What Vendor Guarantees Should I Require for AI Due Diligence?

Updated August 2026 · 6 min read · Deal Room Intelligence Series

Start by conceding the one you cannot have

Stanford researchers evaluated the retrieval-grounded legal AI products sold by LexisNexis and Thomson Reuters and found they “each hallucinate between 17% and 33% of the time,” concluding that providers' claims of hallucination-free output are overstated.^[1]

If the two incumbents of legal information, with proprietary curated corpora and enterprise engineering budgets, cannot guarantee accuracy, no smaller vendor can. A vendor who offers such a guarantee is either not going to honour it or has not read the room.

Conceding this early is tactically useful. It moves the conversation from a promise the vendor cannot make to terms they can, and it signals you have done the reading.

Tier 1 — Guarantees about mechanical behaviour

These are the highest-value asks because they are verifiable, non-probabilistic claims about what the system does, not predictions about judgment. A vendor who declines even these is telling you something important.

Quote fidelity

That any source text displayed with a finding is a faithful reproduction of the underlying document — not paraphrased, not reconstructed. This is the foundation of cheap verification, and it is a mechanical property a vendor can stand behind.

Coverage accuracy

That the coverage report accurately reflects what was processed: documents submitted, processed in full, partially processed, and failed. This is the guarantee that converts silent truncation from your problem into theirs, and silent truncation is one of the most common real-world failures.

No silent truncation

That any document exceeding a processing limit is reported as such rather than analysed partially and presented as complete. Related to the above, and worth stating separately because it is the specific behaviour that hurts — the credit agreement is both the longest document in the room and the one that matters most.

Determinism or disclosed variance

Either the same document produces the same result, or the vendor discloses the variance and its cause. You can work with either. You cannot work with not knowing.

Tier 2 — Guarantees about data

Training prohibition, contractual, with flow-down

Binding on the vendor, subprocessors and the underlying model provider. Watch for carve-outs covering telemetry, evaluation sets and human review of outputs — that is where content actually escapes. “Not by default” is a setting, not a term.

Named subprocessors with change notification

Several parties see your documents. A vendor who cannot produce the list has not mapped their own data flow — a worse finding than any individual name.

Deletion on demand and on termination

Stated timeline, with an answer about backups. Deal documents stop being needed at close and become pure exposure afterwards.

Region pinning where required

Contractual, not best-efforts, where your NDAs or data-protection obligations demand it.

Tier 3 — Guarantees about continuity and reconstruction

Model and ruleset versioning

With the version stamped on every stored result. Without this you cannot reconstruct how a conclusion was reached, and reconstruction is what makes reliance defensible.

Run-record retention and export

Findings, scores, thresholds in force, source quotes, overrides — retained for a stated period and exportable in a portable format. Otherwise changing vendors means abandoning your audit trail.

Notice before a material model change

Not a promise never to change — that would be unreasonable — but notice, so you can note the boundary in your own records and, if a deal is live, finish it on a stable version.

Tier 4 — Commercial guarantees

The warranty carve-out, which outranks the cap

Most negotiations here burn effort on the liability cap. No cap a software vendor will accept is sized for a deal-scale loss, so moving it from one year's fees to two changes nothing real.

Read the exclusions instead. A common structure warrants that the software performs materially in accordance with documentation, then separately disclaims any warranty as to accuracy, completeness or fitness of AI-generated output. Together, that warrants the delivery mechanism and disclaims the product.

What no guarantee can do

ABA Formal Opinion 512 (29 July 2024) holds that lawyers using generative AI must “fully consider their applicable ethical obligations,” including competence and supervisory responsibility.^[2] That duty is non-delegable. No vendor term moves it and none attempt to.

So the honest framing of this whole exercise: you are not buying protection from the consequences of a miss. You are buying disclosure, verifiability and reconstructability — the properties that let you show a managed process afterwards. Negotiate for those and you will get most of them.

A prioritised ask list

1. Training prohibition with flow-down — walk if refused
2. Quote fidelity and coverage accuracy
3. No silent truncation
4. Model versioning with run-record retention and export
5. Named subprocessors, deletion on termination
6. Confidentiality indemnity
7. Determinism disclosure
8. Notice before material model changes
9. Price lock and volume carry-forward
10. The liability cap — last, and do not spend leverage here

Bottom line

Stop asking for accuracy guarantees; they do not exist and asking marks you as a buyer who has not tested the category. Ask instead for guarantees about behaviour — that quotes are faithful, that coverage reporting is true, that nothing is truncated silently, and that the run is reconstructible months later.

Those are the terms a vendor can honestly give, and they happen to be the ones that matter when something is missed.

Check our output against these guarantees →