EU AI Act Engineering Control Map
Version: 1.1.0 Last updated: 2026-07-16 Status: Informative OAIES implementation profile
Purpose
Translate Regulation (EU) 2024/1689 obligations into engineering evidence without presenting legal advice.
Why
Applicability depends on role, system, use case, geography, and effective date.
When
Use at intake, material change, and scheduled regulatory review.
How
- Create one applicability record per intended purpose, jurisdiction, value-chain role, and materially distinct model or system release.
- Have qualified counsel decide scope under Article 2, role definitions under Article 3, prohibited-practice exposure under Article 5, and high-risk classification under Article 6 with Annex I or III.
- If the organization provides or modifies a general-purpose AI model, assess Chapter V separately; downstream system classification does not replace GPAI-provider analysis.
- Bind each applicable duty to an engineering owner, immutable evidence, verification cadence, and stop-release condition.
- Re-open applicability when intended purpose, geography, model, fine-tuning, brand, value-chain contract, or substantial-modification facts change.
Legal boundary and effective-date caveat
This document is engineering implementation guidance, not legal advice. It does not determine whether a system is prohibited, high-risk, a GPAI model, or compliant. The accountable legal function must use the Official Journal text, applicable delegated or implementing acts, harmonised standards or common specifications, Commission guidance, and national competent-authority requirements current on the decision date.
As of 2026-07-16, the principal Article 113 timeline is:
| Date | Engineering planning significance | Caveat |
|---|---|---|
| 2024-08-01 | Regulation entered into force | Entry into force is not the same as full applicability |
| 2025-02-02 | Chapters I and II apply, including Article 4 AI-literacy and Article 5 prohibited practices | Confirm workforce and prohibited-use controls now |
| 2025-08-02 | Article 113(b) provisions apply, including GPAI Chapter V and specified governance, notified-body, confidentiality, and penalty provisions | Transitional rules in Article 111 can alter dates for existing systems/models |
| 2026-08-02 | Most remaining provisions apply | This date is near; release plans must work backward from conformity and evidence lead time |
| 2027-08-02 | Article 6(1) classification and corresponding obligations apply | Applies to specified product-safety-component/product categories; verify exact Article 113(c) scope |
The Commission’s AI Act policy page, last updated 2026-07-07, reports a political agreement on an AI Omnibus that would move certain Annex III high-risk dates to 2027-12-02 and high-risk systems embedded in regulated products to 2028-08-02. A political agreement or policy summary is not a substitute for an enacted amending regulation in the Official Journal. Counsel must track the legislative act and use the legally effective text on the decision date.
For GPAI, Commission guidance states: obligations apply from 2025-08-02 for models placed on the market from that date; Commission enforcement powers apply from 2026-08-02; providers of GPAI models placed on the market before 2025-08-02 must comply by 2027-08-02 under the transitional rule. Verify the specific model’s placement date and current guidance.
Do not encode any date in this section as permanent legal logic. Store applicability_reviewed_at, official_journal_version, pending_amendments_reviewed, and next_review_at; require counsel to update the decision when a binding amendment or authoritative interpretation changes.
Operator-role map
One organization can hold several roles for the same value chain. Record each role independently.
| Role | Typical trigger | Primary engineering evidence | Boundary to resolve |
|---|---|---|---|
| Provider | Develops an AI system/model or has it developed and places it on the market or puts it into service under its name or trademark | Intended purpose, QMS, technical file, evaluation, post-market plan, declarations/registration where applicable | White-labelling, fine-tuning, and substantial modification can move a party into provider duties |
| Deployer | Uses an AI system under its authority, other than personal non-professional use | Instructions followed, input-data relevance, human oversight, logs, monitoring, worker/affected-person notices, FRIA decision where applicable | A deployer that substantially modifies or changes intended purpose may become a provider |
| Importer | EU-established party placing a third-country provider’s branded system on the Union market | Provider/conformity verification, representative, documentation availability, storage/transport controls | Contracting through a reseller does not automatically remove importer facts |
| Distributor | Makes an AI system available in the supply chain other than provider/importer | Required markings/document checks, corrective-action and authority-cooperation records | Modification or rebranding may change role |
| Authorised representative | EU-established mandate holder for a non-EU provider | Written mandate, retained documentation, registration and authority-cooperation procedures | Mandate scope must cover the duties actually delegated |
| Product manufacturer | Places a product on market with an AI system under its own name/trademark | Annex I product integration, safety/conformity file, change control | Determine whether the AI is a safety component and which sectoral conformity regime applies |
| GPAI model provider | Develops and places a general-purpose AI model on the Union market, including through an API | Model technical documentation, downstream information, copyright policy, training-content summary | Open-source exemptions are limited; systemic-risk duties and other exceptions require separate analysis |
Applicability workflow
High-risk engineering duty map
Counsel must confirm the exact duty and role. The following map tells engineering what evidence to prepare.
