The EU AI Act procurement checklist, for teams buying AI.

When your company buys an AI tool, the EU AI Act usually makes you its deployer under Article 3(4). The vendor carries most of the obligations for the system itself, but some duties stay with you, and a few depend on how you use what you bought. This checklist is the buyer's side of the review: what to confirm before you sign, what to put in the contract, and which vendor answers should make you slow down.

Selling AI rather than buying it? This is the list your customer's procurement team works from. The vendor questionnaire guide covers the same ground from your side, question by question.

Step 1: classify the use, not the product

The Act attaches obligations to an AI system and its intended purpose. The same transcription tool is minimal risk when it writes meeting notes and can sit inside an Annex III area when it feeds a hiring decision. Before you assess the vendor, write down in one sentence what you will use the system for.

Then check that use against Annex III. The eight areas are biometrics, critical infrastructure, education, employment and worker management, access to essential services (including credit scoring and life and health insurance pricing), law enforcement, migration and border control, and the administration of justice and democratic processes. If your use falls in one of them, you are buying a potential high-risk system and the rest of this list matters more. The free high-risk checker walks through Annex III and the Article 6(3) exemption in a few minutes.

Step 2: the procurement checklist

Annex I products, meaning AI built into products already covered by EU product-safety law such as medical devices or machinery, follow a later date: 2 August 2028. Every date in one table.

Step 3: your own duties as the deployer

Buying from a well-documented vendor does not finish the job. Two duties apply to almost every buyer today, whatever the risk tier:

• AI literacy, Article 4. Staff who operate or rely on the system need a sufficient level of understanding of what it does and where it fails. Keep a dated record of who was briefed and on what. What Article 4 actually requires.
• Deployer transparency, Article 50(3) and 50(4). If you use emotion recognition or biometric categorisation on people, you must inform them. If you publish deepfakes, or AI-generated text meant to inform the public on matters of public interest, you must disclose that it was generated or manipulated.

If the system is high-risk, Article 26 adds deployer duties from 2 December 2027: use it according to the instructions for use, assign human oversight to people with the competence and authority to act, make sure input data you control is relevant, monitor operation and report risks to the provider, keep the logs under your control, and inform workers' representatives before using it in the workplace. Public bodies, private entities providing public services, and deployers using credit scoring or life and health insurance pricing also owe a fundamental rights impact assessment under Article 27.

Step 4: what to put in the contract

Most of the checklist becomes enforceable only if the contract says so. (Selling AI rather than buying it? The same clauses from the vendor's side.) Clauses worth asking for:

• Classification warranty. The vendor's stated role and risk tier per feature, as of a stated date, attached as a schedule.
• Change notification. Notice before any change to a feature's intended purpose, model provider or classification.
• Cooperation. Help with your own Article 26 and Article 27 duties where the system is high-risk, including access to the instructions for use and relevant logs.
• Incident information. Prompt notice of serious incidents affecting your use. For high-risk systems, providers report serious incidents to authorities under Article 73.
• Disclosure support. Where the vendor provides the user-facing interface, confirmation that the Article 50(1) disclosure is built in and cannot be switched off by accident.

Red flags in vendor answers

• "We are EU AI Act certified." There is no AI Act certificate for limited-risk or minimal-risk systems. Conformity assessment and CE marking exist for high-risk systems only, and for most Annex III systems the assessment is carried out by the provider itself. Ask what exactly was assessed, by whom and when.
• "We only use the OpenAI API, so the Act does not apply." The model provider's Article 53 duties are separate from the vendor's duties for the system it sells. Where the line sits.
• "A human always makes the final decision." In an Annex III area that is not, on its own, an Article 6(3) argument. Ask how often reviewers actually override the system.
• Pre-Omnibus dates. A vendor still quoting August 2026 for Annex III high-risk duties, or 2027 for chatbot disclosure, has not updated its own assessment since June 2026. What the Omnibus changed.
• No date anywhere. A classification with no date cannot tell you which version of the rules it was made against.

What a breach can cost, and who pays

Fines under Article 99 fall on the operator that breached its own obligation: provider duties on the vendor, deployer duties on you. Most operator breaches, including Article 50 and Article 26, sit in the Article 99(4) tier: up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, Article 99(6) applies the lower of the two. The fine calculator shows the ceiling for your own turnover, and the penalties guide explains the tiers.

Keep the file

For each AI system you buy, keep one folder: your one-sentence intended use, the vendor's dated classification and answers, the contract schedule, your Article 4 record and, for high-risk systems, the instructions for use and your oversight assignment. Review it when the vendor ships a material change or when you change how you use the tool. If you also build AI yourself, the compliance checklist gives the provider-side sequence.

Reglynn's free scan gives a likely risk tier for a system in about two minutes: eight questions plus your email. It is a structured self-assessment, not legal advice and not a certification, and it works the same way whether you are classifying a tool you built or one you are about to buy.

Frequently asked questions

Does the EU AI Act apply to companies that only buy AI tools?

Yes, in a narrower way. A company using an AI system under its authority is a deployer under Article 3(4). Article 4 AI literacy applies to deployers now, Article 50(3) and 50(4) transparency applies where relevant, and Article 26 adds deployer duties for high-risk systems from 2 December 2027.

Can a buyer become the provider of an AI system?

Yes. Under Article 25, a deployer that puts its own name or trademark on a high-risk system, substantially modifies it, or changes its intended purpose so that it becomes high-risk can be treated as its provider and take on the provider's obligations.

Should I ask an AI vendor for an EU AI Act certificate?

Only for high-risk systems. Conformity assessment and CE marking under Articles 43 and 48 apply to high-risk systems, from 2 December 2027 for Annex III. For limited-risk and minimal-risk systems there is no certificate; ask for a dated classification and evidence of Article 50 disclosure instead.