AICustom Software DevelopmentMid-Sized Business

Having AI Built Into Your Software? The AI Act May Call You the Provider

The EU moved the AI Act's high-risk deadline to December 2027, and many companies read that as time to wait. One question did not move: if you commission software with an AI feature and run it under your own name, the law can treat you as its provider, not the company that wrote the code. What that means, what already applies today, and what to settle in the contract.

Photo: Google Gemini · AI-generated

New to this topic? Start here

Terms used in this piece

AI Act
The EU regulation on artificial intelligence (Regulation (EU) 2024/1689). It sets rules by risk: some uses are banned, some are high-risk with strict duties, most carry light or no duties.
Provider
Under the AI Act, whoever develops an AI system, or has one developed, and puts it on the market or into service under their own name. Most of the duties sit here.
Deployer
Whoever uses an AI system in their own work, for example a company that switches on an AI feature in software it bought. Fewer duties than a provider.
High-risk AI system
An AI system in one of the areas the law lists, such as hiring, credit decisions or critical infrastructure. These need risk management, documentation, human oversight and a conformity check.
Digital Omnibus on AI
The 2026 amendment (Regulation (EU) 2026/1744) that moved several AI Act deadlines and softened some duties.
Annex III
The list in the AI Act of use cases that count as high-risk on their own, without being part of a regulated product like a machine or a medical device.

Read first

In July the EU moved the AI Act’s main deadline. Duties for stand-alone high-risk AI systems now apply from 2 December 2027 instead of August 2026, and for AI built into regulated products such as machines from 2 August 2028. It is tempting to read that as “we can leave it for now”.

Maybe. But the date was never the hard part for a company that commissions software. The hard part is a definition that did not change at all.

The definition that did not move

Article 3(3) of the AI Act defines the provider as whoever develops an AI system or has one developed, and puts it on the market or into service under their own name. Most of the Act’s duties sit with the provider: risk management, technical documentation, logging, human oversight, the conformity check.

Read that sentence again with a custom project in mind. A manufacturer asks a development partner to build a tool for its HR team, with an AI step that ranks applications. The partner writes the code. The manufacturer runs the tool, internally, under its own name. In the Act’s terms, the manufacturer has had an AI system developed and put it into service. It is very likely the provider. The partner who wrote every line is not.

That surprises people, because it is the opposite of how they think about bought software. If you switch on an AI feature in a product you licensed from a vendor, you are a deployer, with a much shorter list of duties. Commissioning is different. You are not buying a product; you are making one, with help.

We are developers, not lawyers, and the line between these roles has edge cases (Article 25 describes when a deployer becomes a provider, for example by putting their name on a system or changing its purpose). Take the reading above to your own legal adviser. But do not take it to them in 2027.

Most AI features will not be high-risk. Check anyway

The heavy duties only apply to high-risk systems. For most mid-sized companies, the AI features we see requested are not in that group: suggesting a product category, matching a delivery note to an order, sorting incoming emails. We wrote about that kind of feature in an earlier piece.

The list that decides it is Annex III, and one entry catches companies that do not think of themselves as AI users at all. Point 4 covers AI used to recruit or select people, to decide on promotion or termination, to assign tasks based on individual behaviour, or to monitor and evaluate staff performance. A shift planner that learns who is “reliable”, or a tool that scores technicians in the field, can fall under it. If an AI step in your plans makes a judgement about a person, read point 4 before you write the specification.

What already applies today

Two parts of the Act did not wait for 2027.

Since 2 August 2026, Article 50 requires that people know when they are talking to an AI. If your customer portal gets a chat assistant, the interface has to say so, from the first release. A system that generates images, audio, video or text has to mark its output in a machine-readable way, with a grace period to 2 December 2026 only for generative systems already on the market before August. If you had that system built, the provider question from above applies here too.

And since February 2025 there has been a duty, softened by the July amendment, to take measures that support AI literacy among staff who work with these systems. It no longer asks you to guarantee a level of knowledge. It still asks you to do something.

On our own side: when we tested short animated videos on LinkedIn in September, every one carried a line saying it was our own illustration animated with Google Veo, and every AI-generated image on this site is listed as such on our credits page. As far as we read Article 50, an illustration does not legally need that label. We added it because a reader who finds out later feels misled, and that costs more than the label.

What to settle before the first line of code

None of our 21 case studies contains a high-risk AI system, so we have not been through a conformity assessment with a client. We are not going to pretend otherwise. What we can say is where a project like that would go wrong, and it is early.

Write down, for every AI step in the plan, what it decides and about whom. If any answer is a person, assume you are the provider of a high-risk system until someone qualified tells you otherwise. Then put it in the contract: who writes the technical documentation, who keeps the logs and for how long, who designs the human oversight, and who does what when the model is retrained. Article 25 expects exactly that kind of written agreement between a provider and the companies that supply parts of its system.

If you are about to sign a proposal with an AI feature in it, and nobody has asked these questions yet, that is the moment to ask them. The deadline moved. Who is responsible did not.

Go further

Sources

  1. Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-LexThe amendment itself, adopted 8 July 2026.
  2. Article 3: Definitions, European Commission, AI Act Service DeskArticle 3(3) defines the provider, including whoever has a system developed.
  3. Article 25: Responsibilities along the AI value chain, European Commission, AI Act Service Desk
  4. The Digital Omnibus on AI enters into force today, Lewis SilkinPublication and entry-into-force dates, and the new high-risk deadlines.
  5. EU AI Act Update: Digital Omnibus Finalizes 8 Compliance Changes, OrrickSummary of the final changes, including the registration duty that was kept and the extension to small mid-caps.
  6. AI Act: transparency obligations and the AI Omnibus Regulation, HouthoffWhy Article 50 applies from 2 August 2026 and which marking duty got until 2 December 2026.
  7. AI literacy: the Digital Omnibus rewrites Article 4 of the AI Act, Law & TechnologyOld and new wording of the AI literacy duty side by side.
  8. Annex III: High-risk AI systems, artificialintelligenceact.eu (Future of Life Institute)Point 4 is the employment entry quoted in the text.