What should PE operating partners check in AI due diligence?

Check eight things before close: business model exposure to AI, whether the target's AI claims are real, an inventory of AI in use, data rights, governance and incident response, the cost of manual work, the maturity of each AI workflow and who owns it. Score each yes/no check, then price the gaps into the deal.

An AI due diligence checklist covers

• Exposure: whether AI weakens the target's pricing, moat or demand.

• Claims: whether the "AI agents" in the deck pass a live run.

• Risk: shadow AI, data rights, regulation, governance and incident response.

• Upside: manual work AI could take over after close, in dollars, with a maturity level and owner per workflow.

Evidence: Bain & Company says "we have evaluated the impact of AI on more than 1,000 companies in diligence" in its AI diligence article, and that "Most acquirers tell us that an AI diligence has convinced them to walk away from a deal."

Key points

• AI diligence has two halves: the risk AI poses to the target and the upside AI could add after close. Score both.

• Ask for one live run of every "AI agent" the target sells or uses, then check it for planning, tool use, memory and adaptation.

• No AI inventory and no AI use policy is a finding in itself.

• Size post-close upside in dollars of manual work, with the same math the 100-day plan reports against.

What are operating partners checking in AI diligence in 2026?

They check two things: whether AI breaks the target's business and whether its AI story holds up. I opened each source below in October 2026.

1. Disruption first. Bain & Company is a management consultancy that publishes research on private equity for investors and operating teams. Its diligence article warns that "per-seat or cost-plus pricing models could become risky."

2. Governance gaps. Grant Thornton is an audit, tax and advisory firm that publishes annual surveys of business leaders, including private equity leaders. In its 2026 AI Impact Survey, "9% are very confident they could pass an AI governance audit in 90 days."

3. Legal exposure. Mayer Brown is an international law firm that advises private equity sponsors on deals and regulation. Its May 2026 note says "many companies have no centralized record of what AI systems are in use."

4. Inflated claims. Debevoise & Plimpton is a law firm that advises private equity firms and companies on litigation and transactions. Its May 2026 note on AI washing warns that buyers paying a premium for "a target's AI narrative" can find the AI "largely illusory" after closing.

5. Agent washing. Gartner is a research and advisory firm that publishes technology forecasts for business and IT leaders. As reported by W.Media, "Gartner estimates only about 130 of the thousands of agentic AI vendors are real."

6. A dedicated hire. LLR Partners is a lower middle market private equity firm that invests in growth companies. Its AI value creation job posting asks the hire to "Assess AI maturity, data assets, and competitive positioning during diligence."

7. Operators comparing notes. PEI Group is a media company that runs Operating Partners Events, a conference series for PE operating partners. Its June 2026 Napa session asked "Which AI risks tend to get missed in diligence and later show up as execution problems?"

What is the Four-Trait Agent Test?

It is the check to run in the management meeting when a target says it uses or sells AI agents. "A true agent has four traits: planning, tool use, memory and adaptation. If any one is missing, it is an assistant with better marketing." The full test is on the AI assistant vs AI agent page.

Ask for one live run of each claimed agent on real data and watch for each trait:

1. Planning: did it break the goal into steps, or follow a fixed script?

2. Tool use: did it act in another system, such as the CRM or ERP?

3. Memory: did it use context from earlier runs?

4. Adaptation: when a step failed, did it change course?

Then place each workflow on the AI Automation Maturity Ladder. "It is a four-level scale that describes how much of a workflow AI does and how much a person still does."

If the deck says "autonomous agents" and the workflows score L1, value them at L1. The maturity ladder page has the full level rules.

What should you check on business model risk and AI claims?

Start with whether AI breaks the business, then test management's AI claims. Score 1 point per yes.

1. Business model exposure

• [ ] Category set: the deal team has placed the target as revolution, transformation or augmentation.

• [ ] Pricing holds: per-seat or cost-plus revenue still makes sense if the customer needs fewer people.

