An AI vendor due diligence checklist for legal teams
By Asım Serdar Yılmaz · Published · Last updated · 2 min read
I have sat on both sides of this table, having built AI tooling for a legal team in-house and, on other days, signed off on buying it from vendors. What separates a good demo from a safer contract is a short list of questions. What matters as much as the answers is where they end up, which is in the contract itself, under audit rights, breach notice and liability.
The risk here is at least measurable, which is more than can be said for most of what keeps a general counsel awake. Verizon’s 2026 Data Breach Investigations Report put third-party involvement in breaches at 48 percent, though that figure counts all third parties and not AI vendors. IBM’s 2025 Cost of a Data Breach report adds that among breached organizations, about one in five said the breach involved shadow AI, and that heavy shadow-AI users reported costs roughly 670,000 dollars higher than light or non-users, an association rather than a proven cause. Read side by side, the two reports describe what an unvetted AI tool actually is: a third party with unusually broad access to privileged material.
Six questions have earned a permanent place on my list.
Training data provenance. Where did the model’s training data come from, and can the vendor show it had rights to use it? A confident answer to the second half is rarer than it ought to be.
Prompt retention and reuse. Are your prompts and documents stored, for how long, and used to train anything?
Subprocessors and hosting region. Who else touches the data, how far down the subprocessor chain does that go, in which jurisdiction does it sit, and can deletion be verified rather than just promised?
IP indemnity. If an output infringes someone’s rights, who carries that?
Audit trail. Long after the fact, can anyone reconstruct who asked what, and what the system returned?
Exit and data return. When the contract ends, how does your data come back, and in what format? I have come round to thinking that question is too small on its own, so ask instead for the exhaustive portability specification the Data Act puts in the contract, because the workflow doesn’t come back with the data.
A vendor with clean answers tends to welcome the questions. Hesitation on any of them is itself an answer, and it has the merit of arriving before anyone has signed. I read it as a signal rather than a verdict, since a good vendor can simply be badly briefed on the day, and either way the answers still have to land in the contract.
Sources
Next: AI governance for in-house legal teams: five minimum controls →