The EU AI Act: Obligations by Role and Risk Tier

The AI Act now applies, and the timetable most plans are built on was changed in July 2026. Work out your role before you worry about risk tiers: what you have to do depends first on whether you built the system or are using it.

The timetable most people are working from is out of date

The AI Act applies from 2 August 2026, and that date has passed. But the schedule published in 2024 is not the schedule any more. A further EU regulation, 2026/1744, amended the Act in July 2026, and one of the things it changed was when the rules for high-risk systems start to bite. If your plan is built on the original timetable, it is built on a text that no longer applies.

So the bans have been in force since February 2025, and the high-risk work now has more time than it used to. Getting those two the wrong way round is the most common planning mistake here.

Four things sit outside that table and change what you have to do, and when:

• Labelling systems already on the market. Systems placed on the market before 2 August 2026 have to be marked by 2 December 2026.

• A grace period. A high-risk system of a kind that was lawfully on the market before the relevant date is not caught unless it is significantly changed. For systems meant for public authorities the date is 2 August 2030.

• AI in machinery is now dealt with through a separate machinery regulation, 2023/1230, with its own acts to apply by 2 August 2028 — a change that sits outside the AI Act itself.

• National testing sandboxes have to be running by 2 August 2027.

One correction worth stating plainly: the enforcement chapter applied from 2 August 2025 except for one article. Commission fines on providers of general-purpose AI models apply from 2 August 2026.

Work out your role before you worry about the tier

The Act hands out duties by role first. A provider develops a system, or has one developed, and puts it on the market under its own name. A deployer uses a system under its own authority, at work.

Most businesses turn out to be deployers of what they bought and providers of what they built and branded. Both roles can apply at once, to different systems, with different duties. And a deployer who puts its own name on a bought-in system can become its provider.

This is where the supplier contract does real work. What your supplier promises about classification, documentation and conformity decides how much of this actually lands on you, whatever the Act says about roles.

What is banned outright, including two new ones

Some practices are simply prohibited — not a tier to manage, but a ban on putting the system on the market, putting it into service or using it. The original list covers manipulative techniques that seriously distort behaviour and cause real harm, exploiting people's vulnerabilities, and social scoring.

The 2026 amendment added two more, from 2 December 2026: systems that create or manipulate realistic material showing an identifiable person's intimate parts or sexual conduct without that person's explicit consent, and systems that generate child sexual abuse material. Anyone running image or video generation should be checking their safeguards against those two before December.

High-risk, and the exception that narrowed it

There are two ways a system becomes high-risk, and they now start on different dates.

Safety parts of regulated products. A system is high-risk if it is a safety component of — or is itself — a product covered by certain EU product laws, and that product needs an outside body to assess it. The 2026 amendment narrowed this a lot. Systems used only for things unrelated to safety — helping the user, tuning performance, efficiency, convenience, quality control — are out. Systems whose failure would endanger health and safety are in, regardless. And a product that needs assessment only for reasons unconnected with health and safety, such as radio interference, does not meet the condition.

The Annex III list is the route most software businesses meet. There is a way out of it, and it is narrower than it looks. A system is outside only if it poses no significant risk of harm — including by not materially influencing the outcome of decision-making — and only if it does one of four things: a narrow procedural task, improving work a person already finished, spotting patterns in past decisions without replacing human judgement, or preparatory work. If you rely on this, write down why.

General-purpose AI models: a role the tiers do not cover

If you build a general-purpose AI model and put it on the EU market under your own name, you are the provider of that model. That is a separate role with its own duties, in force since August 2025. If you only build someone else's model into your product, you are the provider of a system, not of the model, and you rely on what the model provider gives you.

Changing someone else's model does not automatically make you its provider. Under the Commission's guidelines that only happens where the change is significant, and the rough measure is training compute above about a third of the original. If you fine-tune at scale, answer that question deliberately rather than by default.

The duties are mostly documentary: technical documentation for the AI Office, information for the businesses building on your model, a copyright policy that respects text-and-data-mining opt-outs, and a public summary of what the model was trained on. Models under a free and open licence are exempt from the first two unless they carry systemic risk. Systemic risk is a tier of its own, assumed above 10²⁵ FLOP of training compute, and it brings notification, adversarial testing, risk mitigation, incident reporting and cybersecurity duties. Models already on the market before August 2025 have until 2 August 2027 to comply.

Where Latvia has got to

For most of the economy the Act does its own work: it applies directly, with nothing needed from Latvia. Latvia has legislated in two places that matter.

The financial sector is done. The Financial Market Digital Operational Resilience and Artificial Intelligence Use Law applies the AI requirements to a long list of financial firms — banks, payment and e-money institutions, investment firms, insurers, fund managers and others.

There is a national AI Centre, with a sandbox. The Artificial Intelligence Centre Law sets up a Centre that, among other things, runs a special environment for developing AI systems, with its own rules on handling data.

For a business outside the financial sector, which Latvian authority will supervise it under the Act is not something this article can answer — it is worth checking rather than assuming.

Where to start

The order that works is role, then tier, then date. Work out whether you are a provider or a deployer of each system, remembering the answer can differ system by system. Check against the banned list first, because it is already live and two new bans arrive in December. Then classify against Annex III — and if you think the exception takes you out of it, write down why.