all up and correctly mark in the accounting software what has been paid. ##### 70% of an accountant's month, on a task that requires no decisions Once orders were paid, the accountant received reports from payment processors, courier companies and marketplaces, each in a different format. Based on them, he marked every invoice as closed. Open the report, find the invoice, confirm the money came in, mark it. Hundreds of times a month, as thousands of orders produced just as many invoices. Marking invoices is part of accounting, but it's routine work that always follows the same steps and calls for no professional judgment. The other 30% of the month had to cover the part where an accountant's expertise brings real value to a company: chasing unpaid invoices and late-paying customers, managing cash flow so the company pays its suppliers and taxes on time, analyzing costs and margins, preparing the reports management relies on to make decisions. These are just a few of the tasks the company's financial health depends on, from a much longer list. Marking invoices isn't one of them. “Our accountant couldn't keep up. She had much more important things to do and was constantly stuck marking these invoices.” - Sergiu Maja, managing director, Aquator Transilvania Aquator keeps investing, and sales grow every year. Each increase brings more invoices with it. A task that eats up this much time and grows along with the company shouldn't sit on one person's desk, done by hand, invoice by invoice. Otherwise you end up exactly where Aquator was: the more the company sold, the less time was left to keep an eye on its financial data. #### The alternative: another hire? In a situation like this, almost every business owner's first thought is to hire someone else. It seems like the logical step when the volume is more than one person can handle. An accountant is paid (and ideally should be) for the things a digital tool can't do: interpret the numbers, figure out where margin is being lost and which costs need a second look, recommend a better tax structure, prepare the reports management uses to make investment decisions. These things take an understanding of context and professional judgment. If a second accountant had also ended up marking invoices 70% of the month, Aquator would have been paying two qualified people to spend most of their time on a repetitive, time-consuming task that can be automated and has no business landing on someone with valuable expertise. The same work, split between two people, at double the cost, permanently. And at next year's volume, a third person would probably have followed. When you need a professional's judgment and expertise, you hire. When you have a process that repeats the same way hundreds of times a month, with fixed rules, you automate it. #### What we built for Aquator Transilvania: the invoice-closing agent The agent takes the payment settlement reports from couriers and marketplaces, finds the invoices in the accounting software and marks them as closed, following fixed rules agreed at the start of the project. It processes hundreds of invoices per minute. That speed means something concrete for the company: the number of orders can grow as much as it wants, and processing time stays about the same. Before, doubling sales meant nearly doubling the hours spent marking invoices. Now it's the same few minutes of run time, whether 200 or 2,000 invoices come in. The accountant no longer marks anything. He looks over what the agent did and confirms everything is in order, in five minutes at most. The nearly 118 hours left over go back into reviews, tax filings and reports for management. **Implementation time: 6 weeks. Tier: Plus** (see tier details). #### The discovery stage: mapping the process before writing any code You don't build an agent that runs this cleanly by jumping straight into code. It all starts with an analysis stage we call discovery: the initial conversation where we get to understand the company's process and map it in detail, step by step. An agent does exactly what it was built to do, nothing more and nothing less. If the process is only half understood, it will handle the usual orders correctly and trip over the ones nobody discussed: an unexpected invoice type, a channel that sends data in a different format, an exception to the rule that nobody mentioned. Every situation left out of the initial conversation later turns into an error someone has to fix. That's why we map the process as it actually happens, not as the internal procedure assumes it does. We go through every step, every exception, every special case. At the end of this stage, we document the entire process and include it in the contract. That way, before the build starts, both sides know exactly what the agent will do: what it takes in, which rules it applies, how it handles each exception. No room for interpretation and no surprises along the way. #### What changed, in practice | Before | Now | **Marking invoices** | ~118 hours a month of manual work (70% of a full-time employee's time) | ~5 minutes of review | **Processing capacity** | Limited by how much one person could get through in a day | Hundreds of invoices per minute | **Effect of growing sales** | More orders meant more hours of manual work | More orders, same processing time | **The accountant's workload** | 70% marking invoices, 30% accounting that matters | ~100% of the time recovered, just 5 minutes checking the agent's work | **The extra hire** | Was the next step | No longer needed | #### What it's worth, in numbers A case study without numbers is just a story, so let's put this in lei. Aquator didn't share its salaries with us and we didn't ask, so we're using public data to get an order of magnitude. ##### The money: a hire avoided The real alternative the company had in mind was a second employee. At the national average gross monthly wage for March 2026, 9,902 RON (INS data), an employee in this position costs the company, including contributions, around 10,100 RON a month. The agent that took over that work falls, in Aquator's case, into the Plus tier, with a monthly subscription of 195 EUR, about 975 RON. Compared with a second salary, the company pays almost 9,000 RON less every month for the same result: invoices closed correctly and on time. On top of the subscription there's an initial fee, paid once (consulting, process mapping, building and integrating the agent, training). With a cost gap that large compared to a new hire, the fee pays for itself in roughly the first half of the first month. ##### The value: 118 hours of expertise recovered Here the gain is a better return on someone you already pay, since that person was already on the payroll. The 118 hours that used to go into marking invoices now go back into the work the accountant was hired for in the first place: cost analysis, keeping track of incoming payments, the reports management uses to make decisions. Same person, same salary, but their time goes into what brings value to the company instead of mechanical execution. **A note on the tier:** for Aquator, the complexity of the integrations placed the agent in the Plus tier. In most cases, an agent of this kind fits into the Start tier (135 EUR monthly subscription, about 675 RON), depending on the number of sources and integrations. The conclusion stays the same: the agent's monthly cost is a fraction of what a new hire costs. #### What the client says “Thank you, Cicada Technologies, for an excellent collaboration! The solution you implemented automated a process that used to take up a great deal of our time, bringing a major boost in efficiency and organization. We appreciated your professionalism, responsiveness and openness in finding the best solutions for our needs, and we look forward to the next implementation!” - Sergiu Maja, managing director, Aquator Transilvania #### Who this kind of solution is for Maybe Aquator's story sounds familiar. It doesn't matter whether you sell water filters or something else entirely; the pattern repeats from one company to the next. For example: - You sell through several channels, and each one produces invoices and reports that land, in different formats, on someone's desk at your company; - You have a qualified person you pay accordingly, but who spends a good chunk of the month on a repetitive task that involves no decisions; - You're thinking about hiring to keep up with the volume, but you're not sure what the new person would actually do the rest of the time; - Your company produces and sells more and more, but the administrative side hasn't kept pace with that growth. If any of this sounds like you, an automation agent would take repetitive, manual, time-consuming work off your company's plate and free your people for the tasks that matter, the ones that bring value. #### Where to start It all starts with a consultation call where we learn how you work. Together we look at your needs, at where time or money is being lost in your current processes and at what could be automated. By the end, you have a fairly concrete picture of the opportunities, without having committed to anything. If none of it makes sense for you, you've lost a conversation, not an investment. If you decide to move forward, next comes **the discovery stage**. This is where we map the process in detail, step by step, down to the last exception: where the data comes from, in what format, which rules apply, what happens in the less common cases. It's the part that asks the most attention from both of us, and it decides whether the agent will then run without anyone