AI in M&A diligence, what it does well and what it does not
AI document review has moved from novelty to default in middle market diligence. A practical look at where the tools earn their keep and where they fail loudly.
Five years ago, AI document review in M&A was a novelty product. Two firms in your city had a Kira subscription, one partner ran a demo at a retreat, and nothing changed. Today, AI document review is a default expectation in middle market diligence. The tools are real, the cost savings are real, and the skill gap between associates who use them well and those who do not is widening.
This is a short read on what these tools actually do well and what they still cannot do.
What they do well
Categorization at scale. Given a data room of five thousand documents, current AI tools can sort them into reasonable buckets (employment, IP, real property, commercial contracts, financial, regulatory) with high enough precision that a junior associate's review starts from a much smaller pile. Litera's Kira, Spellbook, Hebbia, and Luminance all do this competently. Harvey, OpenAI's legal-focused product, focuses more on drafting and research than document classification but the categorization layer is solid.
Provision extraction. Once a contract type is identified, the tools reliably extract change of control clauses, assignment provisions, exclusivity terms, term lengths, governing law, and other defined-term-driven clauses. The accuracy on these extractions in 2026 is well above the threshold where a human associate's pass adds value as a check rather than a redo.
Anomaly detection. The tools flag clauses that are unusual relative to the rest of the corpus. A buyer's counsel looking at a hundred customer contracts can quickly find the seven contracts with non-standard termination rights or unusual indemnification language.
What they do not do well
Strategic judgment. The tools surface that a change of control clause exists. They do not tell you whether that clause matters for the deal at hand, whether the counterparty is likely to enforce it, or whether the buyer's leverage with the customer is sufficient to negotiate a waiver. That work is still human work.
Cross-document synthesis. The tools handle a single document well and a category of documents reasonably. They are still weak at the kind of reasoning that says "the customer concentration shown in the financial diligence corresponds to three contracts with assignment restrictions and an MFN clause, and that combination creates real consent risk." Stitching multiple findings into a deal-level recommendation is associate work.
Privileged or sensitive document handling. Most current tools require uploading client documents to the vendor's infrastructure. Some have on-premise or virtual private cloud options. The privilege analysis on uploading client documents to a third party reviewer is not new, but it bears thinking about every time and the answer should be reflected in the engagement letter.
Edge cases the model has not seen. The tools are trained on large corpora of commercial agreements. They are weaker on industry-specific contracts, on contracts written in non-standard formats, on contracts that have been heavily amended over time, and on contracts in less common jurisdictions. Cross-border deal documents are a known weak spot.
What this means for the associate role
The cleanest description I have heard: AI moves the associate from "read everything" to "review what the model flagged." That is not a smaller job. It is a different job. The skill set shifts toward quality review of model output, calibration of confidence thresholds, prompt engineering for the categorization taxonomy, and synthesis across the flagged items.
Firms that have invested in the tools and the workflow are seeing genuine cost savings on diligence-heavy matters. Firms that have purchased licenses but not changed the workflow are paying for tools that nobody uses. The difference is training and process, not the technology.
A small example
I built a couple of small AI-assisted tools on this site. The term sheet decoder pastes a venture term sheet and returns a clause-by-clause plain English walkthrough. The diligence question generator takes a deal description and produces a structured first-round diligence list. They are educational, not transactional, and they exist to demonstrate that the gap between what a non-lawyer needs to understand before signing and what AI can comfortably provide is now small.
The tools are at /tools/. They are free. They do not save anything you submit. They are not a substitute for actual counsel. They are a reasonable picture of where the technology sits in 2026 and where most of the practical value lives for corporate transactions: educational gap closing, first-pass document work, and structured generation of repetitive content.
The work that survives all of this is the work that requires judgment about how a deal should be done. That work pays better than ever.
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