Investor note

Five AI questions to ask in software diligence

May 2026.7 min read

Audience: PE and VC partners and operating partners running underwriting on B2B software, plus corp-dev teams running adjacent diligence.

The market claim

The market is now repricing B2B software on the assumption that agentic AI substitutes for the seat. Through Q1 2026, the iShares Expanded Tech-Software ETF was down more than 21 percent year to date, and the public software basket erased roughly $2 trillion in market capitalisation (FinancialContent, The SaaSpocalypse: AI Agents Trigger a Massive Repricing). The implication for diligence is structural, not cyclical. Multiples have compressed against earnings that have not yet adjusted, and a SaaS-era diligence framework reads the gap as a buying opportunity. Sometimes it is. More often the framework is mispricing the underlying risk.

What the market is mispricing

Two underwriting assumptions need updating before any 2026 software diligence is signed off. The first is that growth and revenue retention are leading signals of business quality. They are now lagging signals. Atlassian’s 2026 disclosure of its first ever decline in enterprise seat counts came after net revenue retention had already softened on the trailing four quarters (The SaaS CFO, Death of Per-Seat Pricing). The second is that the SaaS gross margin floor at 75 to 80 percent is durable. ICONIQ’s January 2026 snapshot put average AI product gross margin at 52 percent, up from 41 in 2024, and public software now discloses 60 to 70 percent as the realistic AI-adjusted range (Monetizely, The Economics of AI-First B2B SaaS; SaaSMag, The AI COGS Problem). The right diligence pack puts five sharper questions ahead of the LBO model.

The five questions

1. What part of the monetised workflow could an AI agent remove entirely?

The relevant test is not whether the product has AI features. It is whether the human workflow on which seat counts depend can be removed by a credible agent build using off-the-shelf models within twenty-four months. Workflow automation is the leading category under replacement pressure, with 35 percent of surveyed enterprise software builders reporting they have already swapped out at least one SaaS tool for an internal build (The SaaS Moat Crisis, Big Ideas DB). If the answer is yes, the recurring revenue base is contracting underneath the growth plan, and the diligence model needs to run on declining seats rather than stable seats.

2. Does AI raise or lower switching costs in this specific business?

Three categories of switching cost survive the AI repricing: proprietary datasets that competitors cannot replicate, regulatory and compliance infrastructure with multi-year switching cycles, and integration depth that makes a system of record expensive to remove (HarbourVest, The Software Industry’s Great Reset). Everything else is eroding. Diligence question: which category does this business sit in, with evidence, and how much of the renewal book is in customers where switching cost is one of the survivors versus one of the casualties? A copilot isn’t a moat. A regulated system of record with proprietary outcome data is.

3. Does the current pricing model survive usage-based AI economics?

Per-seat pricing assumed a stable or growing seat count and near-zero marginal serving cost. Both have moved. ServiceNow Now Assist is tracking to $1.5 billion ACV in 2026 with 50 percent of net new business already on non-seat pricing tied to tokens and consumption, and GitHub Copilot is migrating to pure token billing on 1 June 2026 (Bessemer, AI Pricing and Monetization Playbook). Bessemer now treats hybrid base-plus-consumption as the dominant transition state for enterprise renewals. The diligence question is whether management has modelled the gross margin and net revenue retention profile of its top three products under a forced shift to hybrid pricing, with realistic inference cost curves. If not, the LBO model is anchored on the wrong revenue shape.

4. What happens to gross margin if AI adoption inside the customer succeeds?

This is the AI margin illusion in underwriting form. Revenue rises with AI feature adoption, but inference cost rises faster if the architecture is wrong. The 80-point SaaS gross margin standard is structurally weaker in AI-native delivery because compute, model routing, and vector storage make COGS variable in a way classic SaaS never had to model (SoftwareSeni, Why AI Gross Margins Are So Much Lower Than SaaS). The diligence test is to take the management gross margin guide for year three and subtract the contribution from features that depend on customer-borne inference. The residual is the durable margin. If the residual is below the LBO model assumption, the deal is mispriced.

5. Are the AI features defensive, expansionary, or cosmetic?

The honest answer affects how to weight the AI roadmap in valuation. Defensive features (parity with peers, prevent churn) earn the company a flat curve, not a re-rating. Expansionary features (new pricing tier, new buyer, new workflow ownership) earn a step. Cosmetic features (AI as a label, no economic mechanism behind it) are AI roadmap theatre and earn nothing, even if they ship on time. AI summarisation, smart automation, and generic copilots are now table stakes by 2026 in most categories and serve as margin-improving features rather than competitive advantages (Designli, How to Build a SaaS Moat in 2026: Beyond Feature Parity).

Upside and risk

Upside. A software business that answers questions one and two in the structurally defensible quadrant, has begun the pricing migration before the renewal cycle forces it, and is shipping expansionary AI features rather than cosmetic ones, is mispriced cheap on current multiples. Thoma Bravo and Vista are openly stating that selective software buyouts at compressed multiples are the highest-conviction opportunity of the cycle (Bloomberg, Thoma Bravo and Vista on AI). The diligence has to be sharper than the multiple, not just opportunistic on it.

Downside. The base case for the misdiagnosed deal is not a write-down on growth assumptions, it is a write-down on gross margin and net revenue retention together. That is the worst possible LBO outcome because both inputs to leverage capacity move the wrong way at the same time. A diligence framework that prices only growth will overvalue every transitional software business by the area under both curves.

Watch this, not that

Watch gross margin disclosed in earnings on a constant pricing-mix basis and net revenue retention on a constant seat-count basis. Stop watching headline revenue growth and total AI feature count on the roadmap. The first two move when the underwriting risk is operating. The second two move whether or not the underwriting risk is operating, and have already mispriced the cycle.

Three diligence checks to put on the partner pack:

Customer reference calls that ask the customer’s own engineering team, not its IT buyer, what they would build if they had to replace the product.
A pricing-model sensitivity table that forces management to publish a hybrid-pricing P&L alongside the seat-pricing P&L for the top three products at year three.
A switching-cost classification by ARR cohort, separating regulated, data-graphed, and integrated systems of record from the rest.

If the deal does not survive these three, the LBO model is not the issue. The thesis is.

Sources

diligence, investors, moat, pricing, margin

Back to writing