stepping in. Throughout the project you work with people, not a contact form. You have a dedicated account manager you can call directly when you have a question, and a team that knows your process. Once the agent is live, we keep monitoring it through our platform, where you can also see how it's working at any time. You don't have to do anything. We make sure it keeps working correctly after rollout, because that's exactly our job. If you want to see what could be automated at your company, fill in the form with a few details and we'll get back to you as soon as possible. Fill in the form See also how we automated Aquator's customer communication: the agent that sends filter replacement reminders. --- ### Automation glossary for business owners: the terms you hear everywhere, explained simply URL: https://cicadatech-ai.eu/en/blogs/automation-glossary-for-business-owners Published: 2026-06-30 Author: Robert Călătoae, Senior Software Engineer RPA, AI agent, workflow, LLM and the rest of the automation vocabulary, explained in plain business terms. A simple reference glossary. If you've ever shopped for an automation solution for your company, you've run into a pile of terms: RPA, AI agent, workflow, LLM, no-code, integration, API. Every vendor uses a different mix, often to describe roughly the same thing, and you walk away feeling you'd need a dictionary to understand what you're actually being offered. That's what this glossary is for. We wrote it for the business owner, not the engineer: each term is explained simply, with an example where it helps, without jargon and without assuming any technical background. You don't need to memorize them. It's enough to be able to come back here when you hit a term you don't understand. We start with the core terms, the ones worth understanding well, then go through a quick list of the rest. #### The main terms, in detail ##### Business Process Automation (BPA) Business process automation means letting software carry out, fully or partly, a recurring process in your company that a person used to do by hand. BPA is the umbrella term that covers everything else in this glossary. The key word is process, as opposed to a single task. Writing an email faster is a task. Automatically taking in an order, generating the invoice, sending it to the customer and updating inventory, every single time, is a process. Everything that follows in this glossary - RPA, agents, workflows - is a different way of doing process automation. BPA is the broad category; the rest are types of tools inside it. **Example:** a company that receives dozens of orders a day and processes them by hand can automate the whole chain, from receiving the order to confirming it with the customer. That's business process automation. ##### AI agent An AI agent is a program that interprets changing situations and makes decisions using artificial intelligence, so it can complete tasks that can't be described with fixed rules. With classic automation, you tell the software exactly which steps to take. With an AI agent, you give it a goal, and it works out how to get there, adapting to whatever it runs into. Its strength is handling things that look different every time: documents in different formats, free-form messages, situations that call for interpretation. Because it interprets, it needs oversight on decisions that matter - a well-built AI agent asks a person to confirm when it isn't sure. **Example:** an AI agent that reads invoices from dozens of suppliers, each in a different format, pulls out the data and enters it into your system, sending unusual cases to a person for review. ##### Deterministic agent A deterministic agent is a program that runs a process by fixed rules, predictably: the same inputs always produce the same actions. The name comes from the fact that the outcome is fully determined by the rules - if X happens, do Y, then Z. It doesn't interpret or improvise. It follows a path set in advance. That makes it very reliable on stable, repetitive processes where the steps are clear and repeat the same way. It's predictable and easy to check. It isn't a weaker version of an AI agent. It's the right tool for a different kind of process: one that needs correct execution, the same way every time, rather than interpretation. **Example:** a deterministic agent that, for every new order in an online store, generates the invoice, emails it and updates inventory, always in the same order. ##### Chatbot A chatbot is a program built to hold a conversation with a person, usually a customer, answering questions by text or voice. It's focused on communication rather than on running an internal process. It talks to someone. Keep in mind that "chatbot" describes what the program does (converse), not the technology behind it. A chatbot can be simple, built on buttons and canned answers, or it can have artificial intelligence behind it. That's why the term sometimes overlaps with AI agent, even though the two aren't the same thing. **Example:** the chat window on a website that answers common questions and books an appointment at any hour, without waiting for a human operator. ##### RPA (Robotic Process Automation) RPA means automating repetitive tasks with software that mimics what a person would do on a computer: clicking, filling in fields, copying data from one place to another. The name says "robotic," but the robots here are programs, not physical machines. RPA is reliable and fast on stable processes with clear steps. Its limit shows up when something changes: if a screen looks different or a field moves, the automation can break, because it follows fixed steps and can't adapt. In short, it works great as long as everything around it stays the same. **Example:** a program that takes data from a file every day, enters it into an accounting app and fills in the same fields, so no one has to do it by hand anymore. ##### Workflow A workflow is the sequence of steps, decisions and responsibilities a task, document or piece of information goes through from start to finish. The important part: a workflow describes how work moves, whether the steps are done by people or by technology. A workflow can be fully manual, fully automated or a mix. This is also where it differs from automation, and many people mix the two up. A workflow answers the question of how work moves: which steps exist and in what order. Automation answers the question of which of those steps are carried out by technology, with no human involved. In other words, automation means having technology run all of a workflow or just part of it. **Example:** when a customer fills out a form on your website, the steps that follow (sending a confirmation email, adding the details to a list, notifying the sales team) make up a workflow. If those steps happen on their own, without anyone doing them by hand, the workflow is automated. ##### LLM (Large Language Model) An LLM is an artificial intelligence model trained on huge amounts of text that understands and produces natural language. ChatGPT and Claude are examples of applications built on these models. The LLM is the technology behind many AI agents: it gives them the ability to read free-form text, understand it and respond or act on it. When you hear that a solution uses artificial intelligence to read emails or documents, there's usually an LLM doing the work. **Example:** an agent that reads a customer's email, understands what they want even though it could be phrased a thousand different ways, and drafts a reply - that's possible because of an LLM. #### Quick list of terms The rest of the terms you'll come across often, with a short explanation: - **Artificial intelligence (AI)** - a program's ability to do things that would normally take human intelligence: understand text, recognize images, make decisions based on data. - **Machine learning** - a branch of artificial intelligence where a program learns patterns from data and improves on its own, instead of being explicitly programmed for every situation. - **Agentic AI (autonomous agents)** - a term for AI systems that pursue a goal and decide the steps themselves, acting with more autonomy than classic automation. It's the more advanced form of the AI agent idea. - **Human-in-the-loop (human oversight)** - the principle that a person stays involved in the important decisions an automation makes. When the program isn't sure, it raises its hand and asks for confirmation instead of acting blindly on high-stakes cases. - **Prompt** - the plain-language instruction you give an AI model to get a result. The clearer it is, the better the answer. - **API** - a bridge that lets two applications talk to each other and exchange data automatically. It stands for Application Programming Interface. It's what lets an agent connect to, say, your invoicing software. - **Integration** - connecting two or more systems so they work together and pass data to each other, without someone moving the information from one to the other by hand. - **No-code / Low-code** - tools that let you build automations with little programming (low-code) or none at all (no-code), through visual interfaces. They make it easy to get started but hit limits once the process gets complex. - **OCR** - technology that pulls text out of an image or a scanned document, turning a PDF or a photo into text a program can read. It stands for Optical Character Recognition. - **NLP (Natural Language Processing)** - the field of artificial intelligence that deals with programs understanding and producing human-like text. It's the foundation of chatbots and agents that work