Here is the answer nobody in the sales process will give you plainly: if AI misses something material, you own it. Not the vendor. The liability sits with the professional who relied on the output and the principal who acted on the advice, and no contract available in this market meaningfully changes that. Everything worth doing about it happens before the miss, not after.
Why the loss lands on you
Three structural reasons, none of which are negotiable in practice.
The liability cap is smaller than the deal. Software agreements cap damages at fees paid, or a multiple of them. A screening subscription might cost five figures. A missed change-of-control provision in a mid-market acquisition can cost seven. Even an uncapped indemnity from a small vendor is worth only what the vendor's balance sheet can satisfy, which in this category is usually not much.
The duty is non-delegable. ABA Formal Opinion 512, issued 29 July 2024, holds that lawyers using generative AI must “fully consider their applicable ethical obligations,” including competence and supervisory responsibility.[1] A professional duty is not discharged by selecting a vendor. It attaches to the work product that leaves your desk. The same logic applies to a deal principal's duty to their investment committee or their limited partners.
The miss is discovered by the person harmed by it. This is the one that matters most operationally. A missed provision does not surface at the deal table. It surfaces when a counterparty exercises it, or when an indemnity claim arrives, or during the next transaction's diligence — typically well after close, when the purchase price is spent and the leverage that would have priced the risk is gone.
The specific misses that hurt
Not all omissions are equal. The ones that cause real loss cluster in a recognisable shape: provisions that are individually unremarkable but change the economics of the transaction when triggered.
| Missed item | Why AI misses it | When it surfaces |
|---|---|---|
| Change-of-control in a mid-tier customer contract | Buried in a document nobody prioritised for review | At closing, or when the customer exercises |
| Cross-default linking two unrelated facilities | Requires reading two documents together | When the first covenant trips |
| An obligation in an amendment, not the base agreement | The amendment was filed separately or late | Post-close, on a compliance review |
| Most-favoured-nation pricing in one customer contract | Individually minor, economically material at scale | When pricing changes for anyone else |
| An obligation attributed to the wrong subsidiary | Entity drift across a corporate group | On integration, when the entity map is rebuilt |
Look at the middle column. Almost none of these are model intelligence failures. They are coverage and prioritisation failures — the document was never properly queued, or the connection required reading two things together, or a minor-looking item was correctly deprioritised by a rule that did not understand its economic weight.
That distinction determines where to spend your mitigation effort. Buying a smarter model does not fix a document that was never read.
The two failure modes, and only one is defensible
When a miss is investigated afterwards — by an insurer, a claimant, or your own committee — the question resolves to one of two findings:
Not defensible: the document was never processed, or the category was never covered, and nothing in the record indicates anyone knew that.
The first is a judgment that turned out wrong, which is an ordinary professional risk that processes and insurance are designed to absorb. The second is an unmanaged gap, and it is very hard to characterise as anything else.
Almost all of the protection available to you lies in ensuring that misses land in the first category. That is not a matter of accuracy. It is a matter of whether your process can demonstrate what it covered.
What actually reduces exposure
Make absence of a finding distinguishable from absence of a search
This is the single most important control and the most commonly missing one. In most report formats, “we examined this and found nothing” and “we never looked at this” render identically — as blank space or as a comfortable zero. A system that assigns a category an explicit non-zero floor when nothing was found, and says so on the face of the report, converts an invisible gap into a visible one. That is the difference between the two boxes above.
Score every category on every document, including the boring ones
Coverage failures happen because attention is allocated by folder structure and upload order, which correlate with nothing. A screening stage that scores the entire set against every defined category means no document is unexamined — some are simply examined and ranked low. Those are very different positions to be in when a miss is investigated.
Set thresholds in writing, before the deal
Decide in advance what severity escalates to human review and what does not. A pre-committed threshold turns “we didn't catch it” into “it scored below our documented escalation line” — a defensible risk-appetite decision rather than an oversight. Set after the fact, it looks like a rationalisation, because it is one.
Retain the run record
Document set, categories applied, scores, thresholds, supporting quotes, model version, timestamp, and every human override. This is what allows you to reconstruct, months later, exactly what was known and when. Without it you cannot prove the first box even if it is true.
Keep the human call on materiality
Whether a flagged provision kills the deal is not a question a screening system should be answering, and a process that lets it answer one has moved the decision — and the exposure — to the layer least able to defend it.
What to negotiate, knowing it will not save you
Contractual protection here is real but modest. Ask for it with clear eyes about its size.
- No carve-out for AI output. This matters more than the cap. A vendor who warrants the software but specifically excludes what the software produces has told you where they believe the risk sits. Push on the carve-out, not the number.
- Contractual prohibition on training with your data. Not a policy statement — a term. Policies change without your signature.
- Disclosure of coverage limits. Document length ceilings, unsupported formats, languages handled poorly. You cannot manage a limit you were not told about, and silent truncation is common.
- Model versioning and record retention. If the system that produced your report no longer exists, neither does your ability to reconstruct it.
- Check your PI policy for an AI exclusion. Do this before the deal, not after the claim. This is a fast call to your broker and it occasionally produces a very unwelcome answer.
The honest framing for your committee
Do not present AI-assisted diligence as reducing the risk of missing something to zero. It does not, and the first miss will destroy the credibility of the claim and of everything else in the pack.
Present it as what it defensibly is: a way to ensure every document is examined against every defined category, so that scarce senior attention is spent on the highest-risk material rather than on whatever happened to be in the first folder. The failure mode it removes is the unexamined document. The failure mode it does not remove is the misjudged one.
That is a smaller claim than most vendors make. It is also the one that holds up when something goes wrong — and it is worth a great deal, because the unexamined document is where the expensive misses actually come from.
Bottom line
You will carry the loss, so optimise for the position you want to be in when it happens. That position is: the document was read, the category was scored, the provision was ranked below a threshold you set in advance and wrote down, and a named human made the call.
Get there and a miss is a judgment that went wrong. Fail to get there and it is a gap nobody was managing — and those are two very different conversations to have with a claimant.
Sources
- American Bar Association Standing Committee on Ethics and Professional Responsibility, Formal Opinion 512: Generative Artificial Intelligence Tools, 29 July 2024. americanbar.org
This page describes general risk-allocation practice and is not legal advice. We cite only sources we have retrieved and read — see our methodology.