Your AI can hand you a two-hundred-name buyer list before breakfast. The distance between that list and knowing which twelve will bid is the entire fee. Advisory work is about to be repriced, in both directions.
I’ve spent most of this year writing about where AI savings actually land — the integration tax, the demo-to-production gap, what happens to a margin story once the tools get really good. This piece sits on the other side of the table: what generative AI is doing to the person you would hire to run your exit.
The short version probably isn’t the one you are hearing. Generative AI splits M&A advisory into two halves and prices them very differently — one near zero, the other considerably more expensive than today.
If you are a founder heading toward a transaction, that split becomes your problem, because you will have to choose between the halves with less information than ever.
Picture two AI systems
The first retrieves everything public: filings, industry reports, press releases, transaction databases, earnings calls. It is fast and thorough, and within about a year it will be close to free.
The second holds all of that plus twenty years of buyer interaction — who engaged and who declined, the reasons they gave and the real reasons they had, which arguments for merging actually persuaded a corporate-development committee, and which diligence findings moved price by how much.
Identical base model; very different M&A intelligence.
That gap matters because it points away from the obvious conclusion that an advisor is becoming irrelevant. A model can research companies, find buyers, build market maps, analyze financials, draft a CIM, summarize a data room, pull comps, and read contracts — all in minutes.
If producing information was where advisory value sat, this would be an extinction event. The premise is wrong, and it gets more wrong every quarter.
Adoption and its limits
Deloitte’s 2026 Generative AI in M&A Pulse Study found that 90% of surveyed U.S. private-equity, portfolio-company, and corporate M&A leaders were either actively using or piloting AI; 70% reported active use and 20% were experimenting or piloting. The same study names the top requirement for high-stakes use: human review.
McKinsey’s January 2026 article reports findings from a survey run in February 2025, with users seeing roughly a 20% average cost reduction and 40% reporting deal cycles that ran 30% to 50% faster.
It’s easy to conclude that adoption is near-universal and moving into the later, higher-stakes phases of a deal. But, organizations doing this daily have discovered from experience that unsupervised output can’t be trusted. That finding matters more than any adoption statistic.
The two-hundred-name list
An advisor hands a seller two hundred potential buyers. Ten years ago, the number proved effort; today a model produces it before breakfast, so the count itself carries roughly zero useful information.
The questions that survive are the harder ones: which thirty matter? Which ten hold an acquisition thesis that will survive internal scrutiny? Which business unit sponsors the deal on its own budget? What size deals does this buyer actually complete, against what its press releases suggest? Which obvious candidates will never act?
A real entry reads less like “operates in healthcare software” and more like:”sells into hospital finance departments; has told investors twice that clinical workflow is next; tried to build a scheduling product and killed it eighteen months in; completed four deals between $30 million and $90 million since 2023; just hired a divisional president out of a direct competitor of the target.”
That single name carries more information than the other one hundred ninety-nine.
Calibration over calculation
Berkery Noyes tracked 6,177 software transactions from 2023 through 2025. Financial terms were disclosed for 1,117 — about 18% — leaving 5,060 without public terms. The disclosed deals totaled $662 billion; Berkery Noyes projected an additional $198 billion in undisclosed value, or roughly $860 billion overall. A quarter of the market’s worth is therefore estimated rather than recorded, and that $198 billion means that most of the undisclosed transactions were comparatively small.
Size moves everything. Median revenue multiples ran 2.6x on deals between $10 million and $160 million, 5.4x above that threshold, and 3.2x across all segments. A model computes each one instantly; the hard part is knowing which multiple applies to this company — how durable its growth really is, whether revenue is economically recurring or merely labeled that way by management, what diligence will surface, and what buyers price in when it does.
Retrieval-grounded tools still hallucinate
The standard defense is that serious tools retrieve from real data and cite what they find, so fabrication shouldn’t be possible. Stanford researchers tested exactly that on the leading retrieval-grounded AI legal research products — purpose-built systems from LexisNexis and Thomson Reuters sold into a market where a fabricated citation ends a career. Both hallucinated between 17% and 33% of the time.