with text. - **Structured vs. unstructured data** - structured data is organized in a clear format (a spreadsheet, a database) and easy to read automatically. Unstructured data is everything else - emails, documents, photos - and harder to process without artificial intelligence. - **Discovery (process mapping)** - the stage at the start of an automation project where the process is understood and mapped in detail before anything gets built. We covered why it matters in a separate article. - **Hyperautomation** - an approach where a company automates as many processes as it can, combining several technologies (RPA, AI agents, process analysis). It's more of an overall strategy than a specific tool. - **ERP / CRM** - the kinds of systems automations most often connect to. An ERP manages the company's resources and operations (inventory, invoicing, accounting), and a CRM manages customer relationships. They stand for Enterprise Resource Planning and Customer Relationship Management. #### How to use this glossary You don't need to remember every term above. Very few business owners know them all, and that's perfectly normal. What matters is this: when a vendor talks to you about RPA, agents or workflows, you can tell what they mean and ask exactly what you need to know. And if you want to understand in depth the difference between the three solutions people confuse most often - the chatbot, the deterministic agent and the AI agent - we wrote a full guide on exactly that distinction. --- ### You connected several apps to automate a process. What happens when they stop talking to each other? URL: https://cicadatech-ai.eu/en/blogs/automation-tool-chain-breaks Published: 2026-06-29 Author: Robert Călătoae, Senior Software Engineer Did you connect apps like Zapier or Make to automate a process? Here's why the chain breaks over time and who ends up fixing it. Nobody sets out to build a complicated system. It starts simple: you have a problem, you find a tool that solves it, you connect it to the app you already use. It works. A month later a new need comes up, so you add another app. Then another, to cover a case the first two didn't. No step seems wrong, because each one solves something specific. A year later, though, you find yourself with five or six apps wired together, each passing data to the next, all held together with patience and a few rules only you still remember. As long as everything works, it looks like a success. The problem is that anything built from many connected pieces has a hidden weakness: a single link only has to stop working, and the whole process stops. This article explains why that happens, what it looks like when the flow stops and, above all, who ends up fixing it. Connecting apps wasn't a mistake. The goal is to help you spot the signs early that you've reached the limits of this approach. #### Why an automation built from several apps is fragile First, to be fair: the apps you use for this, whether Zapier, Make, n8n or others, are good tools. Each one does the job it was designed for, and for simple flows they're exactly what you need. None of them is the problem on its own. Trouble starts when you chain several of them together and an important process depends on that chain. The logic is simple. The more connected apps a process runs through, the more places it can get stuck. If your process depends on five apps passing data to one another, it only takes one of them behaving differently from what the next app in the chain expects, and the whole flow stops. The app itself still works; it just no longer communicates properly with the next one. This isn't a rare or exotic problem. A study published by Zapier, one of the best-known platforms of this kind, found that three out of four companies have already had at least one negative outcome caused by disconnected tools, and more than a quarter already use over ten different apps. In other words, apps that stop talking to each other properly are a common pattern, and more and more companies run into it as they pile up tools. #### How the connection between apps actually breaks The process stops without any warning. No alarm goes off, no red light comes on. At some point, the data simply stops getting where it should. Here are the most common ways it happens. ##### An app changes how it works The apps you connect are constantly updated by the people who make them, completely independently of you. One day, one of them changes the way it sends or receives data, renames a field or changes a format. When an app changes its behavior, you don't get any warning. The breakage happens, but you only find out later, when the data no longer arrives where it should. A common example: many companies automatically pull ad spend data from Meta Ads into their reports, relying on a specific field that holds the amount. All it takes is for the platform to rename that field in an update, and the sync keeps running without throwing any error, except it no longer brings in the amount. Everything looks fine until someone notices, a few days later, that the numbers in the report are blank. ##### The process seems to be running, but data is missing This is the most expensive one, because nothing flags the problem. The process seems to be running. The apps are on, nothing looks stopped. But somewhere in the chain, the data no longer flows correctly: some of it gets lost, some arrives wrong, and you only find out when a customer complains or a report comes out wrong. A typical case: a flow pulls orders from an online store platform. The platform adds a new required field to its response, the flow doesn't know how to handle it, and from then on it skips the orders it can no longer read. At the end of the day, the report shows fifteen percent fewer orders. Depending on how the flow is built, nothing may have flagged the problem. The more the process deals with money, invoices or customers, the more a failure like this costs when nobody notices it. ##### Volume grows and the flow can't keep up An automation built when you had ten orders a day can behave differently once you reach a hundred. Connected apps often have limits on how much they can process in a given time. What worked perfectly at low volume starts to miss, lag or skip tasks as the company grows, right when you most need things to work. ##### Small details that stop everything Sometimes it's something that seems trivial. A date written in a different format from the one the next app in the flow expects. A usage limit being hit. Or an authorization linking two apps that expires after a few months, as it was designed to. Nobody renews it, because nobody knew they had to, and the next day data stops moving from one app to the other. Each of these small things can stop an entire flow, and tracking down the cause takes time and patience, because you have to search across several apps to find the link that failed. #### Who fixes it when the flow stops Anything technical can break; that's understandable. The real question here is: what happens next? The underlying weakness of an automation built from several chained apps is that there's nobody whose job it is to keep it running. The promise you start with is that you set up the automation once and then move on to other things. The reality, once the chain gets long, is that you end up fixing it again and again. And the fixing usually falls to you, or to an employee who isn't a specialist and does it in between other tasks, trying to guess where the flow got stuck. Every time you have to step in, you lose time, the very time the automation was supposed to save you. Even the platforms that offer these tools acknowledge the problem. The same Zapier research shows that companies often end up with employees wasting time moving data by hand between systems that no longer communicate, which means the automation ends up adding work instead of reducing it. You slide into this without noticing: you built something to make your work more efficient, and now the effort has moved into maintaining it. #### When a few connected apps are enough, and when they aren't None of this means you should never use tools like these. That would be bad advice. For a simple flow, with two or three apps and low volume, a connection built in Zapier, n8n or Make is a perfect fit: cheap and fast. The fragility there is small, because there are few links, and if an integration fails, you quickly find out which one. The line gets crossed when the process becomes important to the company. When invoices, customers or money depend on it. When it handles high volume. When it has grown to run through many apps, each one added to cover a new case. At that point, the fragility that was acceptable in a small flow becomes a serious risk, and the time spent keeping it running starts to outweigh what you gain. The signs that you've reached that point are easy to recognize. You spend more and more time hunting for the connection that failed. You've already found lost data that you didn't catch in time. The process has stopped at some point without anyone noticing right away. And most importantly, if you or the person who built the automation were away for a while, nobody else would know how it works. #### The difference between a chain of apps and a system built for the process The alternative to several apps wired together is a different approach. Instead of several separate tools passing data along that you have to watch over yourself, you build a single system, designed for your process from the start, as one whole. In practice, the difference is big. A system built for a process checks