The reps and warranties gap that AI document review cannot fix
AI excels at finding what is in a contract. The risk that matters most in M&A is what is missing from the contract. A practical look at the structural blind spot.
The case for AI in document review is now obvious. It identifies clauses, extracts terms, flags anomalies, and produces clean categorization output on data rooms that would otherwise take a junior associate two weeks. Every middle market deal has at least some AI in the diligence workflow now.
The case against treating it as a complete diligence tool is also obvious to anyone who has run a deal where the problem was not what was in the contract but what was missing from it. AI tools are excellent at the presence question. They are weak at the absence question. The two are not the same.
Why this matters
A contract that is silent on a material point can be more dangerous than a contract with a problematic clause. A management agreement that is silent on the consequences of a change of control may be subject to a default rule the buyer does not expect. An IP assignment that is silent on prior work may leave a developer's pre-employment code outside the target's IP estate. A commercial agreement that lacks an exclusivity provision may make the target's revenue projections meaningless if the customer is free to dual source.
The absence-of-clause question is the question a senior practitioner asks instinctively. "Where is the indemnification cap?" "Where is the limitation on assignment?" "Where is the non-compete?" Those are not questions about what the contract says. They are questions about what the contract does not say but the deal model assumes it does.
Why AI tools struggle with absence
Current AI document review tools work by extraction. They identify what is present in the document and categorize it. Their training data teaches them to recognize patterns of what change of control clauses look like, what assignment provisions look like, what indemnification packages look like. They are very good at finding those things when they exist.
What they are weak at is reasoning about what should be there but is not. That kind of reasoning requires a baseline expectation of contract content for the deal type, the industry, and the counterparty class. The tools do not natively carry that baseline. The prompts that try to elicit it ("identify any commercial contracts that lack a standard indemnification provision") get inconsistent results because the model has to first understand what "standard" means in the context of the specific deal and then perform a negative-finding analysis, which is harder than positive extraction.
What this looks like in practice
In a recent deal I worked on, the AI extracted change of control provisions cleanly from a hundred customer contracts. It flagged three with non-standard language for review. It missed the seventeen contracts that had no change of control provision at all. The default rule in the governing jurisdictions varied, and the deal team had assumed the contracts were silent because the counterparty's standard form was silent, which was wrong on more than half of them.
The fix was not the tool. The fix was the associate asking the right question. "Which customer contracts do not have a change of control provision at all?" A senior practitioner would have asked this first. A junior associate using a tool's default output might not.
What corporate counsel should do
Build the absence question into the diligence workflow as a separate pass. After the extraction pass identifies what is present, run a second pass that explicitly looks for what should be present but is not. For each contract type, define the expected provisions and check which contracts lack them.
Train associates to articulate the absence question explicitly. The instinct to ask "what is missing" is not built in. It develops through experience. Until that experience is built up, the question needs to be on a checklist.
Use AI tools for what they are good at. Categorization, extraction, anomaly detection on present provisions. Use human review for what AI is weak at. Absence detection, baseline-against-expectation reasoning, and the deal-level synthesis that determines whether the gap actually matters.
I built a small tool that addresses one slice of this. Paste the reps and warranties section of an APA and a brief description of the deal. The tool returns a list of categories that the reps appear to lack relative to a standard middle market package. It is at /tools/reps-warranties-gap/. Like the other tools on this site, it is educational, not transactional. The gap analysis is a starting point for a careful read, not a substitute for one.
The structural point is that AI tools change the cost of doing certain kinds of diligence. They do not change the questions that have to be asked. The senior practitioner who has been asking the right questions for fifteen years still asks the right questions. The new dynamic is that the cost of running an AI-assisted first pass is now low enough that the senior question can be asked sooner and applied to more contracts. That is a real improvement. It does not eliminate the need for the question.
Walter Allison is a corporate attorney in Denver. He writes here about M&A, private equity, and venture capital structure.
Follow on LinkedIn ·
Firm bio ·
More posts