For a stretch of this cycle, a credible AI story added momentum to almost any technology process. That period is over. Bain's April 2026 read on software investing puts the backdrop plainly: revenue growth that ran near 20 percent a year is trending at half that, net revenue retention is down about eight points since 2021, and dealmaking is at a crawl, with AI hanging over every valuation. Buyers have absorbed enough disappointments to treat the claim itself as a diligence item. The question in committee is no longer whether a company says AI. It is what would remain of the product, the margin and the moat if the claim were tested line by line.
The tests have become specific. Where the capability actually sits in the product, and how much of it is the company's own work rather than a thin layer over a vendor's model. What happens to gross margin as usage scales, once inference costs are carried at full weight. What rights the company holds over the data it trains and retrieves against; Skadden's 2026 M&A outlook lists data rights and provenance among the questions buyers now put first. What breaks, commercially and technically, if a model provider changes its price or its terms.
Exposure is read in both directions. Bain frames the diligence around two questions: how much AI could change the workflow the software supports, and whether AI could displace the product inside that workflow altogether. A buyer wants to know whether AI strengthens the company's position or quietly undermines it. Companies with real usage and retained customers on AI-heavy features hold up well under that reading. Companies whose story runs ahead of their telemetry do not.
The diligence itself now looks like revenue-quality work. Cohorts are examined for whether AI features change retention or expansion. Customer references are asked what they would pay without the feature. Technical sessions walk the architecture and the dependency map, and they run on compressed timelines: AKF Partners describes AI technical diligence that is now done in weeks rather than months. None of this is hostile. It is what underwriting looks like once a theme has matured.
The money is still there for the substance. Kroll counts a record 2,897 software deals announced in 2025, up 35 percent, with $291 billion of value. Strategic acquirers took 71 percent of them, their highest share in nine years, largely to buy AI-critical capability, and they paid 5.6 times trailing revenue against 4.4 times for sponsors, the widest strategic premium in a decade. That premium went to companies whose claims survived the tests above.
For sellers the implication is sequencing. Write the AI narrative the way it will be tested, before launch. Separate what is live from what is roadmap. Reconcile the margin story with the compute bill. Put data rights and model dependencies in order while there is still time to fix them quietly. A claim that survives scrutiny becomes a source of conviction and competitive tension. A claim that needs defending becomes a discount.
The market has not stopped paying for AI. It has stopped paying for the word. The substance, prepared properly, still moves outcomes.