| Provision | Subject | Provider implementation evidence | Deployer implementation evidence |
|---|---|---|---|
| Articles 9 and 17 | Risk management and quality management | Lifecycle risk register, acceptance criteria, change control, supplier controls, competence, corrective action | Local-use risk controls, escalation into provider feedback, operating procedures |
| Article 10 | Data and data governance | Dataset provenance, relevance/representativeness analysis, bias/error examination, preprocessing lineage, access and retention | Input-data relevance and fitness for intended purpose under Article 26 |
| Articles 11 and Annex IV | Technical documentation | Versioned technical file linking intended purpose, architecture, data, metrics, testing, cybersecurity, oversight, changes | Retain provider instructions and local configuration/use evidence |
| Article 12 | Record-keeping | Automatic event-log design, clock synchronization, integrity, retention capabilities | Keep logs under deployer control for the applicable period and purpose |
| Article 13 | Transparency/instructions | Capability, limitation, accuracy, foreseeable misuse, input specification, oversight and maintenance instructions | Operator training and evidence that instructions are followed |
| Article 14 | Human oversight | Competence, authority, interface, stop/override, automation-bias mitigation, two-person verification where specified | Named overseers, staffing, escalation, override and non-retaliation procedures |
| Article 15 | Accuracy, robustness, cybersecurity | Declared metrics/slices, fault tolerance, poisoning/injection/adversarial tests, secure update and dependency controls | Monitor local performance and stop use outside declared operating conditions |
| Articles 43, 47–49 | Conformity, declaration, marking, registration | Applicable assessment route, EU declaration, CE marking and database registration evidence | Verify required system/provider status before use |
| Articles 72–73 | Post-market and serious incidents | Monitoring plan, complaint/telemetry intake, trend analysis, authority-notification runbook | Immediate provider/authority escalation path where applicable |
For Annex III high-risk uses, explicitly evaluate Article 6 exclusions and profiling rules with counsel. For public-law deployers or bodies providing public services, assess the Article 27 fundamental-rights impact assessment before first use; coordinate, but do not conflate, it with a GDPR DPIA.
GPAI provider profile
| Provision | Engineering control | Required artifact |
|---|---|---|
| Article 53(1)(a), Annex XI | Maintain model technical documentation | Architecture/capability description, training and testing process, evaluation results, energy/resource information where required, change history |
| Article 53(1)(b), Annex XII | Give downstream providers sufficient information | Versioned integration pack covering capabilities, limits, input/output modality, acceptable use, evaluation, and system-level risk dependencies |
| Article 53(1)(c) | Copyright-law compliance policy | Approved text-and-data-mining opt-out handling, data acquisition controls, complaint/escalation process |
| Article 53(1)(d) | Public training-content summary | Published summary generated under the applicable AI Office template; legal review of disclosure and trade-secret boundaries |
| Article 54 | EU authorised representative where required | Mandate, contact, retained documentation and authority-cooperation runbook |
| Articles 51 and 55 | Systemic-risk classification and duties | Classification basis, standardized evaluations/adversarial testing, systemic-risk assessment/mitigation, serious-incident reporting, cybersecurity evidence |
Do not use the 10^25 FLOP presumption in Article 51 as the only systemic-risk test. Record Commission designation, capability evidence, and any delegated amendment to thresholds or criteria.
Release gates
- EA-01 Applicability: counsel-signed role, territory, intended-purpose, Article/Annex, transitional-rule, and review-date record.
- EA-02 Prohibited practices: no unresolved Article 5 scenario; any uncertainty blocks launch.
- EA-03 Classification: high-risk/GPAI/systemic-risk decision linked to the exact model and system release.
