More than four software deals in five close with no public price attached. Your AI model didn’t know that, and it gave you a number to the hundred thousand anyway. You’ve been carrying it ever since.
I’ve spent three articles on what your acquirer models that you don’t — the integration tax nobody puts in the LOI, the headcount line item your CFO doesn’t carry, the savings that turn out to be funding somebody else’s capex. Each one was about arithmetic happening on the other side of the table.
This one is about arithmetic happening on yours, and the fact that nobody will ever tell you it was wrong. The AI you’re using certainly won’t.
A buyer who reads your materials and forms a view doesn’t send feedback. They decline, or they never engage at all. The cost of an AI generated valuation, a polished but naive CIM, or an ARR line that quietly includes services doesn’t arrive as criticism. It arrives as four bidders instead of seven, and a process that closes at a number you assume was the market’s verdict.
The arithmetic is fine — the evidence underneath it isn’t
You fed a Large Language Model (LLM) three years of revenue, EBITDA and growth, and asked what the company is worth. Back came a range — $42.3 million to $47.8 million, say, specified to the hundred thousand.
Here’s what that range rests on. Berkery Noyes tracked 6,177 software transactions between 2023 and 2025. Exactly 1,117 disclosed financial terms. That’s 18%. More than four deals in five closed with no public number attached.
And the 18% isn’t a random sample. Larger deals are more visible, and size moves pricing hard: transactions between $10 million and $160 million cleared at a median 2.6x revenue, against 5.4x above $160 million. So, the comparables are incomplete, selected for visibility, and — unless you’re above $160 million — drawn from a size class that prices differently than you do.
The math on top of them is entirely correct.
That’s not a hallucination. It’s false precision, and for a seller it does more damage than a visibly wrong answer, because you’ve now seen the number and you can’t unsee it. It sets the price at which an approach feels insulting. It decides whether a serious buyer at $31 million gets a conversation or a dismissal. And a dismissed buyer produces no feedback, so the number never gets corrected. Eighteen months later you’re still anchored to a figure that was generated in ninety seconds from the wrong comparable set.
I call it the valuation confidence gap: apparent precision divided by the quality of the evidence behind it.
The version of this that should worry you more is that your buyer is running the same kind of tool on you — with the same denominator problem, and no obligation to show you the output.
Grounding doesn’t make it true, and you’re the one who gets retraded
The standard reassurance is that serious LLM tools retrieve from real databases and cite what they find, so they can’t invent things.
That’s been tested. Stanford researchers evaluated the leading retrieval-grounded AI legal research tools — purpose-built systems from LexisNexis and Thomson Reuters, sold into a market where a fabricated citation ends careers. Both produced incorrect or mis-grounded output more than 17% of the time.
Mis-grounded is the category that matters here. It means the citation is real, the document exists, and it simply doesn’t say what the output claims. That failure survives a spot check, because the spot check confirms the document is there.
Now put that in your data room. A buyer’s AI diligence tool reads four hundred of your documents and flags a change-of-control provision, citing the contract. Maybe it’s right. Maybe the provision is in a terminated agreement, or applies to a subsidiary you sold, or says the opposite of what the summary says. The buyer’s associate isn’t going to read the sixty-page agreement the AI tool was bought to avoid reading.
What reaches the investment committee is a finding. What reaches you is a retrade, a larger holdback, or an indemnity you didn’t expect — attached to a fact nobody verified. You will be arguing against a citation, and you’ll be doing it late, under exclusivity, with no leverage.
That’s an argument for reading your own contracts before they read them.
Calculation accuracy is not underwriting accuracy
Financial analysis is the subtlest trap, because numbers look objective.
A LLM will calculate ARR, churn, net retention, cohorts, concentration, working capital and cash conversion — increasingly very well. Underwriting requires knowing what those numbers mean.
One software company has contractual SaaS revenue. Another has services clients who happen to return annually and get counted as recurring. A third is usage-based. A fourth sells multi-year licenses. A fifth bundles software, implementation and hardware into a single line.
Five risk profiles, and five sets of financial tables that look nearly identical.
And the error compounds. An ARR figure that quietly includes recurring services flows into retention, then cohorts, then growth durability, then the multiple you’re asking for. One definitional error, reproduced faithfully across nine exhibits, reads as a consistent picture. Consistency isn’t corroboration, though it looks exactly like it in a well-formatted deck.
Here’s the asymmetry. If you make that error, you don’t find it. Your buyer’s quality-of-earnings provider does, in week six of due diligence, after you’ve signed exclusivity and stopped talking to the other two bidders. At that point the conversation isn’t about whether your revenue is good. It’s about whether your numbers can be trusted — and that question, once asked, gets applied to everything else you’ve represented.
A definitional error costs you a few points on your multiple. A credibility question costs you the process.
Your CIM can be complete and still persuade nobody
A LLM will produce an impressive CIM in minutes. Market overview, company history, growth strategy, competitive landscape, management biographies, financial commentary, investment highlights. The prose is often better than what a junior banker would write.
But a CIM isn’t a writing exercise. It’s an argument built for a specific, sophisticated reader, and it anticipates how that reader will underwrite your business.