whether each step actually ran and whether the data arrived correctly, and when something doesn't match, it sends an alert. A chain of separate tools doesn't do that: it keeps running and reports that everything is fine, even if it skipped half the data. Add one more essential point: a system is maintained by someone whose job that actually is, not by you, in between other tasks. This is where you see the difference between running several connected apps on your own and working with a company that builds your solution, including in code, around your company's actual process. At Cicada Technologies you don't get a tool you have to manage yourself. You get a service with people behind it: a business analyst who maps your process at the start and a dedicated account manager you call directly, not a ticketing system that gets back to you in a few days. Through our own platform, you can see at any time how your solution is working. In practice, the part your company depends on no longer rests on the shoulders of one person on your team. It's handled by a company whose job that actually is. #### Frequently asked questions ##### Why do automations built in Zapier or Make break? Usually the tool itself keeps working. What breaks is the communication between the connected apps. An app in the chain changes how it works, an authorization expires or the volume grows past what the flow can handle, and the process stops. The more apps are linked together, the more places this can happen. ##### How many connected apps are too many? There's no fixed number; it depends on how important the process is. For a simple, low-volume flow, two or three apps are fine. The problem starts when invoices, customers or money depend on an entire chain, and the fragility becomes a risk you can no longer afford. ##### Is Zapier enough for my company? For simple, occasional automations, yes, it's a good tool and a good fit. It stops being enough when the process is critical to the company, handles high volume and runs through many apps, because then the fragility of the chain and the maintenance time start to outweigh the benefit. ##### How do I know when an app has stopped talking to the others? Most of the time, you don't find out right away, because the process gives no sign that anything is wrong. The signs show up indirectly: missing data, a customer complaint, a wrong report. That's why problems in a chain of apps are often discovered late, after they've already done damage. ##### What happens when an app in the chain goes down? If the process depends on it, the whole flow stops from that point on. Data no longer moves forward, and the next steps don't run. Depending on the process, that can mean invoices not sent, orders not processed or information lost, until someone notices and fixes the link. #### Key takeaways An automation built from several connected apps comes from a good instinct and often works for quite a while. But as you add links, the setup gets more and more fragile, because every extra app is one more place where communication can break, most of the time without any warning. Sooner or later, an app in the chain is likely to fail. Do you want to be the one who steps in every time that happens? As long as the process is simple and low-volume, a few connected apps are a good solution. Once it has become important to the company and hard to keep running, you need a system built for your process, maintained by someone else, not you. --- ### Chatbot, deterministic agent or AI agent? A guide for business owners URL: https://cicadatech-ai.eu/en/blogs/ai-agent-vs-chatbot-vs-deterministic-agent Published: 2026-06-22 Author: Attila Gal, Senior Software Engineer Learn how a chatbot, a deterministic agent and an AI agent differ, and how to choose the right one for your company. If you've read anything about automation in the past year, you've seen the words chatbot, agent and AI used almost interchangeably. One vendor promises you an AI agent that takes over your repetitive work. Another sells you a smart chatbot. A third talks about automations. But the same words cover very different things in how they work, what they cost, how reliable they are and, above all, how much risk they bring to your company. The confusion isn't accidental. AI sells, so the label gets stuck on solutions that don't need artificial intelligence to do their job. And for you, the business owner who pays the bill at the end, **picking the right option instead of the fashionable one translates directly into money and into operations that run without surprises.** This guide explains the difference between a chatbot, a deterministic agent and an AI agent, without unnecessary jargon. By the end you'll know what each one does, when it fits and, most importantly, how to choose the right solution for your process instead of the one that sounds most impressive. #### The difference between a chatbot, a deterministic agent and an AI agent, in short A **chatbot** is a program that holds a conversation, usually with your customers, answering questions by text or voice. A **deterministic agent** is a program that runs a process according to fixed rules, the same way every time, predictably, without interpreting anything. An **AI agent** is a program that understands variable situations and makes decisions using an artificial intelligence model, which makes it useful when the work can't be described with simple rules. The three aren't steps on the same ladder, and they don't rule each other out. For many business processes, the deterministic agent is still the safest and most predictable option. Below, we take them one at a time. #### What a chatbot is A chatbot is a program designed to hold a conversation with a person, answering questions by text or voice. It's the most familiar of the three, because you've already run into one: the chat window in the corner of a website, a store's WhatsApp assistant, the automated voice when you call a company. The key point is that the term chatbot describes what the program does, which is hold a conversation, and says nothing about the technology behind it. A chatbot can be very simple, a menu with buttons and predefined answers along the lines of press 1 for opening hours, 2 for our address. Or it can run on artificial intelligence and answer freely, in natural language. In other words, a chatbot can be deterministic or it can have an AI agent underneath. **That's where a lot of the confusion comes from: people use the word chatbot as if it were the opposite of AI, when the two categories actually overlap.** What a chatbot does well is talk to customers. It answers frequently asked questions, books an appointment, qualifies a potential customer and provides basic support around the clock. If your problem sounds like customers message me in the evening and nobody answers, a chatbot is the obvious answer. What a chatbot doesn't do comes down to what it is. It's built for conversation, not for processes. It talks to someone. As a rule, it doesn't take over internal work in your company: it won't reconcile your invoices, move your data from one system to another or close your books for the month. **For that you need an agent, not a chatbot.** **A concrete chatbot example:** A services company gets dozens of messages a day with the same three questions: *what are your hours*, *how much does it cost*, *where are you located*. A chatbot on the website and on WhatsApp answers instantly, at any hour, and passes only the unusual cases on to a colleague. The volume of repetitive questions drops, and the team focuses on what matters. #### What a deterministic agent is A deterministic agent is a program that runs a process according to fixed rules, predictably: the same inputs always produce the same actions. The word deterministic comes from the fact that the outcome is fully determined by the rules. If X happens, do Y, then Z. The program doesn't interpret, doesn't improvise and doesn't make decisions in the human sense. It follows a predefined path that someone laid out in advance. That's exactly where its strength lies. A deterministic agent does exactly what it was told, every time, without variation. On stable, repetitive processes, which make up most of a company's administrative work, **that predictability is worth a great deal.** You know what it does. You can check it. If an error shows up, it repeats identically and gets fixed once, for good. You won't find out that on Tuesday it behaved differently than on Monday for no reason. The technical term you'll often hear for this kind of automation is **RPA**, short for **Robotic Process Automation**, meaning process automation with software robots that imitate human actions: clicks, copying data from one place to another, filling in forms. You'll also come across the term **workflow**, meaning an automated flow that connects steps together: when an order comes in, generate the invoice, email it, update the inventory. Both are, in essence, deterministic. A deterministic agent is good at clear, repetitive processes with predictable steps and a well-defined correct result. Transferring data between systems. Generating and sending documents. Checks that follow the same logic every time. Tracking deadlines. Here it's reliable, predictable and easy to control. It's not good at situations that are new, ambiguous or call for judgment. If the process has unforeseen exceptions at every step, or if the input data looks different every time and needs interpreting, fixed rules start to break down. That's where the AI agent comes in. **A concrete example: ** A company takes orders through an online store. A deterministic agent picks up each new order, generates the invoice in the invoicing software, emails it to the customer and updates the inventory, always in the same order, always the same way. The process is stable, the rules are clear, and the company saves the hours someone used to spend doing this by hand. There's no need for artificial intelligence anywhere, and that's exactly why this solution is the right one. #### What an AI agent is An AI agent is a program that understands variable situations and makes decisions using an artificial intelligence model, to complete tasks that can't be fully described with fixed rules. Unlike a deterministic agent, which follows a predefined path, **an AI agent interprets the context, chooses what to do based on what it encounters and handles inputs that look different every time.