AI Act Article 50 and AI over your ERP

The transparency obligations in Article 50 of the EU AI Act apply from 2 August 2026. They are not limited to high-risk AI systems. A company with no high-risk AI at all can still hold duties under this article.

If you run an AI assistant over ERP data, this page explains which duties apply to you, which ones apply to your vendor, and what changed in the deadlines.

This page is information, not legal advice. Take the decisions with your own adviser.

What does Article 50 require?

Article 50 covers four situations:

1. Direct interaction with people. Chatbots, virtual assistants and other automated systems that a person talks to.
2. Synthetic content. Text, images, audio or video produced by AI.
3. Emotion recognition and biometric categorisation. Systems that assess stress, sentiment or demographic categories.
4. Deepfakes and AI-generated text on matters of public interest.

The duties split between the provider of a system and the deployer of it.

The information must reach the person at the latest at the time of the first interaction or exposure. It must be clear and distinguishable, and it must meet accessibility requirements.

Does Article 50 apply to an ERP assistant?

Situation 1 applies to almost every ERP assistant, because a person types a question and receives an answer from an AI system.

The duty that follows is modest. The person has to know that the answer came from an AI system. A clear label in the interface satisfies it. This is generally the vendor's duty as the provider, and it is worth confirming that the product does it rather than assuming.

Situation 2 applies as soon as the assistant produces content that leaves the building: a drafted email, a generated document, a report narrative. The machine-readable marking duty sits with the provider of the generative system.

Situations 3 and 4 do not usually apply to ERP work.

Is an ERP assistant a high-risk AI system?

Usually no, and the answer depends on what you use it for, not on what it is.

The high-risk categories concern uses such as employment decisions, creditworthiness assessment of natural persons, access to essential services, and safety components of regulated products. An assistant that answers questions about ledger entries is not in those categories.

The same assistant can move into one. If you use it to score applicants, to rank employees, or to decide whether a natural person receives credit, look at the high-risk rules with your adviser. The use case decides, and the use case can change without anybody filing a ticket.

What changed in the deadlines?

The Digital Omnibus package changed the timing for high-risk obligations. As described by the law firms tracking it:

Two points of caution. The scope of Article 50 did not change, only the marking deadline for systems already on the market. And the omnibus amendments required formal adoption by the Parliament and the Council, which the Latham & Watkins note dates as expected by July 2026. Confirm the current status before you rely on the later dates.

What should a company do now?

A short list, in the order we would do it.

1. Write down every AI system you use. Include the assistants inside products you bought. This is the step most companies have not done, and every later step needs it.
2. Check the disclosure. Open each assistant and confirm that a person can tell they are talking to an AI system.
3. Decide about generated content that leaves the company. Emails, letters and reports drafted with AI are the practical case.
4. Ask each vendor two questions. Are you the provider of this system under the AI Act, and how do you meet Article 50? A vendor who cannot answer this in September 2026 is behind.
5. Check your use cases against the high-risk list, and re-check when somebody proposes a new use.
6. Keep the record. The obligation you can prove is worth more than the obligation you met.

The Danish context

Danish companies meet this rule alongside a second change. Since 1 January 2026, the digital bookkeeping requirement extends to sole proprietorships and associations with net revenue above 300,000 kr. for two consecutive years, who must use a bookkeeping system registered with the Danish Business Authority.

The two rules point the same way. The bookkeeping system is the record. The AI assistant reads that record and helps a person act on it. Keeping the boundary clear between the two makes both rules easier to satisfy.

What Article 50 does not require

It does not require you to stop using AI on financial data. It does not require a risk assessment for an ordinary ERP assistant. It does not make your vendor responsible for your use of the output.

It requires that people know when they are talking to a machine, and that machine-made content can be identified as such. Those are reasonable requirements, and a well-built product satisfies them without a project.