AI due diligence on a software company: questions to ask

01

Ask for the AI revenue and cost lines

Check the cost of serving the AI usage behind the reported revenue. Bain notes that AI products have significant variable costs per use that traditional SaaS does not have. To see the spend, it recommends tracking hosting and infrastructure, third-party model costs and the fully loaded cost of the people working on AI research and development.

Docusign and Microsoft describe the same pressure in their annual reports. Docusign warns that the costs of developing and operating AI solutions, "including initial training costs and ongoing processing and inference costs", may reduce profitability. In its fiscal year to June 2026, Microsoft's Cloud gross margin percentage decreased to 66%. Investments in AI infrastructure and growing AI product usage drove the fall, partly offset by efficiency gains.

Ask for AI revenue in the three buckets Bain proposes: traditional AI and machine learning, AI add-ons such as copilots, and agentic products. Ask for the costs against each, separated from the core business. Have management distinguish the costs of serving current usage from spending on development, and explain how those figures relate to reported gross margin.

02

Ask why renewals shrank and deals were lost

Renewal and lost-deal records show whether customers already pay for fewer seats or have chosen an alternative, such as building their own. Dropbox's annual report names the risk of customers reducing their seat licenses over time, "including as a result of increased adoption of AI tools". Guidewire's platform for insurers combines core systems of record with AI capabilities. The company warns that customers "may leverage these capabilities to develop their own solutions that could reduce or eliminate the need for our solutions."

Ask for the renewal records with the change in users or units and the reason recorded when each renewal closed. Ask for the lost and churned deals with the alternative each customer chose, including building in house. PitchBook adds a series to watch: monthly logo churn among small and medium-size customers. It treats that churn as a leading indicator because those customers are the early adopters of good-enough AI replacements.

03

Questions to ask the target's customers

Customer calls let you test management's explanation against the buyer's account of what changed. Bain lists research with customers among the ways to uncover AI's impact on a target. PitchBook writes that enterprise customers expect AI to reduce seat counts and services spending. Choose some customers yourself, as well as the references management offers.

Start with what has already changed. What have your people started doing with AI that they used this product, or the supplier's people, for? Did the number of users you pay for change at the last renewal, and why?

Then ask what the customer will still pay for. Which of the supplier's AI features would you pay more for, and which do you expect to be included in what you already pay? As you do more yourselves, what will you still need this supplier for?

Last, test whether an alternative is real. Have you tried to build or buy something that does part of this work? What would it have to connect to, and what happens to you if it fails?

Follow up by asking what the customer has tested beyond the demonstration. Which required connections work, how would the replacement be secured, and who would run it and deal with failures? Those answers help distinguish an alternative the customer is exploring from one it is ready to adopt.

A prototype can help test how easily functionality can be reproduced. In one diligence, Bain's team built and tested a prototype of a target's tool "outside in". That convinced the sponsor that the target's technology might more easily be challenged by incumbents or new entrants. The customer call still needs to establish whether the customer intends to replace the product and what remains unresolved before it can do so.

04

How to sort a finding, with a made-up example

Sort each answer into evidence, a grounded hypothesis or an unknown, and ask four questions of every claim: What do we know? How do we know it? Who confirmed it? What remains an assumption? My commercial due diligence makes the same split and asks the same questions of every qualification element in a pipeline.

Say management tells you that smaller renewals were entirely due to layoffs at those customers and that AI played no part. The renewal records establish that paid seat counts fell. Management's explanation for the fall is a hypothesis to check.

Ask who supplied the reason, when it was recorded and whether it came from the customer or the account team's interpretation. A note made when the renewal closed preserves that account; its timing alone does not verify the cause. If a customer confirms the staffing reduction and connects it to the seats removed, that supports a grounded hypothesis that layoffs explain the smaller renewal. It does not establish that AI played no part.

Whether AI will reduce that customer's future need for seats remains unknown. Ask what work the customer expects to automate before the next renewal and what that would mean for the seats it pays for. Record that answer as the customer's current expectation. If the investment case assumes stable seat demand, show where it depends on this unresolved question.