** Behind it there's usually an** LLM**, short for** Large Language Model**, an artificial intelligence model trained on huge amounts of text that understands and produces natural language, such as ChatGPT or Claude. That's what lets it read a freely written email, pull out the useful information even when it's worded differently each time, classify a case and act on it. An AI agent is good at work where variability and interpretation are the heart of the problem. Reading non-standard documents and extracting data from them. Sorting requests written in thousands of different ways. Situations where a person would otherwise have to look and decide. Here an AI agent does what a deterministic agent can't. What it doesn't do, and where it needs the most attention, also comes down to what it is. An AI agent isn't predictable the way a deterministic one is. Because it interprets, there's a chance that in edge cases it will interpret wrong. It can reach different results in cases that look similar. It's harder to control and verify, since it interprets instead of following fixed rules. Over time it can drift subtly if the data or context changes. That's why **a well-built AI agent always has safety mechanisms: confidence thresholds, validation steps and human review for high-stakes decisions,** also known as human-in-the-loop. In short, the agent asks a person to confirm when it isn't sure, instead of acting blindly. Many companies start without these safety nets, and that goes a long way toward explaining why they get no results from AI. **An example makes the difference clear:** A company receives invoices from dozens of suppliers, each with a different format, a different page layout and different names for the same things. A deterministic agent would get stuck, because no single fixed rule covers every variation. An AI agent reads each invoice, whatever its format, extracts the supplier, the amount and the VAT, checks them against existing orders and enters the data into the system, while sending questionable cases, like a new supplier or a mismatch, to a person for review. Variability was exactly the problem, and artificial intelligence solves it. #### Comparison table: chatbot, deterministic agent and AI agent Criterion | Chatbot | Deterministic agent | AI agent | What it does | Talks with people, usually customers | Runs a process according to fixed rules | Interprets and decides in variable situations | How it works | Answers, with predefined replies or with AI | Follows rules, doesn't interpret | Uses an AI model to decide | Predictability | Depends on how it's built | Very high, always the same | Lower in edge cases | Control and verification | Simple on simple flows | Easy to verify, reproducible | Needs confidence thresholds and human validation | Reliability over time | Stable on simple questions | Very stable on clear processes | Needs oversight and control mechanisms | A good fit when | You have repetitive conversations with customers | The process is clear, stable and repetitive | The work varies and needs interpretation | Take one thing away from the table. The three aren't competing for the same spot. A chatbot handles communication. An agent, whether deterministic or AI, handles a process. And between the two agents, **the difference is predictable versus interpretive, not old versus modern technology or weak versus powerful.** Each one is better where the other is weaker. #### How to choose the right solution for your company This is the practical part. Instead of starting from I want AI or I want an agent, start from your process and ask yourself a few questions, one at a time. ##### Is my problem communication or execution? If the answer is communication, meaning customers are waiting for answers and writing in on several channels, you're looking at a chatbot. If the answer is execution, meaning steps that need to be done, data that needs to be moved, documents that need to be produced, you're looking at an agent. Many companies ask for an AI agent when what they needed was a chatbot, and vice versa. ##### Does the process follow the same steps every time? If it does, if you can describe what needs to happen as when X happens, do Y, then Z, and that covers the vast majority of cases, then a deterministic agent is enough. **And it's enough as the correct choice, not as a compromise.** It gives you predictability, low cost and control. You're not paying for intelligence your process doesn't need. ##### Does the input need interpreting? If documents, messages or situations come in many different forms and someone would have to look at them and decide case by case, only then do you have a solid reason for an AI agent. Variability is the only strong argument for choosing an AI agent. Without it, AI adds complexity and extra risk with no matching gain. ##### How expensive is a mistake? The bigger the consequences of a mistake, whether money, tax compliance or a customer's trust, the more predictability matters. On sensitive processes, a deterministic agent that's verifiable and reproducible, or an AI agent paired with human validation, is a wiser choice than AI left to decide on its own. Run your process through these four questions and you'll notice a pattern. **For a good share of a company's administrative and operational processes, the clear, stable and repetitive ones, the right answer is a deterministic agent.** That's not because artificial intelligence is weak. On that type of process, predictability and control count for more than interpretation. Artificial intelligence belongs exactly where a fixed rule can no longer cope, and no further. You can run this exercise on a concrete process from your company in the process automation audit: an agent labels each step as deterministic, AI or human and shows you where the risks are. This is also where you see the difference between someone who sells you the newest thing and someone who looks at your process first and only then recommends the tool. The second approach saves you money and protects you from risks you didn't even see. #### Frequently asked questions ##### Is an AI agent the same thing as a chatbot? No. A chatbot talks with people, while an AI agent carries out tasks by interpreting variable situations. A chatbot can have artificial intelligence behind it, but its purpose is still communication, not taking over a business process. ##### When do I need an AI agent and when don't I? You need artificial intelligence when the work involves interpreting inputs that look different every time: non-standard documents, freely worded messages, decisions that depend on context. If the process follows the same steps every time, a deterministic agent is usually enough and preferable, because it's more predictable and easier to control. ##### Is a deterministic agent a weaker solution than an AI agent? No. They're different tools for different types of processes. On a stable, repetitive process, a deterministic agent is actually the better fit, because it offers predictability and control, which are harder for an AI agent to achieve. The AI agent isn't superior, just suited to a different kind of work: the variable kind. ##### Can an AI agent make mistakes? Yes. Because it interprets, there are edge cases where it can decide wrong. That's why a well-built AI agent includes confidence thresholds and human validation for high-stakes decisions. It escalates to a person when it isn't sure, instead of acting blindly. ##### Can I combine all three? Yes, and that's often the best solution. A company can have a chatbot for customer communication, deterministic agents for stable processes and an AI agent only where variability calls for it. What matters is that each one is used where it makes sense, not everywhere. #### Instead of a conclusion The three terms you hear everywhere, chatbot, deterministic agent and AI agent, aren't steps on the same ladder. They're different tools for different problems. A chatbot communicates. A deterministic agent runs clear processes, predictably and verifiably. An AI agent interprets variable situations, with more power but also more cost and more care required. For a business owner, the practical takeaway is simple. Don't start from the tool. Start from the process. Ask yourself what you need to solve, how stable the process is and how costly a mistake would be. Most of the time, for your company's day-to-day processes, you'll find that **the right solution isn't the most spectacular one, **but the one that fits your need best, and that is, surprisingly often, a deterministic agent. --- ### Automated a process with ChatGPT or Excel? Here's where the makeshift solution breaks URL: https://cicadatech-ai.eu/en/blogs/chatgpt-excel-automation-makeshift-solution