The Legal AI 'Carve-Out Blindness' Problem

Your AI Contract Review Tool Is Missing the Most Important Language in the Deal

Here is a problem that is not theoretical. A mid-market M&A associate uses an AI contract review tool to flag issues in a stock purchase agreement. The tool produces a clean summary. Indemnification caps flagged, rep and warranty baskets flagged, governing law noted. The associate reviews the summary, catches a few additional issues on a quick read, and sends comments to the other side. Six months after closing, a fraud claim surfaces. The buyer's counsel reaches for the liability cap carve-out — and finds that the language the seller's team inserted, quietly and without much fanfare, strips the carve-out of any practical effect in the exact scenario now in dispute. Nobody caught it. The AI did not flag it. The associate did not know to look.

This is carve-out blindness, and it is one of the most structurally serious problems with the current generation of AI contract review tools.

Why Carve-Outs Are Not Boilerplate

There is a common misconception, apparently baked into the training data of most commercial AI review tools, that carve-outs are subordinate language — the fine print after the real clause. They are not. Carve-outs are where deals actually get made.

The fraud carve-out from a liability cap is not a formality. In SPA negotiations, the battle over whether the cap applies to intentional misrepresentation, fraudulent inducement, or common law fraud — and how "fraud" is defined for that purpose — is often the hardest-fought language in the entire indemnification article. See RAA Management, LLC v. Savage Sports Holdings, Inc., 2010 WL 975629 (Del. Ch. 2010), which turned substantially on how fraud-based claims intersected with contractual indemnification frameworks. The difference between "fraud by the party" and "fraud by any representative of the party" is worth millions of dollars in post-closing litigation. AI tools trained to recognize liability caps as risk items consistently fail to treat the carve-out from that cap with equivalent analytical weight.

The same logic applies to IP ownership provisions. A standard work-for-hire clause in a technology services agreement reads cleanly. The carve-out for the vendor's pre-existing IP, retained tools, or general methodologies is where the real negotiation happens — and where the real risk sits. Courts have grappled with this repeatedly. When carve-out language around pre-existing IP is ambiguous, disputes follow: Aymes v. Bonelli, 980 F.2d 857 (2d Cir. 1992) being an early marker, but modern SaaS agreements generate similar fact patterns constantly. An AI tool that flags the work-for-hire operative clause but treats the background IP carve-out as standard is identifying the frame while missing the painting.

Regulatory carve-outs from exclusivity clauses are perhaps the most underappreciated example. In distribution agreements and licensing deals involving healthcare, financial services, or defense contracting, the exclusivity grant typically contains carve-outs for regulatory requirements — the licensee can deal with competitors if a government contract or regulatory directive requires it. How that carve-out is scoped, whether it requires notice, whether it triggers a most-favored-nation adjustment, whether "regulatory requirement" is defined narrowly or broadly — all of it is heavily negotiated. AI tools that summarize the exclusivity grant and note the regulatory carve-out as "standard" are making a legal judgment they are not equipped to make.

The Structural Problem: Trained on Volume, Blinded by Frequency

The reason AI tools have this problem is not accidental. It is architectural.

Large language models used in legal review are trained on high volumes of contract language. Carve-outs appear frequently. They appear in similar syntactic positions — after "except," "provided that," "notwithstanding the foregoing." Because they appear frequently and in patterns, models learn to recognize them as a type of clause rather than as substantive negotiating positions. Frequency becomes a proxy for standardization, and standardization becomes a proxy for low risk. The model's behavior reflects the training signal: carve-outs are common, carve-outs are structural, carve-outs are noise.

This is exactly backwards. Carve-outs are common because they are important, not in spite of it. The sophistication of transactional lawyers over decades has produced a drafting culture where contested issues are resolved through exception language. The model interprets the output of that culture — widespread carve-out usage — as evidence that carve-outs are routine. It is a fundamental inference error, and it has not been corrected in any major commercial tool available as of this writing.

What This Costs Transactional Lawyers

The practical cost falls hardest on mid-level associates and in-house counsel who inherit AI-reviewed drafts. When you receive a contract with an AI-generated summary or issue list, you implicitly trust that flagged items are the risk items. The cognitive load shifts from independent review to validation of the AI's output. That is exactly what these tools are sold to produce — efficiency through triage. But when the triage systematically deprioritizes carve-out language, you inherit a draft where the operative clauses have been scrutinized and the exceptions have been accepted in silence.

The Uniform Electronic Transactions Act and the CFPB's increasing scrutiny of AI-assisted contract processes (see the CFPB's 2025 guidance on AI in consumer financial agreements) both reflect a regulatory environment that is beginning to take seriously how automated tools shape contractual outcomes. Neither framework has fully addressed the transactional lawyer's reliance problem, but the direction is clear.

Fix the Workflow, Not Just the Tool

Until the tools improve — and they will need to be retrained with carve-out language explicitly weighted as primary rather than subordinate — the workflow fix is manual and non-negotiable. Treat every carve-out from a defined term or operative clause as independently reviewable. Build checklists that require separate sign-off on exception language in liability caps, IP provisions, exclusivity grants, and non-compete definitions. Do not let an AI summary substitute for clause-by-clause review of exception language.

The fraud carve-out is not a footnote. In the right dispute, it is the entire case. AI tools that cannot see that are not just incomplete — they are creating liability for the lawyers who rely on them.