Working with AI vendors

You are still the deployer

When you buy an AI product you gain a capability and keep your duties. Regulators, customers and the people your decisions affect will look at you, not the vendor. That is why procurement is a governance step, not just a buying step. Do your checking before you sign, because problems are cheaper to find then.

Before you buy: due diligence

Sixteen questions in nine groups cover the ground. Highlights:

• Governance: who is accountable at the vendor, and which frameworks or certifications (for example ISO/IEC 42001 or the NIST AI RMF) do they follow?
• Data: what was the system trained on, and do they use your data to train or improve it? Where is data stored and processed, by which sub-processors, and for how long?
• Transparency: can they share a model or system card with intended use, known limits and test results?
• Testing: what accuracy, bias and robustness testing have they done? How do they test security, including prompt injection and leakage?
• Oversight and logs: what override features exist, and what logs do you get?
• Change management: how and when will they tell you about model changes, incidents or breaches?
• Legal: which EU AI Act role and tier do they consider the product to be, and how do they support your duties as a deployer?
• Exit: can you export your data and stop using the product, and what happens to your data afterwards?

The vendor due diligence tool marks which questions you must ask for your risk level, tracks the answers and creates email text you can send.

In the contract

Turn answers into obligations. A short checklist of topics for counsel to draft: data use and training rights, security commitments, incident notice, audit rights, change notice for model updates, performance commitments, liability, and exit and data return. For personal data, add data processing terms. The contract clause checklist and the data processing checklist list the topics.

What a certificate does and does not tell you

A vendor with an ISO/IEC 42001 certificate runs an AI management system that an auditor checked. That is useful evidence. It does not show that a particular product is lawful or right for your use. Each system still needs its own assessment. Similarly, a vendor’s compliance claim is a statement, not proof: ask for the evidence.

After you buy

• Monitor. Track performance and complaints, and for decision tools track outcomes by group.
• Watch for change. Vendors update models. Ask for advance notice and re-test after major changes. Add a vendor-version check to monitoring.
• Review. Re-run due diligence at renewal.
• Keep an exit plan. Know how you would leave and what you would need to take with you.
• Keep records. Store questionnaires, contracts and evidence so you can show what you did.

Common mistakes

• Accepting “we are compliant” without evidence.
• Not asking whether the vendor trains on your data.
• No change notice, so a model update alters risk without warning.
• Discovering AI features switched on inside software you already use.

Example

A company bought a CV-ranking tool. Due diligence showed the vendor would train on its applicant data and keep logs for only 30 days. The company asked for an opt-out in writing, longer log retention and a clause requiring notice of model changes. Months later the vendor updated its model and fairness dropped; because of the clause, the company knew quickly, paused the tool and had the vendor roll back. See the hiring tool worked example.

What to do now

• Open a vendor file for the next AI product you buy or renew.
• Send the must-ask questions.
• Put incident notice, change notice, data use and exit into the contract.