Does the AI Act apply to us? Four questions that settle most of it

Provider or deployer: why the answer changes everything

The AI Act divides the world into providers and deployers. If you build an AI system and put it on the market, or build it for your own use, you're a provider. If you use someone else's AI system under your own responsibility, you're a deployer ^[1].

Think of it like a car. The manufacturer is the provider. The taxi company using it is the deployer. Both have duties, but they're different duties.

For high-risk AI, the provider does the heavy lifting: risk management, data governance, technical documentation, automatic logging, conformity assessment and a quality management system ^[2]. The deployer makes sure it's used right, monitors it in practice and tells people when decisions affect them ^[3].

The high-risk requirements under Annex III apply from 2 December 2027 ^[4]. Until then, the transparency rules in Article 50 already apply, and that's where most organisations meet the AI Act first.

Decisions about people: the high-risk lists

Annex III is the list everyone worries about. AI systems on this list are high-risk if they make decisions about people in areas like biometrics, critical infrastructure, education, hiring, workforce management, access to public benefits, credit scoring of individuals, pricing of life and health insurance, emergency triage, law enforcement, migration and the justice system ^[5].

If the AI makes the decision, or feeds into it in a way that materially influences the outcome, it counts. A tool that only formats data or runs a spellcheck doesn't.

For deployers, Article 26 sets out what you must do before you start using a high-risk system: use it according to the instructions, assign someone to keep an eye on it, monitor how it performs, tell trade union representatives and employees before you use it at work, and inform people affected by decisions ^[3].

That last one matters. If your AI turns someone down for a loan, a job or a flat, they have a right to know it was an AI, and to understand why.

Content that could pass as real: Article 50

Article 50 has applied since 2 August 2026 ^[6]. It says three things.

First, if people are talking to an AI, they must be told, unless it's obvious from the context ^[6]. A chatbot on a website needs a notice. An obvious game character doesn't.

Second, providers must mark AI-generated content in a way machines can read ^[6]. That's on the people who build the tools, not on everyone who uses them.

Third, deployers have two transparency duties. If you use AI to create or manipulate images, audio or video that could be mistaken for real people, objects, places or events—a deepfake, in the regulation's terms—you must disclose that it's artificial or manipulated. If it's part of an artistic, creative, satirical or fictional work, a suitable disclosure that such content exists is enough, as long as it doesn't stop people from seeing or enjoying the work. And if you publish AI-generated text to inform the public about matters of public interest, you must disclose that too, unless a human has reviewed it and someone has editorial responsibility ^[7].

The idea is simple: people should know when they're not dealing with a human or seeing something real.

AI inside regulated products: Annex I and the 2028 date

If your AI is part of a product that already falls under EU safety legislation—medical devices, toys, lifts, radio equipment and the like—it becomes high-risk under Annex I ^[8].

This one has a later date: 2 August 2028 ^[9]. Until then, the existing product rules still apply, but they don't yet treat the AI part as high-risk under the AI Act.

Say you make a diagnostic tool that uses AI and counts as a medical device. After August 2028, the AI part is high-risk, which means the documentation, risk management and logging duties in the AI Act apply on top of the Medical Devices Regulation.

The practical question is whether your technical file and quality system already cover what the AI Act will ask for. If they do, the gap is small. If they don't, now is the time to close it.

What this means for you

Start with the four questions.

Do you build the AI or use it? If you build it, you're the provider. If you use someone else's, you're the deployer. The duties follow from that.

Does it make decisions about people in hiring, credit, benefits, law enforcement or another area in Annex III? If it does, it's high-risk from 2 December 2027, and the logging, monitoring and transparency duties apply.

Does it create content people might mistake for real, or talk to your customers? If it does, Article 50 already applies. Tell people they're talking to an AI, and disclose deepfakes and certain AI-generated text.

Is the AI part of a medical device, a toy, a lift or another regulated product? If it is, it becomes high-risk from 2 August 2028 under Annex I.

If none of that applies, you're likely outside the high-risk net, but the transparency rules might still catch you.

Either way, ask yourself one more question: if someone disputed a decision, a conversation or a piece of content a year from now, could you show what your AI actually did?

The regulation asks for logs. Whether those logs hold up is another matter.

What a digital witness does here

This is where a digital witness fits, particularly for deployers of high-risk AI and anyone whose AI talks to customers or makes content.

Say you deploy a hiring tool that's high-risk under Annex III. Article 26 says you must be able to show what happened: which candidate, which model version, what score, who reviewed it. The thing to seal is each decision record as it's written to the log—the input reference, the model version, the output and the reviewer's name—or log batches every few minutes.

Sealing takes a fingerprint of the record, not the record itself, so the data never leaves your systems. When a candidate asks for an explanation six months later, or an auditor wants to see the log, anyone you share it with can check that it's exactly what was written on the day. Change one character and the fingerprint won't match.

The same idea works for chatbots under Article 50. The customer's screenshot is often the only record of what the bot said. Sealing each answer, or each conversation when it closes, gives you your own record. It doesn't settle whether the company is bound by what the bot promised, but it does show what it actually said.

For AI-generated content, the witness doesn't replace the marking duty in Article 50, but it does give you a record of what was generated, when and by which system, in case the marking is stripped or questioned later.

What it doesn't do is show that a decision was fair, that an answer was correct or that a log is complete. It shows that a record hasn't changed since it was sealed, even by someone with administrator rights. Less to trust, more to check.

You can start with WitniumChain through the REST API, the SDK or the MCP server. The content stays with you; only the fingerprint is sealed. The first 2,000 witnesses a month are free. Full documentation is at witniumchain.com/docs.

How it would work

1. Your system sends a fingerprint of the decision record, answer or content to WitniumChain and gets a witness ID back.
2. The fingerprint is sealed on a chain where any later change would show, even one made by us.
3. When it's questioned, anyone you share the record with calculates the fingerprint again and checks it against the sealed one.

What to watch

2 December 2027 is the date for high-risk systems under Annex III ^[4]. If you're deploying one, the monitoring, logging and transparency duties start then.

2 August 2028 is the date for AI in regulated products under Annex I ^[9]. If your AI is part of a medical device or another product on that list, the high-risk requirements apply from that day.

In the meantime, Article 50 already applies ^[6]. If your AI talks to people or makes content, the transparency duties are live now.