Published: 2026-06-21 Author: Robert Călătoae, Senior Software Engineer What happens when a process you automated in ChatGPT or Excel starts making mistakes, and how to tell you need a different solution. At some point you had a process that was eating your time - a weekly report, some data processing, messages sent to customers - and you figured, rightly, that there had to be a better way. You opened ChatGPT or set up an Excel file with formulas and maybe a macro, linked a few steps together, and it worked. The process moved faster, and you solved a problem without waiting for a budget, a vendor or approvals. That's not a mistake. It's exactly the right instinct for a business owner looking for solutions with what's on hand. And you're not alone: more and more business owners, at companies of every size, are trying to automate repetitive processes with the tools they already use. The trouble comes later, when the solution you built starts to creak - and usually you can't tell right away why. This article shows where automation with ChatGPT or Excel breaks down, why it happens and how to recognize the point where you've outgrown the tool. The goal isn't to tell you that you did it wrong. It's to show you where these tools stop, before a mistake costs you. For some perspective: a 2025 MIT report looked at hundreds of AI initiatives in companies and found that, despite tens of billions of dollars invested, the vast majority of organizations get no business results from them, precisely because these solutions are fragile and don't fit the way work actually gets done. In other words, if the solution you built has started to let you down, you're not an isolated case. More often than not, it's a pattern. #### Why ChatGPT and Excel look like the right answer at first Before getting into where they break, it helps to see why they work at first - because they do work, no question about it. ChatGPT and Excel are right there, you already have them, they cost next to nothing extra and you don't depend on anyone to use them. You can shape them quickly: write an instruction, build a formula, link two columns and get a result in minutes. For a simple process, that really is enough for a good while. You get speed right away, with no investment and no waiting. That's exactly why they're the natural choice. A business owner short on time isn't looking for the perfect solution, just the one that gets them unstuck today. ChatGPT and Excel do exactly that. Starting with them isn't the problem. The problem is that these tools were designed for something other than carrying a business process that repeats day after day, month after month - and you only see that later. #### Where the makeshift solution starts to break Things don't get complicated with an announcement. The solution doesn't suddenly stop. It starts making mistakes here and there, needing more and more attention, depending on more and more things that all have to go right at the same time. Here are the three ways that happens. ##### Hidden errors that cost money The most dangerous problem is the one you can't see. An Excel formula that pulls from the wrong cell after you inserted a row. A ChatGPT answer that looks right but contains a made-up number or misreads an unusual input. These mistakes don't stop the process, which makes them far more expensive - they slip into invoices, reports and decisions, and you find them late, after the damage is done. A marketing specialist would call this functional risk: the solution doesn't perform the way you expect, but because it seems to work, nobody checks. The more the process touches money or customers, the more a silent error costs you. ##### The time maintenance swallows A makeshift automation works perfectly on the day you build it. Then something around it changes: the supplier who sends you data changes the file format, an app changes how it responds, someone adds a new column, ChatGPT gets an update and starts answering differently. Each time, something breaks, and you or someone on your team has to step in and fix it. So the solution that was supposed to save you time ends up eating it - not all at once, but in small, repeated fixes that are hard to predict. These solutions fail because of fragile workflows and a poor fit with day-to-day operations. They're built on sand, and sand shifts. ##### Everything depends on one person The third problem is the hardest to spot. A homemade solution usually lives in one person's head - yours, or the employee who built it and knows how it's wired, what each step does and where to look if it breaks. As long as that person is around and remembers, everything seems fine. But if they leave the company, go on vacation or simply forget a detail a few months later, you're left with a mechanism nobody understands anymore. There's no documentation, no one to take it over, no system - just a personal build you can't hand off. For a company that wants to grow, that dependence on one person is a hidden weak spot. #### Why this happens All three problems have the same root: ChatGPT and Excel weren't designed as process systems. They're general-purpose tools, very good at what they do - Excel is a calculation tool, ChatGPT is an assistant that understands and produces text, among other things. Neither was built to carry, unsupervised, a business process that runs indefinitely, with inputs that change and a lot riding on the result. If you want to understand the difference between the types of solutions, we explained in detail the difference between a chatbot, a deterministic agent and an AI agent. A system built for a process has stable rules, catches exceptions, flags when something is off, can be checked and can be handed over to someone else. A makeshift solution has none of these safety nets. It works as long as everything around it stays the same and the person who remembers how it works is around. Otherwise, it's exposed. There's also a common thinking trap here. A makeshift solution seems free, because you didn't pay anything visible for it. But its cost isn't in the purchase price. It's hidden elsewhere: in maintenance time, in errors caught too late, in the risk of getting stuck. Specialists call this total cost of ownership - what a solution really costs you over the whole time you use it, not just up front. Seen that way, free at the start can turn out more expensive than it looks. #### How to tell you've outgrown the tool Not every makeshift automation needs replacing. Plenty of simple processes live happily for years in an Excel file with a few formulas, without causing headaches. The question isn't whether your solution is good or bad, but whether you've passed the point where the tool is still enough. Here are a few signs that you have: - You regularly spend time fixing something that should run on its own; - You've already found at least one error that reached a customer or an important report; - The process has stopped at some point because one person was missing or didn't remember how it worked; - Only one person can explain how the solution is built; - As the company grew, the solution started struggling with higher volume; - You rely on it for something that matters - money, compliance, customers - but have no way to check that it works correctly every time. If you recognized yourself in two or more of these situations, it doesn't mean you did something wrong. It just means your process has grown beyond what a general-purpose tool can carry, and it's worth thinking about it differently - as a system built for that process, not a workaround kept alive with effort. #### Frequently asked questions ##### Can I automate a process with ChatGPT? Yes, up to a point. ChatGPT handles one-off text tasks well - summaries, drafts, replies. The limit shows up when you want that to become a stable process that runs on its own, with changing inputs and results you need to be able to rely on. That's where fragility and hard-to-catch errors come in. ##### Why doesn't Excel automation work in the long run? Excel works great for calculations and simple workflows, but a macro or a file full of complex formulas is sensitive to any change: an inserted row, a new column, a changed format. Each change can quietly break a formula, and the resulting errors are hard to notice. Over time, maintenance ends up costing more than it saved. ##### How do I know if my makeshift solution is still good enough? Look at three things: how much time you lose fixing it, whether it has already caused errors with consequences, and whether it depends on a single person. If any of the answers worries you, you've probably outgrown the tool. ##### Isn't a professional solution just an extra expense? It has a visible cost, that's true. But the right comparison isn't with zero. It's with the hidden cost of the makeshift solution - maintenance time, errors, the risk of getting stuck. Looked at as a total cost over the whole time you use it, the math often changes. ##### Why do so many companies get it wrong when they automate on their own? Because they put a general-purpose tool to work on a job that needs a system built for the process. The 2025 MIT report shows that most AI initiatives in companies produce no results for exactly this reason: fragile workflows and a poor fit with how work is actually done. #### In place of a conclusion If you automated a process with ChatGPT or Excel, you did something good: you looked for a more efficient way with the means you had. These tools often take you surprisingly far. But only up to a point. Beyond that point - when errors slip in silently, when maintenance