What the self-assessment tool really does, and where its classification stops
The EU AI Act Self-Assessment tool is a guided procedure in five steps that, from the answers you give about an AI system, returns a risk class, a list of compliance gaps with article and deadline, and a report to download. It is an orientation aid, not a legal classification: its rules are a simplification of the AI Act, and on this page I read them next to the text of the Regulation to show you where they coincide and where they diverge. The backend code documents it as the "Audit" tier at €15 a month (comment of the controller and of migration V063); this is not a price list, and the current price is read where the subscription is bought.
A note on the dates the tool shows: high risk under Annex III applies from 2 December 2027 and under Annex I from 2 August 2028, under Regulation (EU) 2026/1744, which moved the original dates; transparency under Article 50 stays at 2 August 2026 and general-purpose models at 2 August 2025. When a recommendation sentence of the tool cites a date or an article different from those on this page, this page prevails and, above it, the consolidated text of the Regulation. This is not legal advice.
The five steps, as they are built
I read the code of the page and of its components on 13 November 2026. Each step collects a piece of the profile and none is optional for reaching the risk class: the table says what it asks and what comes out of it.
| Step | What it asks | What it produces |
|---|---|---|
| 1. Identification | Name and description of the system | Label of the report |
| 2. Purpose | Intended purpose (classification, recommendation, scoring, generation, translation, detection, prediction, automation), sector, area of deployment (EU, extra-EU, global) | The inputs of the classifier |
| 3. Classification | Areas of Annex III (the eight), level of autonomy, effect on fundamental rights, special categories of data, human oversight, whether it is a general-purpose model and with systemic risk | Risk class: unacceptable, high, limited, minimal |
| 4. Gaps | Nothing: it uses the previous answers | List of gaps with article, severity and deadline |
| 5. Report | Optional: an email address | Downloadable .txt file |
The structure is sensible: the eight areas of step 3 correspond to the macro-areas of Annex III (biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, justice). The limit is that Annex III reasons by use cases ("evaluate the creditworthiness of natural persons"), not by sector, and step 3 asks the user to say for themselves whether their system falls within an area: the correctness of the answer, and therefore of the class, depends on that question. The use-case procedure is in the article on Annex III.
How it classifies: the function, executed
The classifier is a function of cascading rules, with the precedence "unacceptable, high, limited, minimal". If the call to the backend fails, for example because there is no active subscription (the controller answers 402), the page falls back on that function, so it is the behaviour everyone sees. The listing below is an extract of the repository file, reproduced without changes, plus a short program that applies it to five cases; I ran it with Node 24, which executes TypeScript by stripping the types, and the output is identical in two runs.
In the code, the block to watch is the last one before the "limited" branch, the one with scoring or classification combined with finance, healthcare or hr: it is a heuristic that classifies as high risk any scoring or classification in three sectors, even without an Annex III area selected. Remove that block and only the following would remain high-risk: systems with a declared area, those that affect fundamental rights with full autonomy and those that process special categories of data without human oversight. The comment // Art. 5 at the head of the first two blocks should not suggest a literal reading of the article: it is an approximation, as the output below shows.
// Extract from frontend/src/app/pages/dev-tools/eu-ai-act-self-assessment/services/eu-ai-act-assessment.service.ts
type RiskClass = 'unacceptable' | 'high-risk' | 'limited-risk' | 'minimal-risk';
type Richiesta = {
sector: string; intendedPurpose: string; deploymentRegion: string; autonomyLevel: string;
annexIIICategories: string[]; affectsFundamentalRights: boolean;
processesSpecialCategoryData: boolean; hasHumanOversight: boolean;
};
function deriveRiskClass(req: Richiesta): RiskClass {
// Art. 5 - prohibited practices
if (
req.sector === 'law-enforcement' &&
req.annexIIICategories.includes('biometric') &&
req.autonomyLevel === 'fully-autonomous'
) {
return 'unacceptable';
}
if (
req.annexIIICategories.includes('biometric') &&
req.deploymentRegion !== 'extra-eu' &&
req.autonomyLevel === 'fully-autonomous'
) {
return 'unacceptable';
}
// Annex III - high-risk
if (req.annexIIICategories.length > 0) {
return 'high-risk';
}
if (req.affectsFundamentalRights && req.autonomyLevel === 'fully-autonomous') {
return 'high-risk';
}
if (req.processesSpecialCategoryData && !req.hasHumanOversight) {
return 'high-risk';
}
if (
(req.intendedPurpose === 'scoring' || req.intendedPurpose === 'classification') &&
(req.sector === 'finance' || req.sector === 'healthcare' || req.sector === 'hr')
) {
return 'high-risk';
}
// Art. 50 - limited risk (transparency obligations)
if (
req.intendedPurpose === 'generation' ||
req.intendedPurpose === 'translation' ||
req.intendedPurpose === 'recommendation'
) {
return 'limited-risk';
}
return 'minimal-risk';
}
const base: Richiesta = {
sector: 'other', intendedPurpose: 'other', deploymentRegion: 'eu', autonomyLevel: 'human-in-loop',
annexIIICategories: [], affectsFundamentalRights: false,
processesSpecialCategoryData: false, hasHumanOversight: true,
};
const casi: [string, Partial<Richiesta>][] = [
['FAQ assistant that generates text', { intendedPurpose: 'generation' }],
['customer scoring in an online shop', { intendedPurpose: 'scoring', sector: 'retail' }],
['fraud detection on payment cards', { intendedPurpose: 'classification', sector: 'finance' }],
['CV screening (employment area)', { annexIIICategories: ['employment'], sector: 'hr' }],
['medical image classification model, area not flagged', { intendedPurpose: 'classification', sector: 'healthcare' }],
];
for (const [nome, delta] of casi) {
console.log(nome.padEnd(62), '->', deriveRiskClass({ ...base, ...delta }));
}
// Output:
// FAQ assistant that generates text -> limited-risk
// customer scoring in an online shop -> minimal-risk
// fraud detection on payment cards -> high-risk
// CV screening (employment area) -> high-risk
// medical image classification model, area not flagged -> high-risk
Two cases of the output coincide with the reading of the law. CV screening is high risk: Annex III, point 4, point (a), includes systems for the recruitment or selection of natural persons. The assistant that generates text falls in "limited", which for the Regulation means the transparency obligations of Article 50 (informing whoever interacts, marking the synthetic output). Customer scoring in an online shop is "minimal", plausible if it does not decide on credit or employment.
