Skip to content
Zollwerk

EU AI Act · practical

The EU AI Act questions your buyers are already sending.

Article 50 has applied since 2 August 2026. Much of what is published online was written before the Digital Omnibus moved the high-risk deadlines and left Article 50 exactly where it was. This page separates the two, in the order a procurement questionnaire asks.

Dates last checked 3 September 2026 against the sources at the foot of this page.

Which dates actually apply to me?

Six dates matter. Only three of them bind an ordinary AI SaaS that is not high-risk.

DateWhat appliesWho it binds
2 Feb 2025Prohibited practices (Art. 5) and the AI literacy duty (Art. 4).Everyone
2 Aug 2025Obligations for general-purpose AI models, penalties, and the national authorities.Model providers
2 Aug 2026Article 50 transparency obligations apply. Not deferred by the Digital Omnibus.Providers and deployers
2 Dec 2026End of the grace period for marking synthetic output (Art. 50(2)) on systems already on the market before 2 August 2026.Providers
2 Dec 2027Annex III high-risk obligations — moved back from 2 August 2026 by the Digital Omnibus.High-risk only
2 Aug 2028Annex I high-risk obligations, for AI inside already regulated products.High-risk only

The Digital Omnibus postponed the Annex III high-risk deadline to 2 December 2027. It did not postpone Article 50. A checklist telling you that everything moved to 2027 is describing an obligation you probably do not have, and skipping the one you do.

The questions, in the order they get asked

Am I a provider or a deployer?

Both are defined in the Regulation, and you can be both at once. A provider develops an AI system, or has one developed, and places it on the EU market under its own name or trademark. A deployer uses an AI system under its own authority in a professional capacity.

  • If you ship a product with a chat feature under your own brand, you are the provider of that system — even though you did not train the model.
  • You are simultaneously a deployer of the model you call.
  • Personal, non-professional use falls outside the deployer definition entirely.

Do I have to tell people they are talking to AI?

Yes, where they interact with it directly, unless that is obvious. Article 50(1) puts the duty on the provider: a person must be informed they are interacting with an AI system, unless it would be obvious to a reasonably well-informed person. In practice that means a visible disclosure in the interface, not a sentence in your terms.

Do I have to mark what the model generates?

If your system produces synthetic audio, image, video or text, Article 50(2) requires that output be marked in a machine-readable format as artificially generated or manipulated. The duty sits with the provider. Systems already on the market before 2 August 2026 have until 2 December 2026 for this one obligation.

  • Exempt: output where the AI performs an assistive function for standard editing.
  • Exempt: source code, and short sequences of numbers, symbols or letters.
  • Exempt: machine-to-machine output with no human exposure.

Does the deepfake labelling rule apply to me?

Only if you publish. Article 50(4) is a deployer duty covering deepfakes, and AI-generated text published to inform the public on matters of public interest. Text that has undergone human review or editorial control is exempt.

  • Human review means deliberate examination by a person with relevant knowledge and professional judgement. A spell-check does not qualify.
  • Content that is evidently artistic, creative, satirical or fictional is disclosed in an appropriate manner rather than labelled flatly.

What does an EU buyer actually ask in the questionnaire?

Three things, repeatedly: which models you use and where they run, what sits between a user and the model, and what you can show after the fact. These arrive as security and procurement questionnaires long before any authority contacts you — which is why the deadline that costs you money is usually your next enterprise deal, not an enforcement date.

  • Which model providers and subprocessors are in the path, and in which country they run.
  • Whether prompts or customer data are used to train anyone’s model, and how long they are retained.
  • What controls exist on spend, on personal data leaving your servers, and on prompt injection.
  • What record you can produce of a specific call, months later.
  • A data processing agreement and a current subprocessor list.

What Zollwerk ships toward these

Mechanisms, mapped to the need they serve. Read the middle column as “what exists in the codebase”, never as “therefore you are done”.

What is being asked of youWhat the kit ships
Tell a user they are interacting with AI (Art. 50(1)).Article 50 disclosure mechanics wired into the app shell.
Show, after the fact, what a specific call did.An append-only audit ledger recording every model call, including the ones that were refused.
Say which models ran, and where.One gateway every call passes through, self-hosted on your own EU server, with the full source in your repo.
Answer the subprocessor and data-processing questions.A compliance pack: AVV, subprocessor list, checklists — templates your lawyer reviews rather than writes.
Keep a coding agent inside the rules it agreed to.A hard daily spend ceiling, PII redaction and prompt-injection screening, enforced below the level the application can edit.

These are technical building blocks. They do not make you compliant, and no software sold at this price can — that is a judgement about your specific product, made with your lawyer.

Pre-order — €99
30 seats · ships September 2026 · refundable until deliverySee the full pricing

Sources

Every claim above comes from one of these. Check them yourself; they are short.

This page is a summary of published sources, written by a developer, and it is not legal advice. It cannot tell you whether your product is high-risk, and it does not create a lawyer–client relationship. Zollwerk is software and document templates: technical building blocks. Review your obligations with your lawyer.