AI for Private Equity Due Diligence: Workflow, Costs, and Limits

A mid-size virtual data room runs past 2,000 documents, and a deal team has weeks, not months, to read it. AI now handles the reading-and-extraction layer of private equity due diligence well: indexing the room, pulling clauses with citations, and cross-checking what management claims against what contracts say. It does not yet replace the judgment calls, and it fails in specific, predictable ways you need to design around. This post lays out the workflow that works, what it costs in real API spend, where the models break, and what the SEC expects you to document about it.

What AI does well in a data room today

The honest answer is: four jobs, all in the reading-and-matching layer. None of them is "do the diligence."

• Contract abstraction with citations. Given a folder of customer agreements, extract the clauses that matter to the deal (termination for convenience, change of control, exclusivity, auto-renewal) and return each one with the exact quote, the document name, and the page number. Anthropic's documentation explicitly recommends requiring verbatim citations and returning a "not found" response when the source does not support a claim, and both behaviors are configurable in the prompt (Anthropic, reducing hallucinations).
• Concentration and consistency tie-outs. Compare what the CIM says about customer concentration against what the top-20 contracts actually say, and flag mismatches. This is pattern matching at a scale humans are bad at: one practitioner describing a Claude-based diligence workflow notes that a change-of-control provision can sit buried inside a renewal summary, and a model reading every page catches what a skimmed summary misses (Claude Formations).
• Question lists and issue logs. Turn extraction results into a structured list of follow-ups for the seller's counsel, grouped by workstream, with the source document attached to each question.
• First drafts. Produce the first draft of the diligence section of the IC memo, and the first draft of the QoE question list. A reviewer then edits rather than writes from a blank page.

The same pattern works downstream of the deal, for portfolio and LP reporting, which we covered in what AI for LP quarterly reporting costs and where it stops.

What is conspicuously absent from that list: the quality-of-earnings analysis itself, legal judgment on ambiguous clauses, and the investment decision. More on why below.

A review workflow, step by step

This is the workflow we implement for funds through our Claude for financial services practice. It assumes you are working through the API or a controlled workspace, not a consumer chat window.

1. Index the room before reading anything. Extract text from every file and build an index: document name, workstream, page count, document type guess, date. A mid-size data room often exceeds 2,000 documents, and folder names are unreliable, so the index, not the folder tree, becomes your map (Claude Formations).
2. Gap-check against a checklist. Compare the index to a standard diligence checklist for the deal type. Missing insurance certificates, missing amendments to the top customer contract, an employee agreements folder with three files for a 400-person company: these gaps are findings in themselves.
3. Run clause extraction per workstream. For each workstream, send the relevant document set with a prompt that demands verbatim quotes and page numbers, and a "not found" result rather than a paraphrase. Keep amendments and side letters attached to their parent agreements; a change-of-control clause modified by a later amendment is exactly the thing a summary pass gets wrong (King & Company).
4. Tie out claims against documents. Feed the CIM or management presentation claims ("no customer is more than 8% of revenue") against the extracted contract and revenue data. Every mismatch goes on the issue log with its source.
5. Draft, then verify. Generate the memo section and question list drafts. A human reviews every cited quote against the source page. This step is not optional, and the cost section below shows it is also not expensive.

Two technical constraints shape this workflow. Claude reads PDFs as both text and images, with a limit of 600 pages and 32 MB per request, so large files must be split, and page numbers map to the PDF page, not the printed folio (Anthropic, PDF support). And the models are not calculators: any deterministic math, such as summing revenue across contracts, should be done in code with the extracted values, not by asking the model to add (King & Company).

What a deal actually costs in API spend

Here is the worked example, using current list prices from Anthropic's pricing documentation: Claude Opus 5.5 at $4 per million input tokens, Sonnet 5.5 at $2, and Haiku 5.5 at $0.10 to $0.50 depending on prompt size (Anthropic, pricing).

Take a 2,000-document room averaging 30 pages each: 60,000 pages. Anthropic's docs put a three-page PDF at roughly 7,000 tokens when read visually, about 2,300 tokens per page, versus roughly a seventh of that for plain extracted text. Reading every page visually is therefore about 140 million input tokens per full pass.

