GUIDE · AI ACT

How to classify your AI systems by risk level

After this guide you will be able to build an inventory of your company's AI systems, apply a sequence of checks to assign each one a risk level, and record the decision so that it can be verified. The procedure helps you decide what to do next; the final conclusion should be confirmed with a consultant.

Updated october 2026
7 min read
Altovar Technical Editorial Team

1. Why classification comes first

The AI Act organises obligations according to risk. A prohibited system, a high-risk one, one subject to transparency obligations and one posing minimal risk each call for very different actions. If you do not know which category a system falls into, you cannot decide what to document or what to ask the supplier for. That is why classification is the first operational step, even before choosing a consultant for the whole compliance project.

Risk depends on use, not on the product name. The same text generation tool may have low impact if it helps write an internal draft, and become a sensitive matter if it is used to evaluate people. Classify therefore what you do with the tool, not the brand of the technology, and record the difference between one use and another.

2. Building the inventory of systems

Start with a complete list of the systems that use AI, even if you did not buy them as AI products. Many management, email, document management and HR software packages include built-in AI features: suggestions, summaries, scores, automated searches. Ask department heads which tools they use and for what tasks, and compare the answers with your list of active contracts and licences.

For each entry, note four pieces of information. The point that is often missing is the last one: the effect on people. A system that proposes a priority order for customer requests affects different people from one that helps compose an internal reply, and the difference matters for classification. If a system has no direct effect, describe it anyway: it is often the indirect effect that makes the difference to risk.

FieldGuiding questionGeneric example
Who uses itWhich function or department uses itAn administrative office
For which taskWhat it produces or suggestsDrafts of replies to customers
On which dataWhether personal or sensitive data is involvedCustomer records
Effect on peopleWho is affected and in what wayNo direct decision, support only

3. The sequence of checks

Once the inventory is complete, always apply the checks in the same order. The first concerns prohibitions: if a use falls under a prohibited practice, nothing else matters and the use must be stopped. The second concerns whether it appears on the list of high-risk use cases or acts as a safety component in an already regulated product. Only after that can you assess transparency obligations and, for everything else, minimal risk.

Following the same sequence avoids the temptation to close a case by saying the system is harmless. A quick check is enough to rule out clear cases. Ambiguous cases should be discussed with the consultant, the supplier and the people who really know the work process in which the tool is embedded. Note each doubtful case with the open question, so that the consultant can close it without repeating the work.

  1. Check that the use does not fall under a practice prohibited by the regulation, such as manipulation that exploits a person's vulnerability.
  2. Check whether the system is a safety component of an already regulated product, or whether its use appears in the list of high-risk cases.
  3. If the system interacts directly with people or generates synthetic content, identify the transparency obligations.
  4. Determine your role: developer, provider placing it on the market, or simply a deployer, meaning a user of the system.
  5. If none of the previous steps applies, record the system as minimal risk, with the reasoning written down.
  6. Assign an internal owner who approves the classification.
  7. Review the classification each time the use, the supplier or the type of data processed changes.

4. The most common mistakes

The first mistake is classifying by supplier: the tool comes from a large brand, so it is safe. Classification depends on the use you make of it, not on the reputation of whoever produces it. The second is forgetting the AI functions inside software already in use, which are often switched on by an update and which no one has ever assessed. The third is thinking that internal use takes the system outside the scope: a system that evaluates candidates or employees remains relevant even if no customer ever sees it.

The fourth mistake is treating classification as closed once and for all. Changing the way the tool is used, connecting it to new data or assigning it to another function can alter the risk level. Set a periodic review and a clear criterion for when a change requires the assessment to be redone. Remember that the date of the review is part of the decision, not an administrative detail.

5. How to verify the classification and what to keep

A good classification is recognisable because every decision has a written rationale that a third party can read and understand. Carry out a sample check: pick three systems and ask a person who did not classify them to retrace the steps. If their conclusion differs, the reasoning is not clear enough and should be rewritten. Also verify that the supplier confirms the information about the system's functions.

The classification register is the most useful document you will produce at this stage. For each system keep the description, the use, the data involved, the role, the result of the checks, the rationale, the owner and the date of the decision. Keep earlier versions of the register too when decisions change: whoever reviews it will want to see how the conclusion was reached, not only the conclusion itself.

6. How to start

This week, do just one piece of work: a list of the AI systems you already know you use. Ask three people, one from administration, one from sales and one from HR, which tools they use that resemble an assistant, a generator or a system that suggests decisions. Even incomplete answers are useful, because they show where information is missing.

With the list in hand, pick the system that affects the most people and apply the steps of the sequence. That first case will show you which information is missing and will serve as a model for the others. Take the result to your compliance consultant and to your IT supplier to verify which functions are active. If the result is unclear, it is better to postpone the case than to close it with an approximate answer.

Frequently Asked Questions

Does a system need classifying even if it is free?

Yes. Being free does not exclude a system from the regulation. What matters is what it does and how you use it. A free tool that helps select candidates deserves the same attention as a paid one. Include free tools used by employees for work in the inventory too, because they are often the ones that escape controls.

If the supplier declares the risk level, can I rely on it?

The declaration is a starting point, not proof. The supplier describes the product in general terms, while the risk depends on your use. Compare its indication with the purpose you have in mind, record any differences and ask for confirmation in writing. If the declaration does not match your case, the assessment is yours to make, with the consultant.

Do I need software to carry out the classification?

It is not essential. A well-structured worksheet, with the same columns for every system, may be enough for a small company. Management software can help when systems are numerous or change often. In any case the tool does not decide on your behalf: the rationale and approval remain the responsibility of an accountable person.

Who must approve the classification?

An internal manager with authority over how the system is used, usually the head of the department that deploys it, supported by whoever manages IT and data protection. The approval must be explicit and dated. If nobody takes ownership of the decision, the classification remains a draft, and drafts do not protect the company when someone asks for explanations.

How often should the classification be reviewed?

There is no single frequency for everyone. A review should be carried out when the use, the data processed, the supplier or the way the system operates changes, and in any case at regular intervals decided by the company. For systems that affect people, more frequent checks are advisable. Record the dates of the reviews together with the decisions.

About this content

General informational guide. Regulations and technical requirements change over time: always check the official sources and assess your specific case with a qualified technical or legal advisor.

GET IN TOUCH

Talk to our team.

We'll show how Altovar integrates into your infrastructure. No commitment, no sales pitch. Just a technical conversation.

BOOK A DEMO →