The dangerous failure is misgrounding: the citation is real, the document exists, and it says nothing like what the output claims — which passes a spot check. In a deal, when a tool flags a change-of-control provision and cites the contract, verifying means reading the sixty-page agreement the tool was bought to avoid, then multiplying that by four hundred documents.
Retrieval-grounded systems can still produce serious errors in high-stakes professional settings. When taken together with Deloitte’s finding that human review remains the top requirement for high-stakes use, it’s the clearest signal in the data about where advisory value migrates.
Deal memory as a moat
Public data tells a model which companies make acquisitions: it reads earnings calls, lists past deals, and maps portfolios. That helps. But acquisition decisions get made by organizations, and organizations are made of people with histories.
An advisor two decades into a market may know that the obvious buyer looked at a nearly identical company last year and walked because absorbing it became politically difficult; that an unlisted buyer just hired a business-unit president with a mandate to enter this exact market; that an acquirer known for deals twice this size has quietly started doing tuck-ins.
That is what separates model intelligence from model intelligence plus proprietary deal memory. Every process, including the failures, generates data that should strengthen the next one — a buyer whose process died in week nine tells you more than three who closed on terms.
The operational problem here is unglamorous and mostly unsolved: a stated reason for declining is easy to write down but frequently lies, while the real reason surfaces on a call three months later and gets written down nowhere. That is an organizational-design failure before it’s a technology one, which explains why so few firms have fixed it despite universal agreement that it matters.
It’s important to remember that deal memory also decays — an entry from 2019 may describe an organization that no longer exists — and acquirers are building their own stock of it with a structural advantage, since they sit on every conversation they’ve ever had.
The edge is very real, but it carries a shelf life.
What the failure base rate shows
Any argument for advisory judgment has to confront the record.
Lev and Gu analyzed roughly 40,000 acquisitions across four decades and found that 70%-75% failed to improve one or more outcomes such as post-acquisition sales growth, cost savings, or the buyer’s share price; advisors sat in the room for many of them.
The causes they identify are the interesting part, because none is an information failure: targets too large to absorb, unrelated businesses with no real economic overlap, broken companies bought on the belief that good management fixes them, and incentives that reward closing rather than succeeding. Each starts as a conviction problem — somebody wanted the deal and built the case backwards from the decision.
An advisor who produces supporting material for an already-made call adds to that failure rate instead of reducing it. An advisor willing to say this buyer will not pay what you think, or this revenue will not survive diligence, or you should not do this deal, works on the only variable that has ever moved the number.
AI makes the first kind far cheaper to replace and the second considerably more valuable.
Accountability
An AI system can state an enterprise value of $45 million without bearing any consequence if the company could have sold for $60 million. It can recommend Buyer A when Buyer B would have paid more, or recommend accepting an offer while staying out of the room where the CEO and board make the call.
A client buys an answer, and inseparably from it, somebody who will be present and answerable when the answer turns out to be wrong.
For a founder selling twenty years of work, that’s not abstract: the decision is irreversible and happens once. A second opinion that costs nothing and answers for nothing is worth what you paid.
Who gets repriced first
Advisors whose value rests on producing materials and running routine analysis face genuine commoditization, because AI does exactly those tasks very well. Advisors who refuse to adopt the tools lose to peers who research more broadly and prepare more thoroughly.
The equation worth tracking is:
Experienced M&A judgment × AI capacity × proprietary deal intelligence
Good judgment becomes substantially more productive; poor judgment becomes substantially more scalable. The terms multiply, so a zero anywhere collapses the result — which is why an advisor with great instincts and no tooling usually lands in the same place as one with great tooling and no instincts.
Two sellers
Seller A uses a model to value the company, write the CIM, generate two hundred buyers, draft outreach, and analyze offers.
Seller B works with experienced advisors who use AI to stress-test the revenue base, challenge the valuation, test positioning against real buyer priorities, surface non-obvious strategic buyers, track how buyer behavior shifts through the process, and compress diligence.
Both run on AI; only one is buying judgment.