Two cases diverge, and I declare them because they are the ones where you can go wrong by trusting. Fraud detection on payment cards comes out as "high risk", but Annex III, point 5, point (b), expressly excludes from the systems for evaluating creditworthiness "AI systems used for the purpose of detecting financial fraud". The classification of medical images comes out as "high risk" because of the sector heuristic, whereas in the Regulation a healthcare system is high-risk if it is a safety component of a device governed by Annex I (for example Regulation (EU) 2017/745) and subject to third-party assessment, with the date of 2 August 2028; there is no rule that says "classification + healthcare = high risk". The function errs on the cautious side, but it errs.
The function also lacks three elements that matter: the exception of Article 6(3) (an Annex III system is not high-risk if it does not pose a significant risk of harm, for example because it performs a narrow procedural task or a preparatory task), the opposite rule that it is always high-risk if it performs profiling of natural persons, and the distinction between Annex I and Annex III products. The two "unacceptable" rules (biometrics with full autonomy) are a rough approximation: the Regulation prohibits precise practices (Art. 5(1)), and the biometric systems of Annex III are high-risk, not prohibited in themselves.
The gaps it reports, and the report
Step 4 produces gaps with article, severity and deadline. For a high-risk system it lists, among others, human oversight (Art. 14), the risk management system (Art. 9), technical documentation (Art. 11, Annex IV) and conformity assessment with registration (Art. 43), with the deadline "2 December 2027"; for a system that generates content, the marking obligation of Article 50(2), with the deadline "2 August 2026" and the note on the 2 December 2026 deadline for systems already on the market (Art. 111(4)); for a general-purpose model, the documentation of Article 53 with the deadline "2 August 2025". The articles and dates correspond to the consolidated text I read.
The report downloads as a text file with risk class, gaps and date of generation. In the current code the button to receive it by email is not connected to the backend: the page marks the sending as successful without making a request. Download the file. And do not mistake it for technical documentation: Annex IV asks for nine points, from the general description to the post-market monitoring plan, and a summary of gaps does not contain them. For the real file start from the article on technical documentation.
When it is enough and when it is not
The tool is useful for a first pass: it forces you to answer the right questions in the right order and gives you a list of articles to read. It is not enough when your system sits close to a boundary, like the two cases above, or when the role is uncertain: if you are the one changing the purpose of someone else's system you become the provider (Art. 25), and the tool does not ask about that. For those cases start again from the article on providers and deployers and from the 17-item work list, which lists obligations and dates by role.
Frequently asked questions
Is the class the tool gives me final?
No. It is the result of simplified rules, which err on the cautious side in some cases (fraud, healthcare) and do not cover the exception of Article 6(3). Treat it as a list of articles to read. In case of doubt, the classification is decided by the provider, who documents its assessment (Art. 6(4)), and by the supervisory authorities.
Is it of any use for a system already on the market?
For the first pass, yes. Remember, though, that high-risk systems already on the market before the date of application are subject to the Regulation only if they undergo significant changes in their designs, except those intended to be used by public authorities (Art. 111(2)): the tool does not calculate that.
How much does it cost and how much does it save?
The price is read where the subscription is bought; the code documents it at €15 a month. I have no measure of the time saved or of the cost of a conformity assessment, and the Regulation fixes no costs: no estimate of savings above has a source, which is why I give none.
What you can do now
You know what the five steps ask, you can read the risk class as a first approximation, and you recognise the cases where the function diverges from the law: financial fraud, healthcare, the exception of Article 6, profiling. With the class in hand, the next step is to establish your role and the procedure that applies to you, which you find in conformity assessment. To know whom a high-risk system answers to, see the article on the supervisory authorities.
Sources
- Regulation (EU) 2024/1689, consolidated text of 27.7.2026 - Articles 5, 6, 25, 50, 53, 111 and 113, Annexes III and IV, consulted on 2026-11-13
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ L of 24.7.2026 - dates of Article 113 (2 December 2027 and 2 August 2028), consulted on 2026-11-13
The behaviour of the tool is described by the repository code (assessment service, components of the five steps, backend controller), read on 2026-11-13; the extract was run with Node 24.11.0 with identical output in two runs. I did not run the path through the backend, which requires a subscription. The price is not verified on a sales page.

Commenti
Caricamento commenti...
Sign in to comment