Two different jobs get lumped together
"Reviewing a data room" actually covers two different jobs. The first is retrieval: finding every document that mentions a specific counterparty, clause type, or date range across thousands of files. The second is judgment: deciding whether what was found changes the deal terms, the price, or the risk allocation in the purchase agreement.
AI document analysis is good at the first job and should never be asked to do the second alone.
Retrieval, done well, saves real time
In a hypothetical mid-market M&A process, imagine a target with 15 subsidiaries, each with its own set of supplier agreements. The buyer's counsel needs to know which of those agreements contain most-favored-nation clauses. Searching that manually means opening every contract in every subsidiary folder. An index built across the whole data room, with search that understands the question rather than just matching keywords, turns that into a single query with a list of citations to check.
The same applies to building a first-pass issues list for a legal diligence memo, or pulling every change-of-control provision into one table before the negotiating team decides how to handle them in the SPA. These are retrieval-heavy, judgment-light tasks. That is exactly where automation earns its keep.
Judgment still belongs to the deal team
Deciding whether a disclosed litigation matter is material, whether an indemnity cap is adequate given the risk profile, or how to price in an undisclosed liability once it surfaces: none of that should be delegated to a model. These decisions depend on context an AI tool does not have (the fund's risk appetite, the relationship with the counterparty, what else is happening in the deal) and carry consequences that require a named, accountable person behind the call.
The practical split: let AI produce the first pass, complete with citations back to source documents, and treat that output the same way a partner would treat a junior associate's draft. Useful, fast, but reviewed before it goes anywhere near a negotiating table.
A due diligence checklist question worth adding
When evaluating any AI tool for a data room, ask specifically whether it distinguishes between "I found this text" and "this text means X for your deal." A tool that blurs that line, presenting inference as if it were extraction, is more dangerous than no tool at all, because the output looks authoritative even when it is wrong. A tool that keeps the citation visible next to every claim makes it easy for a human to catch the difference. That distinction is worth testing directly in a proof of concept, not just taking on faith from a sales deck.