Article · Governance & Risk
Is your AI system high risk? How to classify each use
Boards usually phrase it as “does the AI regulation apply to us”. That is not the question. It applies. What is still open is which category each live use falls into, because everything else hangs off that answer. And the first thing an authority asks for is not your classification. It is how you reached it.
A note on method. In the EU AI Act and AI agents I describe the framework in general terms and say there that the category has to be settled case by case. This piece is that case by case. I go down to the specific criteria because that is how the regulation is drafted, and I checked them in the consolidated version the Publications Office of the Union puts out, which I do not link for two reasons: it is not one of the sources this site cites, and it warns on its own front matter that it is a documentation instrument with no legal effect, because the authentic versions are the Official Journal ones. I still avoid citing article numbers: for the exact wording there is the regulation.
And one currency warning that changes how everything below should be read. The regulation was amended in July 2026, and with that amendment the high-risk classification rules and the list of areas are not applicable law yet: they arrive on 2 December 2027 for the listed uses and on 2 August 2028 for systems inside a regulated product. The full calendar is in the AI Digital Omnibus. What does fall on you today is knowing which category each use sits in, because without that answer you cannot even claim the delay reaches you, and because the file that spares you the 2027 rush gets written now. I add what I flag in the roles piece too: the official article-by-article explorer can lag behind the amended text.
The two doors into the high-risk category
The regulation does not classify systems. It classifies uses, and there are only two doors into the high category.
The first is the regulated product door, and it takes two things at once. One, that the system is intended to be used as a safety component of a product covered by the Union harmonisation legislation the regulation itself lists, or that it is one of those products. Two, that the product has to undergo a third-party conformity assessment before being placed on the market. Both, not either. This door mostly concerns whoever manufactures machinery, lifts, medical devices or toys, and they recognise it immediately because they already live inside a conformity file.
The July 2026 amendment narrowed that door, which is worth knowing if you manufacture. AI systems used solely for non-safety-related aspects in the areas of user assistance, performance optimisation, service efficiency, automation, convenience or quality control do not qualify as safety components. The carve-out is that systems whose failure or malfunction can endanger health and safety do qualify, so the test comes back to what happens when it breaks. And a product required to undergo a third-party conformity assessment solely because of risks other than health and safety, such as radio spectrum or electromagnetic interference that does not affect them, fails the second condition.
The second is the list of areas. In addition to the above, systems falling within the areas the regulation lists in its annex are considered high risk, and there are eight:
- biometrics, in so far as their use is permitted
- critical infrastructure
- education and vocational training
- employment, workers’ management and access to self-employment
- access to and enjoyment of essential private services and essential public services and benefits
- law enforcement
- migration, asylum and border control management
- administration of justice and democratic processes
Eight areas, not eight use cases. Within each one the regulation lists specific situations, and being in the sector is not enough: what counts is whether your use matches one of them. Seven of the eight areas are broken down that way. Critical infrastructure is exhausted by its own heading.
For a company in the bracket I work with, the first door almost never opens, and only three of those areas show up in practice. Employment and workers’ management, where any agent touching hiring, promotion or shift allocation lands. Access to essential private services, which is how customer creditworthiness assessment gets in. And education and vocational training, if you sell or deliver training. The rest of the back-office catalogue I deploy, reconciling invoices, classifying orders, assembling case files, comes through neither door.
The back door: on the list and still not high risk
This is where most of the doubtful cases get settled, and it is the part almost nobody shows. A use inside one of those eight areas is not considered high risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision-making.
That exemption opens if any one of four conditions is met, and one is enough:
- the system is intended to perform a narrow procedural task
- it is intended to improve the result of a previously completed human activity
- it is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review
- it is intended to perform a preparatory task to an assessment relevant to the listed use cases
And there is a cap that overrides all of it. A system in the listed areas is always considered high risk where it performs profiling of natural persons. That sentence ends the discussion before it starts, and it is the first thing to check.
Translated into what actually gets argued in a room. An agent that receives CVs, deduplicates them, extracts fields and leaves them tidy for a hiring manager to read has a solid preparatory-task argument. The same agent scoring candidates does not. The difference is not the model. It is whether a person afterwards decides by looking at something other than the system’s output.
