How to use AI in private equity due diligence

The practical way to use AI in private equity due diligence is to give it the first pass on a short list of narrow, repeatable jobs and to require a page reference from the data room for every fact it extracts. A named person on the deal team then checks that output against the source before any of it reaches the financial model or the investment committee memo.

This article is for the vice president, principal, or associate running diligence on a live deal, and for the COO or general counsel who decides what the team may do with confidential data room material. It assumes the firm already has a general AI assistant and wants to use it properly on the next deal.

What AI is good at in diligence, and what it is not

Most deal teams have already started. In Deloitte's 2025 survey of 1,000 senior corporate and private equity leaders in the U.S., 86 percent of responding organizations said they had integrated generative AI into their M&A workflows, and 35 percent of those adopters were applying it to due diligence. Bain's survey of 300 M&A executives found that 45 percent of executives used AI tools in M&A in 2025, more than double the prior year, and that about one-third were using AI systematically or redesigning processes for it. The gap between those two Bain figures is the subject of this article, because using a tool on a deal and running a defined process on every deal are different things.

Much of what is published on this topic comes from diligence software vendors, and a speed figure on a vendor page is a vendor claim. We would set those figures aside and look at what the work is.

AI is good at reading a lot of documents against a fixed set of questions. It can pull the same twelve fields out of ninety customer contracts, compare a customer list to a revenue file, and turn a folder of board minutes into a dated list of decisions. It is also useful earlier in the funnel. Bain's 2024 private equity report describes one fund whose professionals look at ten deals to find one worth investigating further and spend a full day on most of those looks, and says generative AI can bring screening time per company down from a day to an hour.

AI is weakest where diligence matters most. It cannot judge whether the management team is telling you the whole story, it does not form commercial conviction, and it has no way to know what should be in the room and is missing. A model reads what it is given. The deal team's job on those three questions does not change.

Which diligence jobs are worth building once and reusing on every deal?

"Have AI read the data room" is too broad an instruction to check. A set of narrow jobs, each with a defined input, a defined output, and a defined reviewer, is something a firm can run the same way every time. These four are where we would start.

Job	What goes in	What comes out	Who checks it
Contract abstract	Customer, supplier, and employment agreements with all amendments	A row per contract with the firm's own fields: change of control, assignment, termination, exclusivity, most favored nation, each with a page reference	Deal counsel or the associate on legal diligence
Customer concentration tie-out	Customer contracts and the seller's revenue by customer file	A table matching each top customer to its contract, with mismatches in name, term, or pricing flagged	The associate who owns the model
Quality of earnings question list	The seller's financial package and the document index	A list of gaps and open questions for the seller and the accounting advisor	The VP running financial diligence
IC memo first draft	Reviewed abstracts, the tie-out, and the firm's memo template	Draft sections in the firm's own format, with sources noted	The principal who signs the memo

Each of these can be packaged once as a skill. Anthropic describes skills as folders of instructions, scripts, and resources that Claude loads dynamically for specialized tasks, and says that on Team and Enterprise plans an organization owner can provision skills for all users. For a deal team, that means the abstract template, the field definitions, and the review checklist live in one place, and deal ten runs the same way as deal one. Our plain guide to what Claude skills are covers the mechanics.

How do you set up a contract review that cites its sources?

The setup follows the same pattern as lease abstraction for brokerage teams, applied to a different document type.

1. Start from the data room index. Work one folder at a time so you know which documents the model has seen and which it has not. Tools have per-request limits. Anthropic's PDF documentation, for example, lists a maximum request size of 32 MB (it notes this varies by platform) and a maximum of 600 pages per request, or 100 on models with smaller context windows, so a full room does not go in at once.
2. Give the model the firm's own field list with a definition for each field. "Change of control" should say whether you mean a consent right, a termination right, or a notice obligation, because contracts use all three.
3. Require a document name, a page number, and a word-for-word quote for every field. Anthropic's guidance on reducing hallucinations recommends the quoting part of this: have the model extract word-for-word quotes first, cite a supporting quote for each claim, and retract any claim it cannot support. The document name and page number are what let a reviewer find that quote quickly.
4. Tell the model to write "not found" when a term is absent. The same guidance says that giving the model explicit permission to admit uncertainty can reduce false information. A blank field a person can chase is better than a plausible guess.
5. Include every amendment and side letter with its parent contract, in date order, and ask for a list of what each one changed.

Tying numbers out: where AI output meets the model

The tie-out is where extraction meets the numbers, and it deserves the most care. The AI can match each top customer in the revenue file to its contract and flag where the legal entity, the term, or the pricing does not line up. It should not be the thing that computes concentration or builds a bridge. Arithmetic belongs in the spreadsheet, where a formula can be audited, and the AI output should arrive as inputs with a source beside each one.

