What to put in an AI vendor contract

McKinsey's late-2025 survey found 88% of organisations using AI and only 7% scaled it into results. Many of the other 81% signed a contract that promised outcomes and specified nothing. The demo answered the questions; the contract is where the answers become obligations. This is a checklist for your counsel, not a substitute for one.

1. Handoff rules, written into the schedule

Which calls and messages go to a person, how fast, and with what information collected. Emergencies, complaints, anything with legal or clinical weight. If the vendor will not put the rules in writing, they do not exist.

2. Your data, your export

Ownership of calls, transcripts, leads and records stays with you. Export of everything, in usable formats, on request and on termination. What the vendor may train on, stated plainly. Where the data is stored and who can see it.

3. Support with a caller-facing definition

Not just uptime. What the caller hears while the engine is broken, how fast a person is reachable, and who owns the fix when a calendar, phone or CRM integration changes.

4. A baseline and a review date

The engagement starts with a measured baseline of your own numbers and names a review point at which results are compared to it. No recovery percentage promised in advance; a measurement promised instead.

5. Termination without hostages

Phone numbers, calendars, knowledge bases and integrations remain yours and keep working through transition. Notice periods that match a monthly retainer, not an enterprise licence.

6. Scope tied to the engine, not to promises

What is deployed, for which workflow, with which integrations, and what is out of scope. Pricing attached to that scope. Changes in writing. A contract that specifies the engine and the baseline is one you can hold a vendor to, including us.