How durable is this revenue? What actually drives retention? How concentrated is the customer base? How much of the growth is organic? What do you have that a buyer can’t build in eighteen months? What happens if you leave? Which risks are real, and which apparent weaknesses can be explained credibly? And the one that decides the outcome: why should this company matter to this buyer in particular?
A language model optimizes for plausibility, completeness and polish. An advisor optimizes for buyer conviction. While those objectives overlap, they aren’t the same.
So, the risk isn’t a visibly bad CIM — anyone spots that, including you. It’s a beautiful, complete, strategically naive one. It covers every section. But it answers no question the buyer was actually asking. And because it looks finished, nobody on your side flags it.
Read your own CIM the way the buyer will: skip to the two issues that would kill the deal and see whether the document addresses them or routes around them.
You are already on somebody’s list, and it’s the wrong somebody
Research cost used to limit acquisition outreach, and that limit did useful work. It forced buyers to choose.
A corp dev team that could research 500 companies a year now identifies 50,000, runs a second model to research each one, a third to draft a personal note to every CEO, and a fourth to handle follow-up. Multiply that by every acquirer in your sector.
Most of those lists are built from product adjacency. You sell healthcare workflow software; the buyer sells healthcare software; you look plausible. That’s semantic similarity wearing a thesis costume, and it’s why the approaches filling your inbox are mostly from buyers with no particular reason to want you.
The one that matters reads differently. Not “we operate in healthcare software,” but: “this buyer sells into hospital finance departments, has told investors twice that clinical workflow is its next expansion, tried to build a scheduling product and killed it eighteen months in, has completed four acquisitions in your size range since 2023, and hired a divisional president out of your closest competitor.” That buyer pays more than the other ninety-nine, and no amount of semantic similarity produces the sentence that identifies them.
Now the part that costs you money. When a message says “I was particularly interested in your recent expansion into…”, it might reflect a week of real work. It might have cost an agent three seconds. You can’t tell. Once personalization is universal, personalization stops signaling anything — and the rational response to fifty plausible approaches a month is to ignore all of them.
Which means the buyer with a genuine thesis and no relationship now has a harder time reaching you than before the tooling arrived. You will never know about the approach you deleted. That’s not a hypothetical cost. It’s the highest bidder in a process you never ran.
Dealslop, and why the seller pays for it twice
There’s a term for output that looks finished and quietly hands the remaining work to whoever receives it. Workslop, coined in Harvard Business Review by researchers from BetterUp Labs and Stanford.
Of 1,150 U.S. desk workers surveyed, 40% had received workslop in the previous month, spending an average of one hour and fifty-six minutes on each instance. From there the researchers extrapolate to roughly $186 per month per affected worker — their own back-of-the-envelope figure, and worth treating as an order of magnitude. The measured part is the part that matters: all of it describes work that was completed, delivered, and looked fine. Nobody complained. The effort just moved downstream.
M&A has its own version. Dealslop. A valuation carried to one decimal on comparables that don’t apply. A CIM of polished generalities. A financial package resting on an undefined revenue line. A management presentation that answers questions nobody asked.
In an office, workslop costs the recipient two hours and some resentment. In a sale process, the recipient is a buyer who won’t tell you what went wrong. They read the materials, form a view, and decline — or price the uncertainty into their opening number.
You pay twice. Once in the bidders who never engage, and once in the ones who do, at a discount that reflects work they now expect to have to redo. Neither cost appears anywhere you can see it.
You see interest as thin — you conclude the market is soft.
The base rate your buyer is underwriting against
Baruch Lev and Feng Gu analyzed roughly 40,000 acquisitions worldwide over forty years and found 70% to 75% failed — measured against post-acquisition sales and margin trends, stock performance, and goodwill write-offs.
That’s the record before generative AI touched anything, and your buyer’s investment committee knows it. It’s why diligence has been getting harder, not easier, and why every unverifiable claim in your materials costs you more than it did five years ago.
It also explains something sellers consistently misread. A buyer who asks you to substantiate a number isn’t being difficult. They’re operating inside an institution that fails three times in four and has finally started to care why. The causes Lev and Gu identify are judgment failures, not capacity failures: targets too large to absorb, unrelated businesses with no real economic overlap, broken companies bought on optimism, incentives that reward closing rather than succeeding.
The sellers who do well in this environment are the ones who make the buyer’s judgment easier rather than harder — who arrive with the evidence already assembled, so the committee’s skepticism has somewhere to land.
The counterargument
None of this makes AI bad for sellers.
McKinsey describes one business software company that used a third party’s AI-powered scouting platform — gen AI paired with a proprietary database of more than 40 million public and private companies — to score more than 500 targets in under a day, ending in three completed acquisitions. In McKinsey’s survey, respondents using gen AI in M&A report an average cost reduction of roughly 20%, and 40% report deal cycles 30% to 50% faster. PwC cites research across 61,000 early-stage ventures where LLM screening ran 537 times faster than human analysts at comparable categorization quality and outperformed them on cluster separation and compactness by 70%.
