Most organisations deploy AI they did not build. That does not transfer the risk. Air Canada, Rite Aid and the employers in Mobley v. Workday all used someone else's technology; they answered for it anyway. Your contract is where you turn the vendor's promises into evidence.
Know what the vendor is
Ask three questions before signing:
- Is the vendor a provider under the EU AI Act or a developer under Colorado/Texas law? If so, it owes you documentation, instructions for use and cooperation with your oversight and incident duties (EU AI Act Articles 13, 16 and 25).
- Does it train on your data? Get the answer in writing, and the switch to turn it off.
- Where does the data go? Sub-processors, regions, retention.
The clauses that matter
Add these to your standard terms or a short AI addendum:
- Description of intended purpose and known limitations, matching what you will actually do with it. Using a tool outside its intended purpose can make you the provider.
- Documentation and instructions for use sufficient to meet your duties (EU AI Act Article 26 requires deployers to use systems in accordance with the instructions).
- Testing and fairness evidence: the vendor's bias audit or equivalent, refreshed annually, and cooperation with audits you run on your own data (NYC Local Law 144 needs this).
- Transparency support: content marking, chatbot disclosures and model cards you can pass to your users.
- Data protection: a processing agreement, no training on your data without consent, deletion on exit, breach notice timelines.
- Security: patching, access control, prompt-injection defences, and the right to see a penetration-test summary.
- Incident cooperation: notice to you within a fixed number of hours of a malfunction or serious incident, with the information you need for your own reporting (EU AI Act Article 73; New York's RAISE Act uses 72 hours for frontier developers).
- Change notice: advance warning of model updates that could change behaviour, and a way to pin or roll back.
- Records and audit: the right to obtain logs and to have your regulator inspect.
- Exit: export of your data and configuration.
The agency problem
In Mobley v. Workday, the court let claims proceed against the software vendor on the theory that a tool making or substantially shaping hiring decisions acts as the employer's agent. For you as a deployer that cuts both ways: the vendor may share liability, but you cannot outsource it. Indemnities are worth negotiating; they are not a substitute for your own testing.
Small vendors and free tools
You will not get a bespoke addendum from a free chatbot. You can still: read the terms for training and retention, choose business tiers where data is excluded from training, and restrict what staff may paste in. Record that decision in your inventory.
Re-check on renewal
Vendors change models, terms and sub-processors. Put "AI clauses reviewed" on the renewal checklist, and re-run your fairness test after any material model change.