How AnnexWise derives its verdicts

Every classification is produced by a deterministic rules engine that encodes Regulation (EU) 2024/1689, as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744), article by article. No language model sits between your answers and your verdict: the same answers always produce the same result, and every rule below is traceable to the article it implements.

Classification order

1. Article 5 screen. Any selected prohibited practice classifies the system as prohibited — nothing else is evaluated for tiering, and the €35M / 7% fine tier applies.
2. Article 6(1): regulated-product safety components are high-risk only where both limbs are met — the system is a safety component of (or is itself) a product covered by the Annex I harmonisation legislation, and that product is required to undergo third-party conformity assessment. A self-assessed product (a Class I medical device, most toys, machinery outside Annex IV of the Machinery Regulation) does not take this route. The Article 6(3) derogation does not apply to systems that do.
3. Article 6(2) + Annex III. A match with any Annex III category is high-risk, unless the Article 6(3) derogation is claimed — in which case the system is classified out of high-risk and a counsel-review flag is attached, because that derogation is a documented-judgment call, not a checkbox.
4. Article 50 transparency triggers attach duties regardless of high-risk status; with no high-risk match they classify the system as limited risk. One deliberate exception: for prohibited systems, tiering stops and no Article 50 obligations are listed — deployment must stop, so there is nothing left to make transparent.
5. GPAI (Articles 53/55) duties attach orthogonally when you provide a general-purpose model; systemic-risk scale adds the Article 55 regime.
6. Anything else is minimal risk (Article 4 AI literacy still applies).

Rule inventory

Prohibited practices — Article 5(1)

• Subliminal or purposefully manipulative techniques — Interfaces designed to distort behaviour and cause significant harm
• Exploiting vulnerabilities (age, disability, social or economic situation) — Targeting children or financially distressed users to their detriment
• Social scoring by general behaviour or characteristics — Scoring citizens/customers leading to unrelated detrimental treatment
• Predicting criminal offences from profiling alone — Risk-of-crime scores based solely on profiling or personality traits — systems supporting a human assessment grounded in objective, verifiable facts are excluded
• Untargeted scraping of facial images — Building face databases from internet or CCTV footage
• Emotion recognition in workplaces or education — Inferring emotions of employees, job candidates or students — e.g. video-interview 'confidence scores' (outside medical/safety uses)
• Biometric categorisation of sensitive attributes — Inferring race, political views, religion or orientation from biometrics — labelling/filtering of lawfully acquired datasets in law enforcement is excluded
• Real-time remote biometric ID in public spaces (law enforcement) — Live facial recognition in publicly accessible spaces — save the three narrowly authorised law-enforcement objectives of Art. 5(2)–(7)
• Generating non-consensual intimate imagery of identifiable persons (prohibited from 2 Dec 2026) — “Nudifier” functionality, or image generation where such misuse is reasonably foreseeable and reproducible in the absence of reasonable, proportionate and effective safeguards — added to Article 5 by the 2026 Omnibus

High-risk categories — Annex III

• Biometric identification or categorisation — Remote biometric ID, emotion recognition (where not prohibited)
• Safety component in critical infrastructure — AI managing energy grids, water, traffic, digital infrastructure
• Education and vocational training — Admissions, exam scoring, proctoring, learning-outcome evaluation
• Employment and worker management — CV screening, interview scoring, promotion/termination decisions
• Access to essential services — Credit scoring, insurance pricing (life/health), benefits eligibility
• Law enforcement — Evidence reliability assessment, recidivism risk, polygraph-type tools
• Migration, asylum and border control — Visa/asylum application assessment, border risk assessment
• Administration of justice and democratic processes — Assisting courts in interpreting facts/law; influencing elections

Transparency triggers — Article 50

• People interact directly with the AI (chatbot, voice agent)
• Generates synthetic audio, image, video or text content
• Generates or manipulates deepfake content
• Uses emotion recognition or biometric categorisation (permitted cases)

Scoring

The compliance score is the weight of satisfied obligations divided by the weight of all applicable obligations, on a 0–100 scale. Weights reflect severity: critical items (e.g. conformity assessment, technical documentation, human oversight) weigh roughly twice as much as procedural ones. A prohibited classification scores 0 by definition.

Verification & change control

• Every rule is covered by a golden test suite — canonical scenarios with expected verdicts that run on every change to the engine.
• Any change to classification logic, obligations or weights bumps the rule-set version. Reports permanently record the version and an input fingerprint, so any verdict can be re-derived and audited later.
• The rule set tracks amending regulations, Commission guidelines, AI Office templates and harmonised standards; the legal-state date above tells you exactly how current your verdict is, and the changelog below records every legal change we have incorporated.
• Genuinely gray areas (e.g. the Article 6(3) derogation) are surfaced as counsel-review flags rather than silently resolved.
• Verdicts are living: an assessment saved under an older rule-set version flags itself in your inventory, and one click re-runs the same answers under the current law with a per-obligation diff — deadline moves, new duties, new flags — before anything is overwritten.

Application timeline the engine applies

Deadlines quoted in verdicts and reports follow the timeline as amended by Regulation (EU) 2026/1744 (in force 27 July 2026), which postponed the high-risk regime and added a new Article 5 prohibition. These are fixed calendar dates — the proposed standards-availability trigger was dropped by the co-legislators.

Rule-set changelog

Every change in the law that this engine incorporates is recorded here, permanently and publicly, with the rule-set version that shipped it. If the law moves and this log has not, that is a bug — tell us: support@annexwise.com.

AnnexWise is compliance software, not a law firm. This page documents how the software reasons so that your counsel can audit it; it does not replace their judgment.