That last figure matters. The screening wasn’t only quicker. It was better. And a faster, better-targeted buyer universe is good news for a seller, because it raises the odds that the buyer who should want you actually finds you.
Deloitte’s 2026 study puts gen AI adoption in M&A at 90% of organizations, with 52% using it in post-close integration — and human review still the top requirement for high-stakes uses. When read next to the Stanford error rates, that looks less like institutional caution and more like a correct reading of the failure rate.
The distinction isn’t human versus machine. It’s AI used to sharpen judgment against AI used in place of it. And on your side of the table, the second one is a lot easier to do by accident.
What to check before any materials leave the building
Models are excellent at questions with checkable answers: which of these files contradict each other, where does the change-of-control language sit, which customers explain the revenue decline. They’re unreliable at questions requiring someone to be accountable for being wrong: what is this worth, should this buyer be on the list, will this capability still matter in three years.
Make sure you have answers to these five questions before anything goes out:
1. What’s the denominator behind my number? How many comparable transactions actually disclosed terms, and were they in my size class? If you can’t answer, the range is an opinion with decimals — and it should not be setting your reserve price.
2. What exactly is inside the ARR line? Contractual SaaS, repeat services, usage, multi-year licenses, bundled implementation. Write the definition down before anyone builds a cohort chart on top of it. This is the single most common place a process turns from a valuation conversation into a credibility conversation.
3. Have I read the contracts a diligence tool will read? Change of control, assignment, most-favored-nation, auto-renewal, termination for convenience. You want to know what’s in them before someone else’s model summarizes them, incorrectly, to an investment committee.
4. Does my CIM answer the two questions that would kill the deal? Not “does it have a competitive landscape section.” Identify the objections, then check whether the document meets them or routes around them.
5. Who is the buyer that pays more, and can I write the sentence that says why? If the answer restates product adjacency, you have a list, not a thesis — and you’re running a process that treats your best bidder as one of a hundred.
Turn the last one around and it stops being a checklist and becomes an advantage. Most sellers arrive with a generated list of acquirers. Arriving with five names and a real sentence after each is a materially different conversation.
What’s actually scarce now
The real equation isn’t the one in the productivity deck:
Net AI M&A Productivity = Automation Gain + Decision Quality Gain − Attention Externality − AI-Induced Decision Error
The first two terms are measured and quoted everywhere. The last two aren’t measured at all, which is a poor reason to treat them as zero. Both have already appeared in this piece.
Attention externality is the cost of universal outreach, and it lands on you. Fifty plausible approaches a month are somebody’s productivity gain. Together they train you to stop reading, which means the buyer who actually had a thesis about your company can no longer reach you.
AI-induced decision error is the outcome you get wrong because the analysis was AI generated rather than built. The reserve price set from the wrong comparable set. The retention curve resting on a contaminated ARR line. The approach you deleted. Each read as a finished answer, which is exactly why none of them gets checked.
Information used to be the constraint in this business. Finding buyers was hard, intelligence was expensive, analysis took weeks. AI is removing all three, fast.
What remains scarce is judgement and conviction — and in a sale, what you’re actually selling is a buyer’s conviction about you. Someone on their side has to say: this company matters, this thesis holds, this price survives committee. And someone has to answer for it when that call is wrong. A model produces a number and bears no responsibility. It doesn’t sit in front of an investment committee in week ten explaining why the revenue wasn’t what the deck said.
Your job as a seller is to make that person’s case easy to make and hard to attack. Everything above is a way of removing the reasons they’d have to hedge.
What would change my mind. If decision-quality measurement catches up — if acquirers start publishing post-close outcomes against that 70% to 75% base rate and the number moves — then the speed gains are buying something real, and generated work product is less of a liability than I’m arguing.
That data doesn’t exist yet. Until it does, the buyer across from you is reading polished, confident, unverifiable material all day, and has learned to discount it.
Give them something they can verify. That’s the whole strategy. And AI can’t do it for you.
References
Berkery, Noyes & Co. (2026, January 7). Full year 2025 software industry trends. https://berkerynoyes.com/full-year-2025-software-industry-trends/
Deloitte. (2026). 2026 Deloitte generative AI in M&A pulse study. https://www.deloitte.com/us/en/what-we-do/capabilities/mergers-acquisitions/articles/m-and-a-generative-ai-study.html
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. See also 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://www.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. Preprint: https://law.stanford.edu/wp-content/uploads/2024/05/Legal_RAG_Hallucinations.pdf
Niederhoffer, K., Kellerman, G. R., Lee, A., Liebscher, A., Rapuano, K., & Hancock, J. T. (2025, September 22). AI-generated “workslop” is destroying productivity. Harvard Business Review. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
Peck, E. (2025, September 24). AI “workslop” sabotages productivity, study finds. Axios. https://www.axios.com/2025/09/24/ai-workslop-workplace-efficiency-study
Yenulevich, R., Desai, K., & Janssen, B. (2026, May 27). M&A: Why dealmakers need their own AI capability. PwC. https://www.pwc.com/us/en/services/consulting/deals/library/dealmaker-ai-capability-deal-origination-execution.html
By Ivan Ruzic, Ph.D. · Originally published on Medium