The duty to document it is written for the provider
This is the asymmetry that causes most of the confusion in boardrooms, and it is worth stating precisely. Whoever considers that a system in the listed areas is not high risk shall document that assessment before the system is placed on the market or put into service, is subject to the registration obligation the regulation sets out, and shall provide that documentation to national competent authorities on request. That written duty sits with the provider.
The reading is mine and I give it as such: the literal duty not being yours leaves you no better off. If you are the deploying company, nothing on that route forces you to write the assessment, but you will not be able to hold the line in front of an authority that your use is not high risk without the same piece of paper. And three moves turn you into a provider without your deciding to, one of them changing the intended purpose of the system, which is exactly what happens when a tool bought for one thing ends up used for another. Which ones they are and what they drag along is in provider or deployer. The day you cross one, documenting stops being good practice and becomes literally your duty, with the file that should have existed from the start.
What the classification record has to say
The file fits on a one-page record per use that anyone on your board can read and sign. No legal opinion needed. What it does have to contain:
- The system and its version, and who provides it.
- The declared intended purpose, written in business terms rather than brochure terms.
- What it decides and about whom: whether the output affects an identifiable person, and in what.
- What happens if it gets it wrong, in one sentence, with the realistic worst case.
- The door examined and the result: regulated product, listed area, or neither.
- If the exemption is invoked, which of the four conditions and why, with the concrete argument.
- Whether there is profiling, answered before anything else.
- Date, author and who signs.
- The review trigger: which change forces the record open again.
That last point separates a living file from one for show. Classification expires on its own. The intended purpose changes, a new data source gets connected, the assistant that used to suggest starts ranking, and the use is now a different one even though the contract still says the same thing. The record has to name those cases. And the inventory where all the records live is what shadow AI breaks: you cannot classify what you do not know is running.
What this record does not settle
Four boundaries, because all four get mixed up in the same meeting. It is not the data protection impact assessment: controller and processor are roles from another regulation answering another question, the one about who decides over personal data, and I cover them in AI agents and data privacy. It is not a certificate: classifying well and holding an accredited badge are different things, as I separate in do you need ISO 42001 certification?. It does not tell you from when whatever comes out of the record binds you, which is calendar and sits in the Omnibus piece. And it does not cover the literacy of the people working with the systems, the one duty this matrix does not settle because it looks at people rather than the system, which I develop in the AI literacy obligation.
What the record does do is turn regulatory non-compliance, one of the five risks I order in AI risks for business, into something you can show. The case-by-case evidence that goes with it is in auditing AI agent decisions. The governance framework it all fits inside, in AI agent governance, and the way into the subject, in AI agents for business.
One last warning on method, because I have seen more classifications fail on this than on misreading the rules. The record is written against the real purpose, the one the use has in your operation today, not the one in the procurement brief it was bought under. An agent bought as a reading assistant and used to rank candidates is, for classification purposes, an agent that ranks candidates.
Frequently Asked Questions
Is an agent that reconciles invoices high risk?
On its own, no: reconciling invoices comes through neither the regulated product door nor any of the eight listed areas. But that has to be written down rather than assumed, because the authority’s question is not which category you are in, it is how you determined it. A one-page record with the intended purpose, what it decides, about whom, and the door examined settles that case in ten minutes.
Who signs the classification in a mid-sized company?
Whoever answers for the use, not whoever built it. In the operations I have seen work, the process owner signs it, supported by whoever runs compliance if that role exists, and it gets reviewed in the same committee that approves switching the agent on. Leaving it to the vendor is comfortable and is exactly what does not survive an inspection.
My vendor says their system is not high risk. Is that enough?
It works as a starting point and as documentation you can ask them for, because the written duty to document that assessment is theirs. It does not work as the final answer, because the category depends on the use you put it to, and you define that use. If you employ it for something other than the intended purpose they declared, their assessment stops describing your case.
How often should the classification be reviewed?
On the calendar, once a year is enough in most cases. On events, whenever the intended purpose changes, a new data source is connected, the agent’s write permissions are widened, or the output moves from suggesting to ranking. The event trigger is the one that actually protects you. The annual one only catches what slipped through.