eats your time, when everything hangs on one person - a better tool won't fix it. You need to look at the problem differently: as a process that deserves a system built for it, one that's stable, can be checked and doesn't depend on anyone's memory. Only you can say where your process is on that road. What matters is knowing the road exists and recognizing the signs when they show up. --- ### Why 9 out of 10 companies get no results from AI, and what the successful ones do differently URL: https://cicadatech-ai.eu/en/blogs/why-companies-get-no-results-from-ai Published: 2026-06-20 Author: Bogdan Rus, CEO Almost every company uses AI, but few get results. See why most projects fail and what's behind the ones that succeed. Almost every company uses artificial intelligence in some form today. One person writes copy with it, someone else uses it for reports, a third put an assistant on the website. Look around and it seems everyone has already jumped on board. And yet, ask what concrete result it has brought the company, and the answers trail off into generalities. The vast majority of companies that use artificial intelligence get nothing out of it that shows up in profit - serious estimates from 2025 put it at about 95% of organizations with no measurable result. In other words, adoption is nearly universal, and results are rare. The useful question isn't why so many companies fail, but why some still succeed. What do they do differently? #### More money doesn't mean better results The first explanation everyone jumps to is budget. The companies that succeed spent more money, bought pricier tools, hired more people. It sounds logical, but it's wrong - and the difference isn't academic, because it decides whether a company without a corporate budget is beaten before it starts. In reality, the big companies, with the deepest pockets and the most specialists, are the ones that get stuck most often. They throw money at dozens of initiatives and end up with prototypes that never produce anything. If money were the lever, big corporations would have won long ago, and a mid-size company wouldn't stand a chance. The door is open for an ordinary-size business too. If budget isn't what separates success from failure, the difference lies in something you can control no matter how big you are: how you put artificial intelligence to work. Here's where most companies go wrong, and then what the successful ones do differently. #### Why most fail Failure doesn't come from bad technology. It comes from three mistakes in approach that show up again and again, whatever the industry. ##### They use a general-purpose tool where a system is needed The most widespread mistake is also the easiest to understand. Someone in the company opens a general-purpose AI tool and uses it to work faster - writing an email, summarizing a document, making a draft. Useful for that one person, but with no effect on the company as a whole: the person works faster, yet the company runs exactly as it did before. It's the difference between using AI as a personal assistant and having it run a process from start to finish. The first makes your day easier. The second changes how the company works. Most stop at the first and wonder why it doesn't show up in the numbers. ##### They put artificial intelligence on top of a process that's already broken The second mistake costs more, because it looks like progress. The company takes a process that was already running badly - slow, messy, full of exceptions nobody ever sorted out - and puts artificial intelligence on top of it, hoping the technology will fix it. It won't. Automate a broken process and you get a broken process that runs faster. Mistakes now happen in seconds instead of minutes, at larger scale, and travel further before anyone catches them. Technology doesn't bring order to a badly designed workflow. It just speeds up the mess. ##### They invest where it shows, not where it pays The third mistake is about where companies focus when they automate. Enthusiasm goes to the visible areas, while the easiest gains from automation often sit in the back office, in routine processes: invoicing, reporting, repetitive admin work, where a process handled properly brings direct savings that are easy to measure. The reason has more to do with perception than with math. A visible project feels like a step forward and looks good in a presentation, while automating an administrative process impresses no one - even though that's exactly what would have brought the savings that show up in the numbers. #### What the successful ones do differently The companies that end up on the right side of the line do the exact opposite. And none of their choices requires a corporate budget - just clearer thinking. ##### They start from a process, not from technology Successful companies don't start by asking how to use AI. They start with a specific process that eats their time or money and ask how to make it work better. Technology comes last, as the answer to a precise problem, instead of coming first as a solution looking for a problem. The difference seems small, but it changes everything: one starts from a need, the other from enthusiasm. ##### They rethink the process before automating it This may be the most important choice, and the one that most clearly separates success from failure. Before putting artificial intelligence on a process, successful companies rethink it - they look at how the work is really done, cut the unnecessary steps, sort out the exceptions and only then automate a clean workflow. They don't bolt technology onto the old way of working. They build a better way of working that technology can carry. It's the difference between automating chaos and automating order. The first speeds up the problems; the second solves them. We covered separately why every good automation starts with an analysis. ##### They leave repetitive tasks to automation and decisions to people Successful companies don't fall for the illusion that artificial intelligence can handle everything on its own. They have it do the repetitive heavy lifting - processing, sorting, moving data - and keep people exactly where their judgment matters: on exceptions, on high-stakes cases, on decisions where a mistake is costly. It isn't either-or. Each does its own part, and that safety net is exactly what lets them trust the system. ##### They lean on someone who has done it before Successful companies rarely build everything themselves from scratch. They lean on the experience of someone who has been through the process before and knows where the traps are. It's a matter of risk, not pride: the numbers show that solutions built with a specialized partner succeed about twice as often as those built in-house. Trying it alone, without ever having seen what the road looks like, is exactly how most companies end up in the failure statistics. #### What this means if you don't have a corporate budget If you run a mid-size company, everything said so far works in your favor. The line between success and failure isn't drawn by budget, size or the number of specialists on staff. It's drawn by four choices: you start from a concrete process, you rethink it before automating it, you keep people on the important decisions and you lean on someone who knows the road. All four are within your reach, and not having a corporate budget doesn't condemn you to the wrong side of the line. The only thing that keeps you there is the wrong approach, and an approach can be changed at any time. #### Frequently asked questions ##### Why do most artificial intelligence projects fail? Because of three mistakes in approach: companies use a general-purpose tool where they'd need a system tied to the process, they put artificial intelligence on top of a process that was badly designed to begin with, and they invest in visible areas instead of the ones that bring concrete results in the numbers. None of these has to do with budget. ##### Do I need a big budget to get results from AI? No. The companies with the biggest budgets often fail, because money isn't the lever. What matters is the approach - a process chosen wisely and rethought before automation - not the amount invested. ##### What do companies that succeed with artificial intelligence do differently? They start from a concrete process, not from technology. They rethink the workflow before automating it, keep people on the important decisions and lean on the experience of someone who has been through the process before. ##### Does company size matter for success with AI? Less than you'd think. Small and mid-size companies decide quickly and have less bureaucracy, which makes implementation easier. Speed and clarity matter more than size or resources. ##### Is it better to build a solution myself or work with someone? Solutions built with a specialized partner succeed about twice as often as those built on your own. The experience of someone who knows the traps greatly reduces the risk of failure. #### The takeaway At first glance, the numbers are discouraging: almost everyone uses artificial intelligence, and almost no one gets results from it. Look again, though, and the picture changes completely. Failure doesn't come from weak technology or lack of money. It comes from a way of using it - a general-purpose tool placed on top of some process, without changing anything about how you actually work. Success comes from the exact opposite: a process chosen wisely, rethought before it's automated, with people kept where their judgment matters and with the support of someone who knows the road. None