- EA-04 Evidence: each applicable duty has an owner, artifact, verification result, and retention location.
- EA-05 Human oversight: overseers demonstrate authority, competence, stop/override access, and workload capacity.
- EA-06 Post-market: complaints, drift, serious-incident escalation, provider notification, and corrective-action loops pass a tabletop exercise.
- EA-07 Change control: substantial-modification and intended-purpose triggers are implemented in release intake.
Evidence contract
The decision record is the EU AI Act applicability and evidence file. It records provider/deployer/importer/distributor/GPAI roles; territory; intended purpose; Article/Annex basis; dates; controls; counsel decision. The EU AI Act compliance owner owns completeness; the evidence is invalid when intended purpose, model role, or value-chain contract changes. Preserve legal/applicability decisions, system and data facts, contracts, evidence versions, approvals, and next-review triggers in the system record.
Failure response and recovery
Trigger: a prohibited-practice signal or missed high-risk duty is found.
Immediate response: stop the affected use, preserve logs, notify legal/compliance, and apply required incident/conformity processes. Preserve the EU AI Act applicability and evidence file, affected trace IDs, timestamps, and decision logs before mutation. Open an incident when users, data, money, authorization, or a release decision may have been affected; closure requires a regression case and verified control change specific to eu ai act engineering control map.
Decision authority
The EU AI Act compliance owner accepts the operational decision. The qualified legal counsel provides independent challenge for high-risk scope, failed gates, or exceptions. Policy engines can enforce approved boundaries; accountable operators, counsel, privacy, risk, and business owners decide applicability and residual risk.
Tradeoffs
| Choice | Benefit | Cost |
|---|---|---|
| Conservative scope | Lower compliance risk | More controls |
Anti-patterns
- Calling every AI system high-risk.
- Claiming compliance from a checklist alone.
Enterprise considerations
- Legal owns interpretation; engineering owns evidence.
- Do not claim conformity or certification without formal assessment.
Framework relationship
For EU AI Act Engineering Control Map, this profile translates governance questions into engineering records. Qualified authorities still decide legal duties, conformity, and risk acceptance.
| Source | Relationship for EU AI Act Engineering Control Map | Boundary |
|---|---|---|
| NIST AI RMF | GOVERN 1.1 and MAP 1.5 | Framework mappings organize evidence but do not determine legal obligations. |
| ISO/IEC 42001 | 42001 may support evidence but does not establish EU conformity | Certification, where pursued, is scoped to a management system and does not certify each AI output or legal compliance. |
| Domain threat/control source | Regulation-specific security duties remain in Articles 15 and 55 | Test only the threats applicable to the documented system and release |
Checklist
- Article 2 scope and every Article 3 operator role are recorded by legal entity and release.
- Article 5 prohibited-practice review is signed and unresolved uncertainty blocks deployment.
- Article 6 and Annex I/III classification basis includes exclusions and intended-purpose facts.
- GPAI provider and Article 51 systemic-risk analyses are separate from system classification.
- Article 113 dates and Article 111 transitional rules are reviewed against the current official text.
- Applicable Articles 9–15 evidence is linked to owners and immutable release identifiers.
- Conformity, registration, transparency, FRIA, and deployer duties are dispositioned.
- Serious-incident, corrective-action, and authority/provider notification paths are exercised.
- No statement claims compliance, conformity, certification, or legal sufficiency solely from this profile.
References
- European Union, Regulation (EU) 2024/1689, Official Journal text (accessed 2026-07-16).
- European Commission, Regulatory framework for AI (accessed 2026-07-16). Use guidance only with the binding Regulation and current implementing/delegated acts.
- European Commission, Guidelines for providers of general-purpose AI models (accessed 2026-07-16).
- NIST, AI RMF 1.0, GOVERN and MAP (accessed 2026-07-16); cross-framework implementation aid only.
- ISO, ISO/IEC 42001:2023 (accessed 2026-07-16); certification does not establish EU AI Act conformity.
Changelog
| Version | Date | Change |
|---|---|---|
| 1.1.0 | 2026-07-16 | Added operator-role, Article/Annex, effective-date, high-risk, GPAI, systemic-risk, evidence, and release-gate implementation maps with explicit legal boundaries. |
| 1.0.0 | 2026-07-16 | Initial complete profile. |