Seller A will never know what the process failed to surface. The auction produces four bids when nine serious buyers were actually out there, and from inside it looks exactly like a healthy four-bidder sale. Nothing tells them otherwise
The question was always whether AI stands in for expertise, or extends it.
Five questions for an advisor
If you are within two years of a transaction, take these first four into your next advisor conversation; any capable model makes them cheap to ask and expensive to bluff. The fifth becomes answerable only once an offer lands, and it tells you whether the first four were answered honestly.
· Which fifteen of the two hundred company buyer list will actually move? Volume is free; conviction mapping isn’t. Ask for names, the internal sponsor, and the size range each buyer has genuinely completed -not what its press releases suggest.
· What did you learn from your last failed process? Any advisor can show closed deals. The one worth hiring can describe a process that died in week nine, name the buyer, and say what it revealed; that memory is the moat.
· Which diligence finding will move the price, and by how much? Pre-registration of judgment — and a vague answer is a tell.
· What would you tell me not to do? If the answer just flatters the deal, you’re not buying anything of value.
· Once an offer lands: what does it say about the buyer’s conviction, and how hard can we push? Reviewing the offer is a task requiring judgment. It can’t be rehearsed in a first meeting — only discovered when the bid arrives.
The value of verified expertise
Generative AI has made one thing extraordinarily cheap: the appearance of expertise.
A founder can generate an investment-banking-quality document, and a junior corporate-development analyst can produce a sophisticated-looking acquisition case; none of it means the judgment underneath improved.
The more AI commoditizes the artifacts we take as evidence of expertise, the more verified expertise is worth. Credentials were always a weak proxy for that; and now deliverables have become a worthless one.
Two items remain undecided and I’m watching both:
· whether human review remains at the top of Deloitte’s requirements list next year, and
· whether corporate acquirers’ in-house deal memory starts closing the gap on the advisor’s.
If the first breaks, this argument breaks with it. If the second does, repricing gets faster and less forgiving — for advisors and for the founders choosing between them.
Good advisors will produce far fewer things by hand. The job reduces to deciding what matters, what is credible, what to ignore, who is serious, where the pressure sits, what the company is really worth, and what the client should do.
AI makes those advisors faster; it will also make their judgment more valuable.
References
· Berkery, Noyes & Co., LLC. (2026). Full year 2025 software industry trends. https://berkerynoyes.com/full-year-2025-software-industry-trends/
· CFO Dive. (2026). Generative AI reduces M&A costs by 20%, McKinsey says. https://www.cfodive.com/news/generative-ai-reduces-merger-acquisition-costs-20percent-mckinsey/812514
· Deloitte. (2026). GenAI in M&A 2026 pulse survey. https://www.deloitte.com/content/dam/assets-zone3/us/en/docs/services/merger-and-acquisitions/2026/gen-ai-pluse-survey.pdf
· Kordestani, K., Silva, R., & Berbel, J. (2026, January 14). Gen AI in M&A: From theory to practice to high performance. McKinsey & Company. https://www.mckinsey.com/capabilities/m-and-a/our-insights/gen-ai-in-m-and-a-from-theory-to-practice-to-high-performance
· Lev, B., & Gu, F. (2024). The M&A failure trap: Why most mergers and acquisitions fail and how the few succeed. Wiley.
· Lev, B., & Gu, F. (2024, November 13). We analyzed 40,000 M&A deals over 40 years. Here’s why 70–75% fail. Fortune. https://fortune.com/2024/11/13/we-analyzed-40000-mergers-acquisitions-ma-deals-over-40-years-why-70-75-percent-fail-leadership-finance/
· Magesh, V., Surani, F., Dahl, M., Suzgun, M., Manning, C. D., & Ho, D. E. (2025). Hallucination-free? Assessing the reliability of leading AI legal research tools. Journal of Empirical Legal Studies, 22(2), 216–242. https://doi.org/10.1111/jels.12413
Will AI Replace M&A Advisors? The Profession Is Splitting in Two was originally published in DataDrivenInvestor on Medium, where people are continuing the conversation by highlighting and responding to this story.
By Ivan Ruzic, Ph.D. · Originally published on Medium