Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice
EU AI Act High-Risk Decision Systems: The December 2, 2027 Deployer Duties
The short answer: The EU AI Act's Annex III high-risk list, as carried in the kit's checker data, covers employment and worker management, access to education and vocational training, creditworthiness evaluation, and access to essential public and private services. The high-risk deployer duties for those systems — use the system as instructed, keep human oversight, keep the logs — were postponed by the Digital Omnibus (Regulation (EU) 2026/1744) and apply from December 2, 2027. Article 50's transparency duties already apply since August 2, 2026. December 2027 is therefore the second EU clock: later than the US January start, close enough to plan in the same programme.
Check your duties and dates — free Build your first notice — free
What is an Annex III high-risk system, in this guide? It is an AI system whose use falls in one of the Annex III areas the checker names — the decision areas where the EU places its strongest deployer obligations. This guide restates the kit's checker data on the Annex III timeline and the three deployer duties it names; the fuller EU regime beyond that data is outside the kit and outside this guide.
The three deployer duties the checker names
The checker summarises the high-risk deployer position in one parenthesis: use the system as instructed, keep human oversight, keep the logs. Each maps onto a discipline US-state deployers will recognise. Use as instructed is the developer-documentation discipline — the system's instructions bound its appropriate use, as in Colorado's developer documentation duty. Human oversight is the meaningful-review standard in another vocabulary. Logs are the record duty: the sequence of what the system did and what the business said, retained and producible.
A deployer already running the kit's US workflows — assessments informed by vendor documentation, a staffed review queue, a hash-chained record — has built the operational substance of all three. The EU work is then alignment and evidence, not a new operating model. That convergence is worth stating to leadership early, because it converts December 2027 from a second compliance programme into an extension of the first.
Which of your systems are Annex III: the area test
Run the area test per system, not per company: does this system operate in employment or worker management (screening, allocation, monitoring-adjacent decisions), education access or vocational training (admissions, placement), creditworthiness evaluation, or essential public and private services? A provider with systems in two areas has two Annex III positions; a general business tool outside the areas is outside this timeline (though Article 50 may still touch it).
Record the mapping with the inventory, exactly as the US listing analysis is recorded per decision. The sector guides in this series cover the US side of the same systems — employment, education, credit — and the EU paragraph belongs on the same register row as the US duty set.
Worked example: Meridian Workforce Platforms
Example — fictional business, real duty structure. Meridian is a fictional platform providing worker-management tooling used by employers, with EU and US operations. Its two-clock plan, set in early 2027:
- Area mapping: of 5 systems in the register, 2 are mapped to Annex III (worker management; a creditworthiness-adjacent screening product), 3 recorded as outside the listed areas with reasoning.
- Article 50 (already live): interaction disclosures and content marking verified across the platform — 100% of user-facing AI surfaces audited, marking survival through exports tested.
- US build reused: the January 2027 US programme delivers the register, vendor documentation files, the staffed review queue, and the hash-chained log — each asset tagged for its Annex III use at creation, not retrofitted.
- Oversight evidence: by December 2027 the review queue will hold 11 months of oversight records — requests, full-case reviews, changed outcomes — producible as the oversight artefact.
- December readiness gate: a November 2027 checkpoint verifies, per Annex III system: instructions on file and followed, oversight staffed and evidenced, logs complete and exportable — the three named duties, each with an owner.
The postponement, stated precisely
The checker data attributes the December 2, 2027 date to the Digital Omnibus on AI (Regulation (EU) 2026/1744), which postponed the Annex III high-risk obligations from their original schedule. Precision about the mechanism matters for planning credibility: the duties are postponed, not removed; the date is December 2, 2027, not 'late 2027'; and the postponement covers the Annex III obligations, while Article 50's transparency duties already apply on their own earlier schedule (August 2, 2026, with the marking grace period to December 2, 2026 for earlier-placed systems).
Planning errors cluster around postponements: teams read 'postponed' as 'cancelled', or move the whole EU file to the new date including the parts already in force. The two-clock discipline prevents both — Article 50 live now, Annex III diarised — and the kit's tracker approach keeps the clocks in one view alongside the US dates on the deadline tracker.
Human oversight and logs: where the evidence will come from
When December 2027 arrives, the oversight duty will be evidenced the way meaningful review is evidenced in the US: named reviewers, full-case information, authority, and records showing engagement. A business that has run a genuine review queue through 2027 under the US duties will hold a year of that evidence as a by-product. A business whose US review step is token will discover the same flaw under two regimes at once — which is why the token-review problem is worth fixing on the US timeline, not the EU one.
Logs have the same dual use. The kit's hash-chained decision record — notices, outcomes, reviews, corrections in sequence — is the US record duty's artefact; as a log of system use and business response, it is also the raw material an Annex III deployer needs. Keep one log discipline, state-keyed and exportable, and both regimes read from it.
Sequencing 2027: one programme, three dates
The practical calendar for a transatlantic deployer: January 1, 2027 — Colorado duties live, California compliance phase begins. October 1, 2027 — Connecticut notices for new deployments. December 2, 2027 — EU Annex III deployer duties for the listed areas. Sequenced as one programme, the US build (inventory, assessments, notices, queues, record) creates the assets the EU date then consumes: the register gains an Annex III column, the vendor documentation file gains the instructions set, the queue's evidence extends.
Use the free checker's EU result as the standing reference for the Article 50 / Annex III position — it returns the same Article 50 answer for every role (the duties attach to interaction and content) and inserts the Annex III note for the listed decision areas, which is exactly the two-clock picture this guide describes.
- Map every system against the Annex III areas — employment, education access, creditworthiness, essential services — and record the finding per system
- Run the two clocks separately: Article 50 live since August 2, 2026; Annex III deployer duties from December 2, 2027
- Collect vendor instructions as the 'use as instructed' artefact, alongside the Colorado-style documentation set
- Build oversight and logs once, under the US duties, to a standard that evidences the EU duties as a by-product
- Set a pre-December 2027 readiness gate per Annex III system: instructions, oversight, logs — each verified, each owned
Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.
Related guides
- ADMT Notices for Hiring and Employment Decisions: A State-by-State Employer Guide
- ADMT Notices for Education Admissions and Placement Decisions: A Guide for Schools and Providers
- ADMT Notices for Lending and Credit Decisions: What Lenders Must Tell Applicants
- ADMT Deadline Tracker: Which State Duties Land When
See all ADMT guides for the full set.
Frequently asked questions
When do the EU high-risk deployer duties start?
December 2, 2027. The Digital Omnibus on AI (Regulation (EU) 2026/1744) postponed the Annex III high-risk obligations to that date, per the kit's checker data.
Which areas are on the Annex III list in the kit's data?
Employment and worker management, access to education and vocational training, creditworthiness evaluation, and access to essential public and private services.
What are the high-risk deployer duties?
As the checker summarises them: use the system as instructed, keep human oversight, and keep the logs.
What EU duties already apply before December 2027?
Article 50 transparency duties, since August 2, 2026: inform people they are interacting with AI, mark AI-generated content (with a grace period to December 2, 2026 for systems placed on the market earlier), disclose emotion recognition and deepfakes.
Does the postponement cancel the Annex III duties?
No. It moves them to December 2, 2027. Planning should treat the date as fixed and use the interval to build the oversight and log evidence the duties require.