EU AI Act high-risk deadline: what has to be ready by December 2, 2027

If you sell an AI product into Europe, the date on your roadmap is probably wrong.

For two years, teams planned around August 2, 2026 for the EU AI Act's high-risk rules. That date came and went without them. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on July 27, 2026 and moved the deadline. The high-risk dates in Article 113 of the amended Act are now:

• December 2, 2027 for the high-risk use cases listed in Annex III: employment, credit scoring, life and health insurance pricing, education, access to essential services and the rest.
• August 2, 2028 for AI built into products already covered by EU product law under Annex I, such as medical devices and machinery.

Annex III is 409 days away. That sounds like plenty until you count what has to exist by then.

I'm an engineer, not a lawyer. This is how I'd sequence the technical work. Have your counsel confirm what applies to you.

What did not move

The delay is narrower than the headlines made it sound. On the Commission's official timeline:

• Prohibited practices and AI literacy have applied since February 2, 2025. The amended Article 4 still requires providers and deployers to take measures to support AI literacy of their staff, although it doesn't require a guaranteed level for each person.
• General-purpose AI model obligations have applied since August 2, 2025.
• Transparency rules in Article 50 have applied since August 2, 2026, and enforcement started that day for GPAI models, prohibitions, transparency and AI literacy.
• December 2, 2026 brings new prohibitions on AI that generates non-consensual sexual deepfakes or child sexual abuse material. It's also the cut-off for providers of generative systems already on the market before August 2, 2026 to meet the Article 50(2) marking rules.

If your product is a chatbot or generates content, you have obligations today, not in 2027.

Does it reach a company in Toronto or Riyadh?

Article 2(1) is direct. It covers providers placing AI systems on the EU market "irrespective of whether those providers are established or located within the Union or in a third country," and providers and deployers in a third country "where the output produced by the AI system is used in the Union."

So the question isn't whether you have an office in Europe. It's whether anyone in the EU uses what your system produces. A Canadian HR-tech vendor with one German customer screening candidates is in scope. A Dubai lender that only scores UAE residents probably isn't.

Is your system actually high-risk?

Annex III is a specific list. The categories I see most often:

• Employment (point 4): recruiting, filtering applications and evaluating candidates, plus systems that inform promotion, termination, task allocation or performance monitoring.
• Credit (point 5(b)): evaluating creditworthiness or setting credit scores for individuals. Fraud detection is expressly carved out.
• Insurance (point 5(c)): risk assessment and pricing for life and health insurance.
• Healthcare access (points 5(a) and 5(d)): public authorities, or someone acting for them, evaluating eligibility for healthcare services and benefits, and emergency patient triage.

Then there's Article 6(3). An Annex III system isn't high-risk if it doesn't pose a significant risk of harm, for example because it performs a narrow procedural task or a preparatory task, or only improves the result of work a human already finished. Two catches. Any Annex III system that profiles people is always high-risk. And under Article 6(4), a provider relying on the exception must document that assessment before the system goes to market, is still subject to registration under Article 49(2), and must hand the documentation over when authorities ask.

My rule of thumb: if you plan to claim Article 6(3), write the assessment now, while the people who designed the feature still work there.

What I'd build first

Providers carry the requirements in Articles 9 to 15. Deployers carry Article 26. The pieces that take longest aren't the documents. They're the plumbing.

1. Logs that can reconstruct a decision. Article 12 requires high-risk systems to technically allow automatic recording of events over their lifetime. Article 26(6) requires deployers to keep the logs under their control for at least six months unless other law says otherwise. For an LLM workflow, that means prompt template version, model version, retrieved sources, output and the human action taken, all tied to one request ID. Retrofitting that into a production pipeline is a quarter of work, not a sprint.

2. Human oversight that exists in the product. Article 14 covers design for providers, and Article 26(2) requires deployers to assign oversight to people with the competence, training and authority to act. "A human can review it" isn't a control. A review queue with an approval step, an override and a record of who approved is.

3. A model change process. Hosted model providers ship new versions on their own schedule. Your risk management (Article 9) and your accuracy and robustness claims (Article 15) assume the model you tested is the model in production. Pin versions where you can, rerun a fixed evaluation set on every change, and keep the results.

4. Notices to people. Under Article 26(11), deployers of Annex III systems that make or help make decisions about people must tell those people. Under Article 26(7), employers must inform workers' representatives and affected workers before using a high-risk system at work. These are copy, UX and contract changes. They're cheap now and awkward later.

5. A fundamental rights impact assessment where required. Article 27 requires one before first use for public bodies, private entities providing public services, and deployers of credit scoring and life and health insurance pricing systems.

A note for financial institutions: under Article 26(5) and 26(6), institutions already subject to EU financial-services governance rules meet the monitoring duty through those rules and keep the logs as part of the documentation they already maintain. If you're a Canadian bank or insurer, OSFI's Guideline E-23 on model risk takes effect May 1, 2027. A good AI model inventory, monitoring thresholds and change controls serve both regimes, so build them once.

What about systems already live?

The amended Article 111(2) says the Act applies to high-risk systems placed on the market or put into service before the high-risk rules start only if they undergo significant changes in their design from that date. Systems intended for public authorities must comply by August 2, 2030 regardless.

It's tempting to read that as "ship before December 2027 and you're grandfathered." I wouldn't build a plan on it. Most AI products change constantly, and swapping the underlying model could easily count as a significant design change. Ask counsel, and build the logging anyway.

What it costs to ignore

Under Article 99(4), breaching provider obligations under Article 16 or deployer obligations under Article 26 can mean fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. Prohibited practices go up to EUR 35 million or 7%. For SMEs and start-ups, Article 99(6) caps each fine at whichever of the two figures is lower.

Where to start this month

1. List every AI feature whose output reaches someone in the EU.
2. Check each one against Annex III, and write the Article 6(3) assessment for anything you believe falls outside it.
3. Pick the high-risk system closest to customers and try to reconstruct one real decision from your logs, end to end.

That third step tells you quickly how far away December 2, 2027 really is.