AI in private equity reads the data room, builds the quality of earnings, model and LBO, and drafts the IC memo and deck. Most firms are still piloting it. On 46 held-out deal tasks, the best agent passes about 80% of the checks a VP runs; the misses are judgement calls that move price.
~20%
of PE portfolio companies have generative AI in production with concrete results (Bain, investors with $3.2T AUM)
7%
of PE leaders say AI is fully integrated across their portfolio (FTI Consulting, 200 respondents)
80.5%
top score on PE Bench, graded on the Excel, Word and PowerPoint files AI agents deliver
Firms measure adoption, not the work
Every survey of AI in private equity asks the same question: how many firms use it. Bain’s 2025 survey of investors with $3.2 trillion under management found nearly 20% of portfolio companies running generative AI in production. FTI Consulting’s 2026 AI Radar found 36% of PE leaders say their portfolio companies use AI day to day, and 7% call it fully integrated. When McKinsey asked a room of about 65 operating partners who had AI in production at scale, “only three hands, or about 5 percent, stayed up.”
None of these surveys asks whether the work is right. A quality of earnings that accepts the seller’s add-backs, or a model with typed-in numbers where formulas belong, costs more than the hours it saves. So we measured it.
How we measured AI on deal work
PE Bench is our benchmark of a private equity analyst’s work. Former private and growth equity investors built five deals: three private companies with full data rooms of up to 158 documents, and two take-privates of public companies built from their filings. Each of the 58 tasks gives the agent the short brief an analyst would get, such as “Bridge Project Voltaic’s FY25 reported result to adjusted EBITDA for pricing from a buyer’s perspective, using the data room.”
Each task is graded on the files the agent hands back, with dozens of checks per task: do the numbers tie, are formulas live, does each figure trace to its source, does the memo match the model. The checks that would move the price carry the most weight. Scores below are the overall PE Bench score: the four deals held out from publication (two private companies and two take-privates), each weighted equally. The fifth deal, Voltaic, is published and scored separately.
| Agent | Score | Cost per task | Minutes per task |
|---|---|---|---|
| Compound | 80.5% | $4.46 | 15.6 |
| Claude Code with Opus 5.5 | 78.1% | $6.79 | 22.0 |
| Claude Code with Fable 5.1 | 73.5% | $13.25 | 23.2 |
| Codex with GPT-6 Astra | 70.1% | $3.35 | 11.3 |
An overall score hides where agents differ. The sections below show where.
AI use cases in private equity, stage by stage
AI now drafts each stage of a deal: it reads the data room, builds the quality of earnings, historicals, comparables, model and LBO, and writes the IC memo and deck. Each stage has its own way to fail.
| Deal stage | Where AI fails on PE Bench | What a person must check |
|---|---|---|
| Data room review | Picks one of two conflicting figures without saying the documents disagree | That each cited page says what the answer claims |
| Quality of earnings | Accepts the seller’s add-backs as presented | Every add-back that moves the price, against the evidence |
| Historicals | Stops at unadjusted EBITDA, or drops the company’s own adjusted figure | Ties to the audited accounts and the company’s EBITDA definition |
| Comparables | Types in market values from a website instead of building them from share counts and debt | Enterprise value built from authoritative sources |
| Operating model and LBO | Keeps the cases from changing the numbers they should drive | Flex one assumption and watch returns move |
| IC memo and deck | Breaks the template’s layouts, or lets the memo drift from the model | Every number in the memo against the model |
The add-back AI agents missed: about $60M of price
In the Voltaic data room, the seller adds back $4.8M of sale costs to FY25 EBITDA. $2.9M of it was spent before the board authorized a strategic review, so it is not a cost of the sale. Claude Code with Opus 5.5 and Codex with GPT-6 Astra accepted the full $4.8M. Compound kept $1.9M and rejected $2.9M.
