Due diligence in the AI era: what buyers rebuild in a software company
AI has not rewritten the software scorecard. Growth, retention, gross margin and efficiency still frame how a buyer reads a software company. What AI changes is how each of those numbers reads, and that is exactly what due diligence will test, on two fronts: the financial review rebuilds your numbers, the technical review reads your code.
- Usage revenue makes ARR harder to define.
- Net retention proves less about customer health.
- AI margins land between software and services, and AI budgets keep overrunning.
- Productivity gains turn into capacity, rarely into savings.
- In technical due diligence, AI written code adds a provenance question.
In financial due diligence, a buyer rebuilds your numbers from invoices, the CRM and the P&L. A strategic acquirer and a private equity fund will not weigh those numbers the same way: the first looks at how your product fits its own, the second at the cash flow it can finance and grow. Both rebuild them the same way. In our deals at dups, AI is now where the rebuilt number comes back lower than the one in the deck. So rebuild them yourself, before the buyer does.
Free scorecard
Is your SaaS ready to sell? 14 questions, about five minutes. Get your score out of 28, your deal killers and the three gaps to fix first.
Take the scorecardYour ARR now needs a definition
AI already sits inside recurring revenue. In Bain's 2026 survey of software companies, about a quarter of respondents classify more than 40% of their ARR as AI ARR, meaning revenue from products or modules where AI is the main value driver.
The harder question is how that revenue is billed. More than half of respondents use usage based pricing for AI, and about a quarter outcome based pricing. Bain treats both as non recurring. So most software companies now carry a mixed book: subscription ARR on one side, usage revenue on the other. The usage part has to be translated into something comparable, and Bain describes three ways to do it:
- Current run rate. Annualise the latest month. Simple, but one spike or one seasonal month distorts it badly.
- Trailing average. Annualise an average, such as the last quarter. Smoother, but slow to show a real trend.
- Recurring usage base. Build a baseline per customer from historical consumption. The most accurate, and the most demanding in data.
In our deals, this is where the first argument starts. If you annualised a peak month, the buyer's number comes back lower. And it comes back after the LOI, when your alternatives are gone.
Before a sale:
- Pick one method, write it down and apply it the same way every month.
- Reconcile it in your ARR bridge (new, expansion, contraction, churn) against the statutory accounts, and put it in the data room.
- Split AI ARR from core ARR. The buyer will ask whether AI revenue is new money or a repackaging of what customers were already paying.
NRR still matters. It just proves less
Net revenue retention held up in the OPEXEngine database: in FY2025 it sat between 103% and 107% in every revenue band. Bain still flags that it is getting harder to hold, because AI pulls retention in both directions and adds noise:
- More expansion. Modules, features and usage growth give more room to grow an account.
- More churn risk. Trials are easier, competitors are closer and switching costs are lower.
- More noise. Usage can drop during a holiday week with nothing wrong at the customer.
Renewal used to be the proof that a customer was healthy. Usage now gives a signal in real time, but hides the reason behind it. In our deals, NRR on its own convinces a buyer less than it did two years ago.
Before a sale: bring the evidence behind the number: usage by customer cohort, active accounts and depth of use, and an explanation for every dip a buyer will find in the monthly data. Explain the volatility before diligence, or the buyer will use it to cut the price.
AI margins sit between software and services
In the OPEXEngine database, product gross margin was 83% in FY2025 and services gross margin 33%. AI lands in between: more than a third of survey respondents run their AI products and features below 50% gross margin. The reason is cost. Inference and compute sit in cost of revenue, and on third party models you do not control that cost. Bain is not convinced AI margins will reach traditional software levels as they scale.
Spending is running ahead of plan as well. Nearly three quarters of respondents spent more on AI tooling and infrastructure this year than planned, and almost all expect to spend more next year.
The bigger AI gets in your revenue mix, the more it pulls down the blended margin a buyer underwrites. Leave it unexplained and the buyer models next year's AI cost increase into EBITDA with their own assumptions. Those will not be generous.
Before a sale: show gross margin by product line with AI separate, and the AI cost line, how it has moved and how it moves in your plan.