of this requires a corporate budget or size. It only requires putting artificial intelligence to work where it belongs, and in the right way. --- ### Don't automate a disorganized process. A digitalization project starts with an analysis URL: https://cicadatech-ai.eu/en/blogs/analyze-before-automating-disorganized-process Published: 2026-06-19 Author: Bogdan Rus, CEO Automating a disorganized process only speeds up the mess. See why every digitalization project starts with an analysis stage called discovery. You have a process that has been driving you crazy for months. It takes too long, it always gets stuck in the same place, and someone on your team loses hours to it every week. At some point you decide: enough, let's automate it and be done with it. It's a good decision, but from here the road splits in two - and which way you go decides whether you'll be happy a year from now or constantly fixing something that shouldn't have been built that way. The first option, the tempting one, is to jump straight to the solution. You take the process as it is, hand it to someone to automate quickly and wait for the result. The second, which looks like a pointless delay, is to stop first and understand the process in detail before touching any tool. This article shows why the second option, even though it takes patience up front, is the only one that leads to a result that holds up. #### What happens when you automate a disorganized process There's a common illusion that automation fixes a bad process. You put technology on top of it and somehow things sort themselves out. In reality, the exact opposite happens. When you automate a disorganized process, you don't bring order to it - you encode the mess and speed it up. Everything that went wrong by hand will go wrong automatically, only faster and at larger scale. Unnecessary steps stay unnecessary. The exceptions nobody ever sorted out stay exceptions, except now they're handled by a tool that can't improvise. A person processing an order can notice when something doesn't fit and make a call on the spot. An automation does exactly what it was built to do, including when the input doesn't match what it expected. That's the heart of the problem. When you build an automation on a process you haven't fully understood, you're automating an assumption about how you think the process works, not the reality. And that assumption almost never covers everything. It works for the usual case, the one you thought of, and breaks at the first exception you didn't foresee. It's one of the reasons many companies get no results from AI. This isn't a lone opinion: IBM, drawing on decades of experience with companies in the middle of digitalization, notes that too often organizations automate disorganized or poorly executed processes, and knowing what to automate and what not to is the first step toward a successful plan. #### Discovery: the step that haste skips In serious automation and digitalization projects, the opening stage, where you understand how the process really works before building anything, has a name: discovery. No tools get touched and no code gets written - you observe, ask questions and map things out. How it starts, where it goes, who does what, where it gets stuck, what happens when an unusual case comes up. It looks like a step that slows things down, but it's the other way around. Discovery is what separates a project that works from one that turns into endless repairs. Any serious purchase a company makes starts the same way, after all: with understanding the problem, before picking the product. Before you buy a machine or switch suppliers, you first get clear on what you really need. Automation is no exception, even if the temptation to jump straight to the impressive part is stronger. There's a simple psychological reason people skip discovery: this phase is invisible and impresses no one. You can't show a process map in a meeting the way you show a running automation. And yet that unseen map is what decides whether the automation runs well or falls over. #### What a good analysis brings to light The value of an analysis isn't the document it produces. It's what it sees that you no longer see, because you're too close to the process. Here's what usually comes to the surface. **The exceptions you've never put into words.** Every process has special cases, situations that don't follow the rule. You and your team handle them on instinct, without ever naming them. An analysis brings them into the open, because that's exactly where automations built in a hurry fail. **How the steps actually connect.** On paper, a process looks like a straight line. In reality it has branches, loops back and steps done in parallel or in a different order than everyone thinks. An analysis shows how the process really flows, because a solution that doesn't fail at the first step that doesn't fit has to be built on the real workflow, not the imagined one. **The knowledge only one person has.** There's usually one person who knows how the process really runs, with all its detours. If that person leaves or is out, the process wobbles. An analysis turns what's in their head into something written down that you can build on. There's one more thing an honest analysis can tell you: that, as it stands, your process isn't ready to be automated. Sometimes a process is so disorganized that putting automation on top of it would only speed up the mess. A good partner tells you that openly instead of selling you an automation that wouldn't hold up - because they're working for your result, not for the invoice. #### Why the analysis takes someone who knows how to do it The temptation is to think you can do the analysis yourself, on the fly, between other tasks. You know the process; it's yours, after all. But that's exactly the trap: because you're too close to it, you see how you think it works, not how it actually works. And the gap between the two is exactly where the problems hide. A good analysis needs an outside eye that has seen dozens of processes and knows what to look for: where exceptions usually hide, which questions surface unnecessary steps, what a process ready for automation looks like compared to one that only seems ready. It's a skill in its own right, not something you improvise between two meetings. At Cicada Technologies, every engagement with a client starts with this discovery stage, before we automate anything. We don't treat it as a formality but as the foundation of the whole project, because we've seen too many times the difference between an automation built on a process understood end to end and one built on an assumption. The first lasts for years. The second needs fixing month after month. If you want a first picture before a conversation, you can describe a process in our process automation audit. It doesn't replace a full discovery, but in a few minutes it shows you which steps work with fixed rules, where AI would be needed and where the risks are. #### Frequently asked questions ##### What is discovery in the automation process? Discovery is the opening stage of an automation or digitalization project, where the process is understood and mapped in detail before anything is built. You look at how the process really runs, where it goes, where it gets stuck and what happens in unusual cases. Only after this phase do you move on to the automation itself. ##### Why can't I just automate a process directly, without an analysis? Because without an analysis, you're automating an assumption about how you think the process works, not its reality. The automation will handle the usual case and break at the first unforeseen exception. The analysis surfaces the exceptions, unnecessary steps and hidden details before they turn into problems. ##### How long does an analysis before automation take? It depends on how complex the process is, but it's usually much shorter than haste would suggest. A clear process can be understood in a few hours of focused conversations. The processes that take longer are exactly the ones nobody has ever looked at as a whole - and those were the ones that needed analysis most anyway. ##### What happens if I automate without an analysis? Most often, you end up in a cycle of repairs: the automation works at first, then breaks at each unforeseen situation, one after another, and you pay in time and frustration what an analysis at the start would have saved you. In some cases, the automation ends up costing more than the value it brings. ##### Does every disorganized process need automation? No. Sometimes a process runs badly only because it's disorganized, and putting it in order solves the problem without any automation. An honest analysis can tell you exactly that - and save you an unnecessary expense. #### What to remember The temptation to jump straight to automation is understandable. You want to be rid of a process that drives you crazy, and analysis feels like a delay between you and the result. But that very haste is what turns automations into sources of trouble, because a tool built on a process nobody understood doesn't fix the mess. It speeds it up. The discovery stage is more than a formality. It's the step that decides whether your automation lasts for years or breaks at the first exception. A process understood end to end, with its exceptions, its quirks and its reality mapped before any technology comes in, is the only kind worth building on. ## Contact - Email: contact@cicadatech.eu - WhatsApp: +40 759 170 588 - LinkedIn: https://www.linkedin.com/company/cicada-technologies/ - Website: https://cicadatech-ai.eu/en