AI in M&A Due Diligence: Contract Review, QA Sampling, and Clean-Team Risks

A frequently quoted estimate, from a 2011 Harvard Business Review article by Clayton Christensen and colleagues, puts the failure rate of mergers and acquisitions between 70% and 90%. The figure depends heavily on how failure is defined, so treat it as a rough range rather than a measurement. Many of the causes trace back to diligence: a change-of-control clause that let a major customer walk, a revenue line that turned out to be one-off, an integration cost nobody estimated.

AI does not fix bad strategy or overpaying. What it does well is the reading. A mid-sized target can have several thousand contracts, and a traditional review samples them. AI makes it practical to review all of them, which changes what diligence can find. This guide covers where AI helps, how to check what it misses, and the confidentiality rules that matter more, not less, once AI is involved.

Where AI helps in diligence

Contract review

The most mature use. Tools trained or prompted to find specific provisions can go through a full data room quickly:

• Change-of-control and anti-assignment clauses (can the customer or supplier terminate or renegotiate after the deal?)
• Exclusivity, non-compete, and most-favored-nation terms that could bind the combined company
• Termination for convenience and notice periods
• Liability caps, indemnities, and unusual warranties
• Auto-renewals and price escalators

Output is a structured table per contract, which lawyers then review, starting with the largest customers and suppliers.

Financial diligence support

Extracting data from invoices, contracts, and ledgers to test revenue recognition, customer concentration, churn, and pricing trends for the quality-of-earnings work. AI speeds up the data work; the accountants still make the judgments. The claims data shows where the money is: in Aon's 2026 claims study, financial statement breaches account for 38% of paid losses on representations and warranties (R&W) insurance policies placed since 2019, and material contracts for 21%. Those are the two areas the extraction and contract work in this section targets.

Other documents

Leases, employment and equity agreements, IP assignments, licenses, and litigation files all benefit from the same extract-and-summarize approach. Open-source license detection in the target's codebase is a well-established automated check for software deals. Security history matters too: the FTC's 2024 action against Marriott and Starwood describes a Starwood breach that began around July 2014 and went undetected until September 2018, two years after Marriott acquired the company.

Target screening

Scanning company databases, filings, patents, and news to find targets that fit a thesis. Useful for widening the funnel, but the shortlist still needs judgment.

Measure what the tool misses

Reviewing every contract with AI is only better than sampling if you know how often the AI is wrong. The simplest check uses the same sampling logic that auditors use.

Here is an illustration. A data room holds 4,000 contracts. The AI flags 380 as containing a change-of-control provision. Lawyers review all 380 (catching false positives) and a random sample of 200 of the 3,620 unflagged contracts. In that sample, they find 3 contracts with a change-of-control clause the tool missed.

• Estimated miss rate: 3 ÷ 200 = 1.5%
• Applied to the unflagged pool: 1.5% × 3,620 ≈ 54 contracts likely missed
• A 95% confidence interval for a 3-in-200 result runs from about 0.5% to 4.3%, or roughly 20 to 155 contracts

Whether that is acceptable depends on what the missed contracts could be. If the unflagged pool is mostly small vendor agreements, perhaps yes. If it includes customer contracts, the team might improve the prompts or rules, re-run, and sample again, or review the unflagged customer contracts manually. Either way, the deal team is making the decision with numbers rather than trusting the tool's confidence. Document the method; it matters if there is a warranty claim later. Timing is a reason to be thorough: Aon reports that about 51% of R&W claims are notified more than 12 months after closing, often after the seller escrow has expired. In the SRS Acquiom 2026 study, 88% of 2025 private-target deals had an escrow or holdback, with a median of 10.0% of transaction value without R&W insurance and 2.8% with it, so there is little seller money to recover from for a problem found late.

Confidentiality: the risk that grows with AI

Deal information is some of the most sensitive data a company handles, and AI adds new ways to leak it.

• NDA terms. Many NDAs restrict sharing information with third parties. An AI vendor processing data room documents may count as one. Check the NDA and get the target's agreement for the specific tool if needed.
• Enterprise terms only. Use tools whose contracts exclude training on your inputs, set retention limits, and state where data is processed. Consumer AI accounts are out.
• Inside information. For listed companies, deal information is often material non-public information. Access logs for AI tools should be as tight as for the data room itself.

Clean teams

When buyer and target compete, antitrust rules restrict what can be shared before closing. Competitively sensitive information, such as customer-level pricing, margins, and strategic plans, typically goes only to a clean team of outside advisers and a few insulated employees. Sharing it more widely, or acting on it before closing, risks "gun-jumping" violations. The FTC's guidance on pre-merger diligence says clean teams should not include anyone responsible for competitive planning, pricing, or strategy. In January 2025 the FTC announced a record $5.6 million gun-jumping penalty against XCL Resources, Verdun Oil, and EP Energy for a 94-day HSR violation. The filing threshold itself moves each year: for 2026, the minimum size-of-transaction threshold for an HSR filing is $133.9 million (effective February 17, 2026).

An AI tool indexed over the whole data room does not know about the clean-team boundary. A business-side user asking a general question could get an answer drawn from clean-team documents. Set up separate indexes and access controls that match the clean-team rules, and have antitrust counsel sign off on the configuration.

Culture: be skeptical of scores

Some tools offer a "culture fit" score from employee reviews and public communications. The data is thin: reviews come from a self-selected minority, often recent leavers, and vary by team and location. Use text analysis to spot themes worth asking about, such as repeated complaints about a particular leadership style or unit, then do the real work through management interviews, retention and engagement data, compensation and incentive structures, and how decisions are made. Our M&A integration playbook covers what to do with those findings after closing.

After closing: tracking synergies

AI is useful post-close for comparing customer lists to find overlap and cross-sell candidates, finding duplicate vendors and software licenses, and tracking synergy delivery against the deal model. The discipline is the same as in diligence: every synergy number should trace back to a source and an owner.

Getting started

1. Pick one high-value review type, usually change-of-control and assignment clauses, and run it on the next deal.
2. Build the sampling check into the workplan from the start, including who reviews the sample.
3. Clear the tool with the NDA and counsel, and set access controls that follow any clean-team arrangement.
4. Keep humans on judgment calls: materiality, valuation, negotiation, and anything going into the disclosure schedules.

For the full diligence process, see our M&A due diligence guide. For evaluating private equity funds and managers, see private equity due diligence, and for deal strategy, strategic M&A for business growth.

This guide is for informational purposes only and is not legal, financial, or antitrust advice. Consult qualified counsel and advisers on any transaction.