Key takeaways
Late-stage breaks are rising. One in five strategic buyers walked away in 2025.
The diligence question shifted. Buyers now test forward earning power, not just history.
Three workstreams decide outcomes. AI substitution, EU AI Act exposure, training-data provenance.
The forward case beats a clean data room. Documentation must be built during hold, not at launch.
The deals reaching signing are not necessarily the deals making it to close
In November 2025, Bain & Company surveyed 303 M&A practitioners across twelve countries. The headline finding made every dealmaking summary of the year: 2025 closed as the second-highest year on record at $4.9 trillion in deal value.
The finding that received far less attention from the same survey: one in five of those executives had walked away from a deal in the past year because of what they had concluded about AI's anticipated impact on the business. Not a price disagreement. Not a financing problem. A buyer-side conviction, formed during diligence, that the asset they were negotiating for was no longer the asset they thought they were buying.
Two true things about 2025 that nobody is reconciling
The consensus story about last year is straightforward. Deal value is up 40%. Two-thirds of GPs surveyed report that buyer-seller pricing is no longer the dominant obstacle to dealmaking. Volumes are recovering across most sectors.
The second story rarely makes the headlines. Herbert Smith Freehills Kramer's 2026 Global M&A Outlook reports that "2025 seems to have been a year where more deals than usual were abandoned, sometimes at a very late stage," and that more planning is now going into how to derisk against this specific failure mode.
Both can be true at once because they describe different stages of the deal. Pricing alignment is what brings deals to signing. The late-stage breaks are happening to deals that already cleared that bar. A few details sharpen the picture:
Deal value rose six times faster than deal count in 2025 — 40% versus 7%. Fewer deals, at materially higher prices.
M&A's share of corporate cash deployment hit a 10-year low of 7% across roughly 700 S&P World Index companies, versus a 9% to 17% range over the prior nine years. Capital is being rationed, even within the rebound.
The failure mode that ended deals in 2022 has been replaced by something else.
Three diligence workstreams now decide whether deals close
Three quarters of strategic acquirers now formally assess AI's impact on the target's business as part of confirmatory diligence — a structured workstream that did not exist in this form in the previous cycle. PwC's 2026 Global M&A Industry Trends calls AI due diligence "essential", describing it as a forward-looking assessment of "AI's potential impact over three to five years.”
Two other workstreams now operate alongside it: regulatory exposure under the EU AI Act, and training-data provenance. Both used to sit inside the broader legal and technical diligence stacks; in the past year they have been pulled out as standalone workstreams with their own scope, owners, and deliverables. Any one of the three can end a deal late.
Buyers now run dedicated workstreams that test the target's future, not just verify its past. When a finding lands that the seller cannot answer, the buyer walks. Pricing rarely reopens at that stage.
1. AI substitution and durability
Buyers test whether an AI-native competitor can replicate the target's core value at credible economics within the underwriting horizon, and whether the target's stated AI integration represents genuine workflow embedding or marketing language.
Bain's companion piece, New Diligence Challenge: Uncovering AI Risks and Opportunities, describes five questions sponsors and dealmakers now ask to test this. The output indicates whether the displacement defense is documented and falsifiable or not.
2. Regulatory exposure under the EU AI Act
Prohibited-practice provisions came into force on 2 February 2025. General-purpose AI obligations took effect 2 August 2025. High-risk system obligations follow on 2 August 2026, with full enforcement in 2027. Article 99 penalties reach €35 million or 7% of worldwide turnover, whichever is higher, for prohibited-use violations.
The Act has extraterritorial reach: any UK or US company providing AI-enabled services into the EU falls within scope. Orrick's November 2025 client guidance requires in-progress acquisitions covering AI products to have revised due diligence before completion.
3. Training-data provenance and IP exposure
This workstream moved fastest in the past 12 months, driven by litigation. The reference point dealmakers cite is the $1.5 billion Bartz v. Anthropic settlement, preliminarily approved in September 2025.
Bloomberg Law observed in May 2026 that "buyers are negotiating technical walk rights, allowing termination of the deal if a deep-dive audit reveals unverifiable model lineage or foundational licensing issues." Targets unable to produce a verifiable history of training-data acquisition channels are treated as carrying impaired assets.
The three workstreams share a common shape. Each produces findings late in the process, after months of advisor work and management attention have been committed. Each generates a question that the seller's historical performance cannot answer, and each gives the buyer a structured basis for walking, not negotiating.
Sellers are still preparing for the failure mode that mattered last cycle
The change happened fastest with training-data risk, because litigation gave buyers a number to point at. It also reset what reasonable diligence looks like:
A specific, falsifiable claim about why an AI-native competitor cannot replicate the target's core value within 18 to 24 months
Proprietary data accumulation curves that demonstrate a real moat, not a stated one
Embedded switching costs documented through customer workflow integration
Customer testimony about what would actually break if the AI features were removed
The forward case has to survive a quant team going through it line by line, not just a partner reading it on a Tuesday. The same logic plays out in investor communications, which is part of why we have written about how brand architecture shapes LP trust across fund cycles.
LP letters that describe what happened are losing ground to letters that interpret what is about to happen, and the underlying skill is the same: anticipate the structured question, pre-position the answer in writing.
The deals closing at premium multiples share three traits
Bain's data shows that 60% of 2025 deals valued above $1 billion were "scope" deals — top-line growth and capability acquisition — versus "scale" deals optimizing for cost synergies. Scope deals are inherently more dependent on the target's future than its past, which is the same shift that put the new diligence workstreams in the room.
The deals that close at the top of their valuation range share three characteristics:
The forward case is documented as rigorously as the historical case. Inside the data room, not just inside the management presentation.
The structured-workstream questions are pre-answered. AI substitution defenses, EU AI Act compliance positions, and training-data provenance logs are prepared with the same precision applied to QofE work.
The story does not drift between the IM and week six. What the buyer's diligence team encounters in the confirmatory phase is the same story that the buyer's principal heard at the IOI, with more depth and the same shape.
For PE sponsors, this means that the last 18 months of hold should be producing artifacts that survive the new diligence, not just KPI improvements that read well in a teaser. The same logic applies to post-acquisition brand and communications strategy, where the architecture decision made during or immediately after diligence shapes the eventual exit narrative years later.
Bottom line: Where the bar is heading next
The bar is still rising. EU AI Act high-risk obligations take effect on 2 August 2026, and the next round of US court rulings on AI training-data fair use is expected in the same window. Buyers are already extending the same forward-looking lens beyond AI — to customer concentration, regulatory exposure, and operational dependencies that look stable on a trailing basis but fragile on a forward one.
By the time the next sell-side process opens, the buyer-side framework will have moved another step beyond what most sellers are preparing for today. The sellers ahead of this used the hold period for documentation, not just performance improvement. By the time a buyer arrived, the answers were already written down.
For sponsors and corporate sellers planning a transaction inside the next two years, the focus should be on what gets put in writing now, while there is still time for it to read as infrastructure rather than a response to pressure. This is the work Collateral Partners is built around — translating a manager's forward case into materials that survive the diligence room before the diligence team arrives.