• Full visual read, one pass, Sonnet 5.5: about $280 in input tokens. On Opus 5.5, about $560. On Haiku 5.5, $14 at the small-prompt rate and up to $70 once individual prompts exceed 100,000 tokens, which triggers the higher tier.
• Text-only pass (where extraction is clean): about 20 million tokens, so $40 on Sonnet, $2 to $10 on Haiku depending on prompt size.
• The pattern we actually recommend: index everything as text on Haiku (about $2 to $10), then deep-read visually the 10 to 20 percent of documents that matter on Sonnet. That lands around $50 to $100 per deal in input tokens, plus outputs.
• Repeated passes: prompt caching cuts re-reads of the same documents to about 5 percent of the base input price on the 5.5 models, so a second and third pass over the room cost almost nothing.

Add output tokens generously and a full multi-pass diligence run is a few hundred dollars per deal. For comparison, 60,000 pages at a very fast two minutes per page is 2,000 human hours. The economics are not subtle. What the money does not buy is correctness, which brings us to the limits.

Where the models break

Three failure modes matter in diligence, and none of them is fixed by writing a better prompt.

Spreadsheet-heavy analysis. FinSheet-Bench, a 2026 benchmark of synthetic financial portfolio data modeled on real private equity fund structures, evaluated ten model configurations from OpenAI, Google, and Anthropic across 24 spreadsheet files of varying complexity. The best model (Gemini 3.1 Pro) scored 82.4 percent, with GPT-5.2 at 80.4 and Claude Opus 4.6 with thinking at 80.2. Accuracy dropped to 48.6 percent on the largest spreadsheet, and the authors conclude no current model is reliable enough for unsupervised financial statement work (arXiv:2603.07316). Diligence is not identical to that benchmark, but any task that is "read a messy workbook and reason across tabs" sits in the danger zone.

Confabulation at scale. NIST's Generative AI Profile calls this confabulation: models "generate and confidently present erroneous or false content," including invented logic and citations that purport to justify an answer, and NIST says the risk is "especially important to monitor when integrating GAI into applications involving consequential decision making" (NIST AI 600-1). A data room is the definition of consequential. Anthropic's own mitigation guidance is the practical counter: require verbatim citations, allow "not found," and decompose big questions into per-document passes (Anthropic, reducing hallucinations).

The last 20 percent is where the deal lives. Even at benchmark scores in the low 80s, one extraction in five has an issue, and the errors are confident and well-formatted. The workflow above survives this because every claim carries a citation a human can spot-check, and because the model never does the final arithmetic or the final judgment.

Adoption data says the industry knows this. Bain's 2025 Global Private Equity Report, surveying investors representing $3.2 trillion in assets, found the majority of portfolio companies were in some phase of generative AI testing, but only about 20 percent had operationalized use cases with concrete results (Bain). Pilots are everywhere; production systems with controls are not.

The confidentiality and compliance frame

Before any document moves, three questions need written answers.

Where does the data go? Anthropic's privacy policy states that inputs and outputs from commercial products (the API, Claude for Work, Claude Gov) are not used to train models by default, with a narrow exception if a user explicitly submits feedback, which can be stored for up to five years (Anthropic privacy center). That covers the provider side. It does not cover your NDA side: many data room NDAs restrict disclosure to named representatives, and a third-party API may or may not qualify. For funds that need tighter control, a private deployment on your own cloud account through Bedrock or Vertex keeps traffic inside your perimeter; the trade-offs are laid out in our guide to private LLM deployment for banks, which applies to funds for the same reasons, and our enterprise AI practice builds exactly these governed deployments.

What will your regulator ask? The SEC's FY2026 examination priorities say the Division of Examinations will "assess whether firms have implemented adequate policies and procedures to monitor and/or supervise their use of AI technologies," including in back-office operations, and will "review for accuracy registrant representations regarding their AI capabilities" (SEC, FY2026 Examination Priorities). The document is a staff statement with no legal force and creates no new obligations, but it tells you what examiners will ask for: a written policy, a supervision chain, and marketing language that matches what you actually do. If a vendor is in the loop, the questions double; our AI vendor due diligence guide for banks and insurers is the checklist we use.