Voltaic FY25: sale costs added back to EBITDA
US$ millions, by agent, on PE Bench
The $2.9M difference was spent before the board authorized a strategic review. At the deal's 21.0x entry multiple it is about $60M of purchase price. From each agent's delivered workbook on PE Bench.
A survey that counts how many firms use AI for diligence would score all three agents the same. The adjusted EBITDA example walks through every add-back in this data room.
Where AI struggles most: private data rooms
Public companies are the easier case: their filings are audited and consistent. A seller’s data room is different: the CIM, the management accounts and the audited financials disagree, and the agent has to decide which number to believe and say why. On Voltaic, the data room PE Bench publishes, Compound scores 79.1% and Claude Code with Opus 5.5 scores 53.3%.
When you test AI for private equity, test it on a private deal’s data room, not on a public company.
Private equity AI tools: what each kind delivers
| Kind of tool | What you get back | Fits |
|---|---|---|
| General AI assistants and chatbots | Answers in chat and short drafts | Questions on a few documents, first drafts |
| Coding agents | Files built by scripts, often with typed-in numbers | Teams with engineers to review the files |
| Relationship and sourcing platforms | Contact and deal records, target lists | Sourcing and pipeline tracking |
| Document search and research tools | Cited answers across many documents | Research across filings and data rooms |
| Finance agents built for deal work | Excel models with live formulas, Word memos and PowerPoint decks, each number traced to its source | Diligence through to the IC deck |
Whichever you use, test it on a deal you have already closed, where your team knows the answers: give each tool the same brief, trace ten numbers back to their source, and change one assumption to see whether the model recalculates.
What AI changes about the companies you buy
AI also changes what a deal team diligences. Arthur D. Little reports that enterprise IT services companies “have lost more than a third of their value since 2021.” Add three questions to diligence of any business that sells hours or software seats:
- How much revenue is really from AI? Sellers relabel existing projects as AI work. Ask for AI revenue by contract.
- What happens to pricing if work takes a tenth of the time? Billing by the hour or by the seat is exposed. Model the price change, not only the cost saving.
- What does the target’s own AI cost? Model and compute costs scale with usage. Check that gross margin holds as usage grows.
What to require before AI work reaches the IC
- A source for every number. Each figure in the memo traces to a model cell, and each hard-coded input to a page in the data room.
- Judgement calls flagged, not buried. Add-backs, revenue recognition and normalizations listed with the evidence for each.
- A live model. Formulas, not typed-in values, and cases that change the outputs when switched.
- The memo matches the model. Every number in the memo and deck ties to the current version of the model.
- A named reviewer. One person on the deal team signs off on each deliverable, as they would for an associate’s work.
FAQ
How is AI used in private equity?
Deal teams use AI to read data rooms, build the quality of earnings and historicals, build operating models and LBOs, and draft IC memos and decks. Firms also use it for deal sourcing, portfolio monitoring and LP reporting. Most firms are still piloting: FTI Consulting found 7% of PE leaders call AI fully integrated across their portfolio.
Is AI accurate enough for due diligence?
On PE Bench, the best agent passes about 80% of a task’s checks, weighted by how much each would move the price. The misses are mostly judgement calls, such as which add-backs to accept. AI can do the first pass of diligence, but a person should review every adjustment that moves the price.
What are the best AI tools for private equity?
For diligence through to the IC deck, choose a tool that returns Excel models with live formulas, Word memos and PowerPoint decks, with every number traced to its authoritative source. Test it on a closed deal your team already knows. PE Bench, our own benchmark, scores agents on this work: the top score is 80.5%.
Will AI replace private equity analysts?
AI does more of the first draft: the spread, the model build and the memo outline. The analyst’s work moves toward checking sources, making judgement calls and defending them to the deal team. On PE Bench, the costliest misses were judgement calls an experienced analyst would catch.
Can a firm test AI on its own deals?
Yes. Run each tool on a closed deal, with the same brief, and compare the files against what your team delivered. A private benchmark built from the firm’s own deals and standards repeats that test every time a new model ships.