Qualify an AI system

⚠ Assembling open-source building blocks and putting the result into service under your own name makes THE COMPANY the PROVIDER.

Why: A provider established outside the EU must appoint an EU authorised representative before placing a high-risk system on the market (art.22). This is asked of every product, including in-house builds, because THE COMPANY entity outside the EU is itself a non-EU provider.

Why: A feature is a separate AI system when it has its own intended purpose. Features that share one purpose and differ only in screens are a single AI system. Each system listed here needs its own Block S and becomes its own register row.

Example — MUSE: (1) CV to Technical Document, (2) Speech to Text / profile need drafting, (3) Consultant-to-mission matching. Three systems, three rows.

✅ “Extracts structured fields from a CV and pushes them into the recruitment software.” ❌ “AI-based recruitment tool.”

(a) INTENDED PURPOSE AS STATED — copy or summarise what the vendor or the build spec says it is for:

(b) ACTUAL USE at THE COMPANY — what do you concretely use it for? If (b) differs from (a), say how:

⚠ A divergence between (a) and (b) can make THE COMPANY the PROVIDER under art.25(1)(c) and can change the classification. Do not smooth it over.

Why: Drives the art.50(2) marking deadline: a generative system placed before 2 August 2026 has until 2 December 2026 to be marked. Nothing else in art.50 has a grace period.

HR / recruitment, finance, health, education, critical infrastructure, legal, general business support…

✅ Predicts which projects will drift · generates a draft reply · recommends a candidate · auto-approves an expense claim.

❌ Displays a dashboard of last quarter · stores documents · computes a payroll total.

✅ Learns or applies a model: trained on past CVs · anything calling GPT, Claude or Mistral · a scoring model fitted to historical data.

❌ Rules a person wrote: an Excel formula · an if-then workflow in a ticketing tool · a fixed threshold alert.

A screening model runs on a server in India, operated by the Indian entity, and scores candidates who will work at THE COMPANY France. The output is used in the EU → IN SCOPE, despite an entirely non-EU technical set-up.

✅ Covered by the research exemption: a model being trained in a lab, never run on real candidates.

❌ Not covered: a pilot running on real applicants “just to test” — testing in real-world conditions is in scope.

⚠ OPEN SOURCE: under art.2(12) the exemption falls away where the system is placed on the market or put into service as a high-risk system, or as a system falling under art.5 or art.50.

Sample: “Deployed and used entirely by the Singapore entity for local clients; no output is used in the Union and no EU staff are affected.”

⚠ (f) ❌ A video-conferencing tool that reads engagement or mood from faces. ❌ A system scoring an agent’s “empathy” from voice tone. ✅ Sentiment analysis of written TEXT (not biometrics). ✅ Drowsiness detection for driving safety.

⚠ (c) Assessing creditworthiness for a specific financial product is NOT social scoring. Social scoring is a general reputation score reused in UNRELATED contexts.

Why: Added by Regulation (EU) 2026/1744. These two prohibitions have their own application date: 2 December 2026, not 2 February 2025. The register dates them separately.

✅ Yes: an image generator with no filter that will produce nude images of a named person on request.

❌ No: a text summariser · a code assistant · an image tool with refusal training and output filtering in place.

Why: Tick Yes where generation of such material is an intended purpose, OR where the outcome is reasonably foreseeable and reproducible without significant technical modification and safeguards are absent. For a deployer the test is actual use.

⚠ For a PROVIDER the test is not intent. Art.5(1a) catches a system whose design, training, architecture or user-facing functionality makes the outcome reasonably foreseeable and reproducible, where reasonable and adequate safeguards are absent. Safeguards that count: data cleaning, refusal training, safe prompt design, output controls, runtime guardrails, content classification and filtering, usage restrictions, abuse detection, notice-and-action. For a DEPLOYER the test is actual use.

Sample: “The tool infers engagement from webcam video during internal training sessions. This is emotion inference in an educational context and falls under art.5(1)(f). Use suspended pending Compliance review.”

Why: This is the only role question everyone answers. The four that follow appear only for the scenarios that need them.

