Plenty of compliance teams spent the first half of this year working toward 2 August 2026, the date the EU AI Act's high-risk obligations were due to apply, until the Digital Omnibus on AI moved it. Regulation (EU) 2026/1744 was published on 24 July. Obligations for stand-alone high-risk systems in Annex III, the list covering AI used in hiring, credit, education and access to essential services, now apply from 2 December 2027, and high-risk AI built into regulated products under Annex I moves to 2 August 2028.
As Jones Walker pointed out in July, most of Article 50 still applied from 2 August, including telling people when they're dealing with an AI system unless it's obvious and marking AI-generated content in a machine-readable way. The one extension covers generative systems already on the market, which get until 2 December 2026 for the technical marking.
A month on, some teams are treating the delay as permission to stop. The work the deadline was forcing, knowing which of your AI systems are high-risk and reviewing them harder, is worth doing anyway, because a mistake in a high-risk system reaches a customer or an employee.
Case-by-case review jams up
A lot of companies review AI use one request at a time. Someone wants to try a new agent, and a reviewer asks some questions and approves it or doesn't. That holds up until the requests pile up, and then the reviewer is the bottleneck and the answers depend on who reviewed what.
A tier gives every use case a level up front, and the level decides the treatment. Nobody debates how much testing a meeting summariser needs, because it's low tier and low gets a registration and a data-terms check. The reviewer's time goes to the few high-tier requests.
Two questions place almost every use case
What goes in? Public or internal information that would do no harm if it leaked sits at one end. Personal data, financial records, contracts and health information are more sensitive, and anything a regulator protects by name is the most sensitive of all.
What can it do with the output?
- A person reads it and decides.
- An agent drafts something a person approves before it goes anywhere.
- An agent acts on its own: it sends, pays, updates a system of record, or makes or shapes a decision about a person.
Anything that acts without approval starts at medium, whatever data it touches.
Low. Non-sensitive data in, and a person reads the output and decides. Internal drafting from public or internal material fits here. Register it and check the vendor's data terms.
Medium. Sensitive data in, or an agent that drafts actions for a person to approve, or one that acts on its own with non-sensitive data. A support agent that drafts replies for a person to send is typical. It gets a named owner, testing against real past cases before launch, and a monthly look at its logs.
High. The agent can act on sensitive data without approval, or it shapes decisions about people. Screening job applicants, setting credit limits, approving refunds above a threshold and anything on the Annex III list all land here. Test it before it touches production and have a person approve its actions until the logs show a track record. Write down how it was evaluated, so you can explain it to a customer or a regulator.
If a use case could fit two levels, put it in the higher one, and move it down after a few months of clean logs.
The transparency check you owe now
While you're sorting, add one column: does a person outside the company interact with this, or see content it made? If your AI systems reach people in the EU, the chatbot disclosure duty applies to those entries today. Check support chats that don't say they're an AI, and ask your vendors how they mark generated images and copy.
Re-tier when permissions change
Tiers drift. A medium-tier agent that quietly picks up broader permissions over a year can end up doing high-tier work under medium-tier review. Re-tier whenever the data or the authority changes, and make granting a new permission the trigger for that review.
Tiers also only cover what's on the list. If you don't have an inventory of the AI your departments already run, build that first.
Between now and December 2027
Fix the transparency gaps a customer can see this month, since that duty already applies.
If you already have an inventory, a first pass through the two questions takes about a day for a mid-size company. Pull out anything that lands on the Annex III list: those systems need their risk management, testing and documentation in place by 2 December 2027, fifteen months from now.
Frequently asked questions
When do the EU AI Act's high-risk obligations apply now?
Under the Digital Omnibus on AI, Regulation (EU) 2026/1744, the obligations for stand-alone high-risk systems listed in Annex III, such as AI used in hiring or credit decisions, apply from 2 December 2027. Obligations for high-risk AI built into products covered by Annex I apply from 2 August 2028.
What applies from August 2026?
Most of the Article 50 transparency duties, including telling people when they're interacting with an AI system unless it's obvious and marking AI-generated content in a machine-readable way. Under the Digital Omnibus, generative systems already on the market before 2 August 2026 have until 2 December 2026 for the technical marking.
What is an AI risk tier?
A level you give each AI use case based on the data it takes in and what it's allowed to do. The level sets how much review the use case gets before and after it goes live, so the most review goes to the uses that can hurt someone.
How many risk tiers should we have?
Three is enough for most companies: low, medium and high. Extra levels tend to add arguments about borderline cases.
Does the delay mean we can stop preparing for the AI Act?
No. If your AI systems reach people in the EU, the chatbot disclosure duty already applies. Confirming which systems count as high-risk under the Act, and getting their risk management and documentation in place, takes months.