What should a PE sponsor ask in AI due diligence?

An acquisition target may have an impressive AI demonstration and very little evidence of financial impact. Another may have modest technology, well-organized data, and a management team with a credible opportunity to improve margins.

Distinguishing between those businesses requires questions that connect technology to how the company earns money, serves customers, and manages risk.

For a sponsor, AI diligence should support three decisions: what to underwrite, what to address before or immediately after closing, and what to leave outside the base case until the evidence improves.

I spent 2019 to 2023 on the sponsor side at NHK Capital Partners, raising from accredited investors and institutional capital. We sourced the capital stack ourselves, from developer capital commitments to senior and mezzanine debt, and held a sizable stake in sponsor equity. Across 8 projects, investor commitments totaled $116.5 million. We owned what we brought to market, so every claim in the investment case had to hold up. The questions below come from that discipline.

Start with the workflows.

Ask management to identify where AI is in production, who uses it, and what happens differently because it is there. Separate tools employees use occasionally from systems embedded in customer delivery, pricing, sales, or operations.

A useful opening request is: "Walk us through a recent transaction from beginning to end, including every point where AI contributes."

That walkthrough should reveal the inputs, the output, the human review required, and the exceptions that send work back to an employee. It should also establish whether the capability is available across the business or operating within a narrow pilot.

A demonstration shows what a system can do under selected conditions. Operating records help establish what it does consistently.

Every material AI claim should connect to a baseline and a financial outcome.

If management reports faster customer service, ask for handling times, resolution rates, repeat contacts, staffing levels, and service quality before and after implementation. If the claim is higher sales productivity, examine conversion, contribution margin, and the cost of supporting the additional business.

The central question is how the operational improvement reaches earnings.

Time saved may support growth, reduce overtime, or allow an open position to remain unfilled. Those are different outcomes, with different evidence requirements. Multiplying estimated hours saved by an employee's hourly cost does not establish an expense reduction.

Ask finance to reconcile claimed benefits to actual results and identify other changes that may explain the improvement, including pricing, staffing, seasonality, and customer mix.

Data availability and permission to use that data deserve separate attention.

The target may hold years of customer records without having consistent identifiers, reliable fields, or clear rights to use those records for a proposed AI application. Information may also sit across systems that are expensive to connect.

Request a practical review of:

• The data required for each material use case.
• Its quality, completeness, and accessibility.
• Contractual restrictions on processing and reuse.
• Access controls, retention practices, and vendor handling.
• The work and expense required to resolve gaps.

Legal and technical reviewers should assess the permissions and protections relevant to the business. The investment team should understand how unresolved issues affect deployment timing and cost.

A company can create substantial value using third-party AI. The diligence question is how much control it retains over the capability and its economics.

Clarify what the target owns, licenses, configures, and maintains. Identify dependencies on individual employees, implementation partners, model providers, and core software vendors.

Then test a few practical scenarios. What happens if a provider raises prices, changes its service, or becomes unavailable? Can the company move its data and operating processes elsewhere? Who can maintain the system if its original developer leaves?

When AI contributes directly to the product being acquired, these questions also inform the assessment of differentiation. A widely available feature may be useful, while offering limited protection from competitors.

The right controls depend on the consequences of a mistake.

An internal drafting assistant and a system that communicates prices to customers require different review standards. Ask management to show how it tests accuracy, restricts access, records failures, and escalates decisions requiring judgment.

The NIST AI Risk Management Framework provides a voluntary reference for organizing this work around governance, context, measurement, and risk management.

For diligence purposes, look for evidence that controls operate in practice: documented tests, assigned owners, incident records, and a workable fallback process. A policy alone cannot establish that employees follow it.

A target's own AI program is only part of the investment question.

Consider whether customers could perform more of the work themselves, whether competitors could deliver comparable services at lower cost, and whether faster delivery will become a basic expectation.

For businesses that bill for time, greater efficiency may also challenge the pricing model. Management needs a view on how it will retain value as the effort required to deliver a service changes.

Test both the opportunity to improve margins and the possibility that customers will expect some of that benefit through lower prices.

Ask for a named business owner, an implementation budget, a realistic sequence, and measures that finance can validate.

Management capacity matters. A company integrating an acquisition, replacing its ERP, and rebuilding its commercial team may have limited room for another substantial operating change.

Before the investment committee discussion, summarize each material AI initiative under four headings: demonstrated results, remaining assumptions, required investment, and accountable owner.

That gives the sponsor a defensible basis for deciding which benefits belong in the underwriting and which should remain contingent upside.