• [ ] Moat holds: a competitor using AI could not rebuild the core product or service cheaply.

Bain puts fewer than 10% of the 300+ companies in its study in the revolution category and about half in augmentation, per the same article. For most targets, the upside work in section 6 matters more than the disruption memo.

2. AI claims

• [ ] Live run seen: every claimed AI agent or feature has been shown working on real data.

• [ ] Four traits pass: each claimed agent passed the Four-Trait Agent Test, or the deck was corrected.

• [ ] Claims match: the website, deck and CIM describe AI the same way the live run showed it.

3. AI inventory and shadow AI

• [ ] Inventory exists: the target has a written list of every AI tool and model in use, with owner and purpose.

• [ ] Policy exists: there is a written AI use policy that staff have seen.

• [ ] Shadow use checked: staff surveys or network logs were reviewed for AI tools used without approval.

Mayer Brown's note states that "The absence of such policies is itself a red flag that warrants further investigation." Angel Ramirez is a technical due diligence practitioner who writes on Substack about AI assessment for private equity buyers. His post asks: "What AI tools are employees using without IT approval?"

What should you check on data, governance and incident response?

Check who owns the data AI touches, what the law says about it and whether anyone can stop a bad run.

4. Data rights and access

• [ ] Terms reviewed: every AI vendor's terms were read for data rights and output ownership.

• [ ] Provenance known: for any model the target trained, the training data source and license status are documented.

• [ ] Data reachable: the systems holding core data have API or export access.

Mayer Brown warns that platform terms employees accept can "govern data rights, output ownership, and vendor liability in ways management has not evaluated."

5. Governance and incident response

• [ ] Regulation mapped: counsel has checked whether any system could be "classified as high-risk under AI laws (including the EU AI Act)," in Mayer Brown's words.

• [ ] Governance owner: one executive owns AI governance and could show controls to an auditor.

• [ ] Incident plan tested: someone can pause an AI system and roll back its output, and has practiced it.

The bar is low: in the Grant Thornton survey, "7% have a tested AI incident response plan."

How do you size the manual-work upside before close?

Estimate the hours of repetitive work in the largest teams and convert them into a yearly dollar figure. Bain names the signal to look for: "whether there are a lot of people in the target's business doing knowledge work."

6. Manual work cost

• [ ] Teams sized: headcount and loaded cost are known for the 3 largest knowledge-work teams.

• [ ] Workflows listed: management can name each team's repetitive workflows, with rough minutes and runs per week.

• [ ] Steps sorted: the top workflows have been sorted step by step. "Sort each step into delete, plain code, agent or human decision."

Staff access is rare before close, so the full Manual Work Leak Audit waits until after close. "A manual-work leak audit is a one-to-two-week review that lists every repetitive task a team does by hand, times each one, and converts the hours into a yearly dollar cost." Before close, ask management for a rough estimate using the same formula, and treat it as a range to verify.

Illustrative worked example. A 12-person team spends 8 hours each a week on manual work at a $60 loaded hourly rate. That is $299,520 a year. Getting half back is worth $149,760 a year. Put it in the value case as a range.

7. Workflow maturity

• [ ] Levels scored: each AI workflow in the inventory has a level from L1 to L4.

• [ ] Runs logged: workflows at L2 or above keep a log of every run and every human override.

• [ ] Evals exist: someone has graded real past outputs, so quality can be measured after close.

8. People and ownership

• [ ] Management ready: Bain says to "assess the AI readiness of the management team and the broader enterprise"; the CEO can name the top 3 AI priorities and their owners.

• [ ] Owners named: each AI workflow has one owner for its outcome and one approver for its output.

• [ ] Process written: the costliest workflows have SOPs a new hire or an agent could follow.

How do you score the target?

Add one point per yes, for a maximum of 24. Score sections 1 and 2 by the end of management meetings and the rest in confirmatory diligence.

Two rules override the total. Any "no" in section 2 sends the AI claims to legal review before signing. Three "no" answers in section 5 go to counsel for the purchase agreement, whatever the score.