A workable rule is that no AI-extracted figure is typed into the financial model without its document and page reference in the adjacent cell, and no figure is used until the associate who owns that model has opened that page.

Drafting the IC memo: first pass by AI, judgment by the deal team

Memo drafting works when the model is given reviewed material and the firm's own template, and is told to use nothing else. The sections that summarize what the documents say (business description, contract overview, customer table, open diligence items) are reasonable first drafts for a model. The investment thesis, the view on management, the risks the team chooses to underwrite, and the recommendation are written by the people who will defend them in the room.

The review step: who checks what before it counts as done?

The review step is the part of the workflow that makes the output usable. NIST's Generative AI Profile defines confabulation as generative AI systems generating and confidently presenting erroneous or false content, notes that outputs may include confabulated logic or citations that purport to justify the answer, and says these risks may be especially important to monitor when the technology is integrated into applications involving consequential decision making. An investment committee vote is a consequential decision. Anthropic's own guidance says its techniques for reducing these errors do not eliminate them and that critical information should always be validated.

So a page reference on its own proves nothing until someone opens the page. We would set three rules.

• Every dollar amount, date, and consent or termination term is checked against the cited page by a named person. Descriptive fields can be sampled.
• An item is closed when the workstream lead has reviewed it. The tool producing an answer does not close it.
• The tracker records who reviewed each item and when, so the memo can be traced back through a person to a page.

Our article on designing the human review step goes further into how to size the check to the risk.

Confidentiality: where can data room documents go?

Confidentiality is a setup question the firm can answer in advance. It also sits at the top of what dealmakers worry about: in the Deloitte survey, 67 percent of respondents named data security as a leading concern.

Three things should be settled before the first document is uploaded.

• The workspace. Diligence material belongs in the firm's own enterprise AI workspace under its commercial terms. Anthropic, for example, states that by default it does not use inputs or outputs from its commercial products, such as Claude for Work and the API, to train its models, with exceptions when a customer explicitly reports feedback or chooses to allow it. Terms differ by vendor and by plan, which we cover in what happens to data you send to an AI model. Nothing from a data room belongs in a personal consumer account.
• The NDA. Have counsel read each confidentiality agreement for restrictions on sharing information with third parties, on who counts as a permitted representative, and on return or destruction of materials at the end of the process. Whether a cloud AI vendor fits within those terms is a legal question for your counsel, and the answer can differ from deal to deal.
• The firm's own policy. If the firm is an SEC-registered adviser, note that the Division of Examinations' fiscal year 2026 priorities say it will assess whether firms have implemented adequate policies and procedures to monitor or supervise their use of AI technologies. The document is a staff statement of examination priorities and says it creates no new obligations, so your compliance lead should decide what it means for a written diligence procedure.

Turning one deal's setup into the firm's standing diligence kit

After the first deal, keep what worked: the field list and definitions, the abstract and tie-out templates, the review checklist, and a short note on what the model got wrong and where. Package each job as a skill in the firm's workspace, assign an owner, and update it after each deal. A firm that does this starts each new deal from the last deal's corrections, which leaves the team more of the exclusivity period for management, the market, and what is missing from the room.

If it would help to talk through the setup for your next deal, you can get in touch.

Common questions

Can AI review an entire data room?

It can read a large share of the documents, but not in one pass and not unattended. Tools have per-request limits on file size and page count, so the room has to be worked folder by folder against a document index, and a person still has to decide what matters and notice what the seller did not post.

How accurate is AI at extracting contract terms in due diligence?

There is no single honest number, because accuracy depends on scan quality, how the fields are defined, and how many amendments sit on top of each contract. The useful test is to run your abstract on contracts your team has already reviewed by hand and compare field by field. Whatever the result, every extracted term should carry a page reference so a reviewer can check it quickly.

Is it safe to upload confidential data room documents to an AI tool?

That depends on the plan, the vendor's data terms, and the confidentiality agreement on the deal. The work belongs in the firm's own enterprise AI workspace under its commercial terms, and never in a personal consumer account. Have your general counsel or compliance lead confirm the terms before documents go in.

Does using AI in diligence violate an NDA?

It depends on the wording of that NDA. Some restrict sharing confidential information with third parties or require that anyone who receives it be bound by the same terms, and a cloud AI vendor may or may not fit within those terms. Counsel should read the agreement with this question in mind on every deal.

How do you check AI output before it goes into an investment committee memo?

Require a document and page reference on every fact, have a named person open the source for every dollar, date, and consent term, and sample the rest. Nothing moves into the financial model or the memo until that person has marked it reviewed.