Productivity is not savings
On paper, the gains are large: nearly 60% of respondents report an R&D productivity uplift of at least 20% since adopting generative AI tools. Very little of it reaches the P&L as lower cost. 77% bank the gains mainly as capacity, only 14% as cost reduction.
Revenue per employee tells the same story. In the OPEXEngine database it rose in FY2024 in three of the four revenue bands, mostly through headcount cuts, then flattened in FY2025 in three of the four revenue bands. Bain says it is too early to see whether the extra capacity turns into revenue.
"We became more productive" is not an answer that survives diligence. Productivity without outcomes gets discounted.
Before a sale: if you banked AI gains as capacity, show where the capacity went: roadmap shipped, technical debt cleared, release frequency, revenue per employee.
What technical due diligence checks
Everything above belongs to financial due diligence. Technical due diligence is a separate workstream, run by engineers who read the code itself. AI adds a question to each of their usual checks:
- Who owns the code. Code written by freelancers, agencies or before the company existed needs a written assignment. Code written with AI tools adds a provenance question: where did it come from, and under which terms.
- Open source licences. More than a third of respondents to Bain's survey say half or more of their new code is substantially AI assisted. Black Duck found licence conflicts in 94% of the codebases it audited in M&A transactions.
- AI and data. Which third party models you use, which customer data goes into them, and under which contract terms. Expect the buyer to ask for it in writing.
Before a sale: run a code scan, build a licence inventory and write down your AI code and data policy.
Your due diligence checklist before you talk to a buyer
- An ARR definition in writing, one annualisation method for usage revenue, and an ARR bridge reconciled to the statutory accounts.
- AI ARR split from core ARR, with a view on cannibalisation.
- Gross margin by product line with AI separate, plus the AI cost line and its trajectory.
- Customer health signals beyond renewal: usage cohorts, active accounts, depth of use.
- A capacity story backed by numbers.
- A code scan covering open source and AI generated code, signed IP assignments, and a written AI policy.
None of this changes your business. It changes how much of its value survives diligence.
Our job at dups is to sit next to the entrepreneur and make sure the numbers a buyer rebuilds match the ones you put on the table. If a sale or a round is on your horizon, talk to us before the buyer starts rebuilding, or see how we run due diligence and closing for sellers.
Sources
- Katherine Zhang (head of OPEXEngine, Bain & Company) and Adrian Kochan (VP strategy operations, OPEXEngine), "Evaluating software company performance in 2026, in the era of AI", Bain & Company webinar, 6 October 2026.
- Bain & Company and OPEXEngine, AI Benchmark Survey 2026: public and private software companies, mid 2026, figures are shares of respondents, sample size not disclosed.
- OPEXEngine database, FY2023 to FY2025: mostly private software companies headquartered in North America, four revenue bands from 10 to 50 million dollars up to 500 million to 1 billion, AI native companies excluded.
- Black Duck, 2026 Open Source Risk in M&A by the Numbers.
This article gives general information. It is not financial, legal or tax advice.
The same scorecard as before: growth, retention, gross margin and efficiency, with the Rule of 40 as a common guide rather than a hard rule. What AI changes, according to Bain & Company and OPEXEngine (October 2026), is how each metric reads. Expect the buyer's financial due diligence to rebuild your ARR definition, your retention, your gross margin by product line and your AI costs, and its technical due diligence to check who owns your code and which licences it carries, including AI assisted code.
Bain describes three methods: the current run rate, a trailing average, or a recurring usage base built per customer. Most companies use one of the first two. In a sale, what matters is to pick one, apply it the same way every month and reconcile it in the ARR bridge against the statutory accounts before signing an LOI.
Mostly not yet. In the AI Benchmark Survey 2026 by Bain & Company and OPEXEngine, 77% of respondents bank measurable AI productivity gains mainly as capacity and 14% as cost reduction. A buyer will ask where that capacity went: roadmap shipped, technical debt cleared, revenue per FTE.
Louis Vanheurck de Tornaco is COO at dups and lecturer in advanced entrepreneurship at Solvay Brussels School. His 2019 master thesis on transitioning from professional services to SaaS, built around the Real Impact Analytics case, won the de Barcy prize.
Let's build your next deal together
Your sparring partner for fundraising, acquisitions, and exits. We bring legal and financial firepower, entrepreneur's speed, and direct access to the right capital.