Which red flags should change the deal?

Each of these five has a fix for the purchase agreement or the post-close plan.

1. AI that will not run live. Management shows slides and recorded demos only. Fix: make one live run per claimed agent a condition of the next stage.

2. Agents that are assistants. The "agent" is missing one of the four traits, the relabeling Gartner calls agent washing. Fix: model it at its real ladder level.

3. No AI inventory or policy. Nobody can list the tools staff use. Fix: run a staff survey and network log review before signing, and add a rep on disclosed AI use.

4. Data the target does not own. Vendor terms grant rights over customer data or outputs. Fix: list the contracts to renegotiate and price the cost.

5. AI value with no baseline. The CIM claims AI savings but nobody timed the work before AI. Fix: value the claim at zero until a baseline exists, then measure it in the first 30 days.

What happens to the checklist after close?

It becomes the starting point of the post-close plan. The scores show which gaps to close first, and section 6 lists the workflows days 1 to 30 will audit in full. The 100-day AI value creation plan covers that work: audit and score, pilot one workflow at L2, then promote what runs clean.

• Claims fixed first. Any agent that failed the four traits gets rebuilt or relabeled.

• Contract next. Each workflow that goes live gets an Agent Handoff Contract setting what the agent may decide alone and how to reverse bad output.

• Revenue workflows. Lead routing, CRM hygiene and pipeline reporting fit GTM Engineering: an audit with a costed workflow map, then a 30-day Sprint with up to 3 workflows live in the CRM.

Most portfolio companies stall between the checklist and the first live agent, because nobody inside has time to write the SOPs. That gap is what Forward Deployed Engineering covers: Phase 1 is an audit and AI roadmap, Phase 2 documents SOPs as the knowledge layer, and Phase 3 builds agents one at a time, each with a handoff pack (SOP, run log, rollback path and owner). Book a free coffee chat to get yours scoped.

FAQ

What is AI due diligence in private equity? It is the part of diligence that tests how AI affects a target: whether it threatens the business model, whether the AI claims are true, what legal and data risks AI use creates and how much manual work AI could take over after close.

When should AI due diligence start? At screening, with the business model check, because a revolution-category target may not be worth the rest of the work. Claim tests belong in management meetings; inventory, data and governance checks belong in confirmatory diligence.

How do you check whether a target's AI agents are real? Ask for one live run on real data and apply the Four-Trait Agent Test: planning, tool use, memory and adaptation. If any trait is missing, the product is an assistant, and the valuation should treat it as one.

What is shadow AI and why does it matter in diligence? Shadow AI is any AI tool staff use without approval. Confidential or customer data may already sit in public AI tools under terms nobody reviewed. Check it with staff surveys and a review of network logs for AI services.

How do you value AI upside before close? Ask management for a rough estimate with the Manual Work Leak Audit formula: runs per week times minutes, divided by 60, times the loaded hourly rate, times 52. Put the result in the value case as a range, then replace it after close with the full audit, where each task is timed on live runs with its owner.

Who should run AI due diligence, the deal team or the operating partner? Both. The deal team owns the business model and legal checks with counsel. The operating partner owns the claim tests, manual-work sizing and maturity scores, because those feed the first 30 days after close.

Use the checklist on your next deal

Score your next target on the eight sections in its management meetings and data room, then carry the sheet into the 100-day plan after close. Kevin Payne is an AI engineer and GTM operator who installs AI workflows for B2B SaaS revenue teams. His three services (GTM Engineering, Forward Deployed Engineering and Content Engineering) are on the services page.

Want to talk through where AI upside is still unproven across your portfolio? Book a free coffee chat.

About the author: Kevin Payne is an AI engineer and GTM operator who installs AI workflows for B2B teams. He has 10+ years in B2B SaaS demand generation and RevOps across Techstars, a16z and YC portfolio companies, and his writing has appeared in HubSpot, GetResponse, Outbrain and more.