Why: Provider and deployer are cumulative roles, not alternatives. Both obligation sets apply: art.9-17 AND art.26 — including art.26(7), information of workers’ representatives, which in France also entails consulting the CSE.

Why: Yes = THE COMPANY is also the deployer. No = THE COMPANY only supplies it to others.

Why: If the register returns Importer or Distributor, it raises the art.23 or art.24 verification duties for you in block 4. There is nothing further to tick here.

Medical devices, in vitro diagnostics, vehicles, aviation, rail, lifts, toys, pressure equipment, personal protective equipment.

⚠ Machinery is no longer in Annex I Section A — Regulation (EU) 2026/1744 moved it. Answer Q5.4.

⚠ BIOMETRICS: Annex III point 1(a) covers remote IDENTIFICATION. A system whose sole purpose is verifying that a person is who they claim to be (1:1) is excluded. Tick the excluded option only where that is genuinely the sole purpose.

⚠ EMPLOYMENT covers more than recruitment: ALLOCATION OF TASKS, MONITORING and EVALUATION of performance and behaviour. A tool that assigns tickets to engineers based on predicted throughput is in scope. So is a tool that flags “underperforming” consultants.

Why: Annex III is a CLOSED LIST of specific use cases grouped under headed areas, and art.6(2) makes a system high-risk only where it is intended for one of the listed uses. Operating in an area does not by itself qualify: an HR chatbot that answers policy questions is in "employment and workers' management" and does none of the things Annex III point 4 lists. The three specific biometric options at Q6.1 and Q6.2 already name an Annex III point and need no answer here; 1:1 verification is expressly excluded from point 1(a).

Why: Art.50(1), owed by the PROVIDER. Applies from 2 August 2026. No grace period.

✅ Yes: a chatbot, a voice server, an assistant people type at.

❌ No: a batch model that scores records overnight, with no conversation.

Why: Art.50(2), owed by the PROVIDER. This is the ONLY limb that carries a grace period: a system placed on the market before 2 August 2026 has until 2 December 2026 to be marked (art.111(4)).

✅ Yes: writes text, drafts documents, makes images, synthesises voice.

❌ No: classifies, scores or extracts data without producing new content.

Why: Art.50(3), owed by the DEPLOYER. Applies from 2 August 2026.

✅ Yes: reads faces or voice for mood · sorts people by biometric traits.

❌ No: sentiment analysis of written text — that is not biometric.

Why: Art.50(4), first subparagraph. A deployer duty. Applies from 2 August 2026.

✅ Yes: a synthetic video of a real executive · a cloned voice · a photorealistic image of a product or document that never existed.

❌ No: a wholly invented illustration of nobody in particular · an obviously cartoon or stylised image · content that defies the laws of physics.

Why: Art.50(4), second subparagraph. A separate duty with a different exception from the deep-fake limb.

✅ Yes: an AI-written article published on our site about regulation or public affairs.

❌ No: internal documentation · marketing copy about our own products · a proposal to one client.

✅ Yes: anything built on GPT, Claude, Gemini, Llama or Mistral — including through a vendor’s product.

❌ No: a small model trained only on our own data for one narrow task.

CVs, emails, user IDs, IP addresses, voice recordings, photographs, login histories, employee appraisals.

Why: GDPR art.3(1). If yes, the GDPR applies — wherever the servers are, wherever the data sits, whoever the people are. Stop here.

✅ Yes: THE COMPANY France runs the tool. It makes no difference that the vendor is American or the hosting is in Ireland.

❌ No: the Singapore entity runs it, for Singapore clients, on its own account.

Why: Drives art.26(7), information of workers’ representatives and affected workers. In France this also entails consulting the CSE — an obligation outside the AI Act.

Why: GDPR art.22 is assessed on its own facts, not on the AI Act risk level. A solely automated decision inside a minimal-risk system is still caught.

✅ Yes: the tool rejects an application and nobody looks at it.

❌ No: it ranks candidates and a recruiter decides.

Sample: “The vendor contract expires in March 2027 and a switch of supplier is under consideration: the qualification will have to be reviewed then.”