← Back to blog

How to test an AI-generated app, by risk class

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.