Who verified the output? Your documentation should show the human review step, the citation spot-checks, and where the model's output was rejected. That record is what turns "we used AI" into "we supervised AI" in an exam.

Common mistakes

• Running a summary pass instead of a clause pass. Summaries smooth over the exact provisions that kill deals. Extract verbatim clauses with citations; never ask for a "summary of key terms."
• Trusting the folder structure. Workstreams should come from the index you build, not the seller's folder names, which are built for the seller's convenience.
• Letting the model do arithmetic. Any number that matters gets recomputed in code from extracted values. The model extracts; the spreadsheet or script calculates.
• Detaching amendments from parents. A side letter changes the parent agreement. Extract them together or the answer will describe a contract that no longer exists.
• Dropping documents into a consumer chat tool. Consumer-tier products may train on inputs. The API and enterprise tiers do not by default. This is a one-line configuration decision with regulatory consequences.
• Skipping the spot-check. Sampling 5 to 10 percent of citations against source pages costs an hour and is the only way to know your error rate on this deal, with this room, on this day.

When AI review is the wrong choice

Be honest with yourself on these:

• The room is small. Under a few hundred documents, a competent associate reads it in days, and the setup overhead buys little.
• The decisive question is judgment, not extraction. Whether a founder will stay, whether a market is real, whether the price is right: no model reads that out of a data room.
• You cannot verify. If nobody on the team has time to spot-check citations, you are not running supervised AI, you are running unsupervised AI, and the benchmark numbers above say what that is worth.
• The NDA forbids it. Some sellers restrict disclosure to a narrow list of representatives. If counsel says the API does not qualify, the answer is no, or negotiate the carve-out first.
• The documents are not machine-readable. A room full of low-quality scans of wet-ink contracts needs an OCR and cleanup pass before any of this works, and sometimes that effort eats the savings.

Frequently asked questions

Can Claude read a whole data room at once? No. PDF support is capped at 600 pages and 32 MB per request, and context limits mean large rooms are always processed in structured passes: index everything, then deep-read by workstream. This is a feature, not just a constraint, because per-workstream passes produce better-cited output.

Is my deal data used to train the model? On Anthropic's commercial products (API, Claude for Work), inputs and outputs are not used for training by default. Consumer products and explicit feedback submissions follow different rules, so keep deal work on commercial tiers and check the current policy before each engagement.

How accurate is AI contract extraction in diligence? Good enough to triage and extract with citations, not good enough to skip verification. Benchmarks on financial statement tasks put the best models near 80 percent, with sharp drops on the hardest files. The correct design treats the model as a fast junior reader whose every claim carries a page number you can check.

What does the SEC say about using AI in diligence? The FY2026 exam priorities say examiners will check that firms have adequate policies and procedures to supervise AI use, and that representations about AI capabilities are accurate. The priorities are a staff statement, not a rule, but they are a reliable preview of exam questions.

Should we buy a PE-specific AI platform instead of building on the API? If you run a handful of deals a year, a platform with a prebuilt data room workflow is probably cheaper than building. If diligence is continuous or you need the system inside your own cloud perimeter with custom checklists, building on the API gives you control a platform will not. Nearly everything in this workflow is retrieval over your documents rather than model training; the trade-off is the same one we walk through in RAG or fine-tuning for banks and insurers. We implement the second kind.

Does this work for legal due diligence too? The same extraction-and-citation workflow applies to contract review for law firms and in-house teams, with its own checklists and privilege considerations. That is a separate implementation we run under Claude for Legal.

Where to go from here

AI data room review works when it is built as a supervised extraction pipeline: index first, extract with citations, tie out claims, compute in code, and keep a human checking the source pages. The API cost is a rounding error on a deal; the control framework is the real work. If you want this running inside your fund's environment, with the checklists, citation verification, and SEC-ready documentation built in, that is what our Claude for financial services team implements. Start with one deal and measure the error rate yourself.