EXECUTIVE SUMMARY
If your exit story calls the business AI-enabled, you are claiming a price. No public number supports it today. The multiples in circulation come from advisers who do not disclose what they measured, and EY describes a premium and a discount in words. What the record does show is the standard advisers state. PwC says companies "must show measurable income statement impact before commanding premiums." Bain reports PE firms rotating toward businesses "less exposed to near-term AI disruption." AI helps in some exit conversations and hurts in others, and the sentence in your CIM does not say which side you are on.
"AI-enabled" bundles three claims: one about the product, one about operations, one about exposure to AI-native competitors. A buyer tests each differently. The discipline we run for this is Claim-to-Evidence Substantiation: list every AI claim in the equity story, give each a line in the model to land on, attach the evidence a buyer can verify without you in the room, then prove it, restate it, or cut it. The Four-Lens Scorecard and the AI Custody Register supply the inputs. The ledger is the output.
THE PROBLEM
I. A Sentence With a Price Attached
Somewhere in the CIM for nearly every company going to market in the next 24 months, there is a sentence that says the business is AI-enabled. Ask the sponsor what a buyer should pay for that sentence, and the room goes quiet.
The documents that should fill the silence are confident about direction and thin on evidence.
In May 2026, EY published a point of view on industrial and manufacturing deals that opens with a flat assertion: companies with embedded AI capabilities are commanding premium valuations, and those without face a growing "AI gap discount." It attaches figures to individual levers, including a 10 to 15 percent increase in enterprise value from AI-powered demand sensing. The figure is not tied to a source, and the piece cites no transaction comparables. Its reference list names only general sources.
An M&A advisory firm's 2026 trend note puts private AI deals at "8x to 15x revenue for well-positioned companies," against 4x to 6x for traditional SaaS. The page does not say how many deals sit behind the range, over what period, or how the firm chose them.
The advisers who work with buyers read it differently. In June 2026 PwC's private equity outlook said: "Companies embedding AI into operations must show measurable income statement impact before commanding premiums." Grant Thornton, the same month, wrote: "Buyers are wary of paying premium multiples for assets that AI could fundamentally reshape." Neither sentence describes a settled price. Both describe buyers asking for proof.
Bain's midyear private equity report adds the portfolio view. Through March 31, "software valuations in PE portfolios declined by roughly 8% overall." And "PE firms are rotating capital and investment resources towards businesses perceived as less exposed to near-term AI disruption and macro volatility." Bain counts "an AI-driven rout in software valuations" as one of three shocks behind the setback, so the 8% is not a clean AI number. The direction of the rotation is still plain.
PwC says firms deploying AI to drive margin improvement "are differentiating themselves in fundraising and exit processes." Bain says capital is moving away from AI exposure. A seller holds one sentence that has to survive both readings. Does AI show up in this business's income statement, or in its competitive set? The sentence in the CIM answers neither.
THE DIAGNOSIS
II. The Sentence Has No Unit
A premium pays for cash flow the buyer expects to receive, and "AI-enabled" points at none. Better adjectives from the banker will not change that.
Read the sentence the way a buyer's diligence team does. It comes apart into at least three claims, because each lands on a different part of the model and gets tested by a different person.
The product claim. AI sits inside what the company sells. The buyer asks whose technology it is, what it costs to run, and whether a person does the work behind it. The SEC has shown what happens when the answers differ from the claim. Its January 2025 order against Presto Automation found that from November 2021 to September 2022 "the only units of Presto Voice that the company deployed used speech recognition technology owned and operated by a third party," and that human order takers "processed the vast majority of drive-thru orders placed through Presto Voice." Presto had told investors, in press releases, Forms 8-K and S-4, and registration statements, that the product eliminated human order taking.
Two other examiners point the same way. In March 2024 the SEC fined two investment advisers, Delphia and Global Predictions, $225,000 and $175,000 for false and misleading statements about their use of AI. In June 2025 Gartner estimated that only about 130 of the thousands of agentic AI vendors are real, and defined "agent washing" as rebranding assistants, RPA, and chatbots without substantial agentic capabilities.
None of those was a private sale. They matter because they show what an outside examiner does with an AI claim. The examiner asks whose model it is and who sits in the loop. A buyer has the same two questions, because the answers decide what the revenue is worth.
The operating claim. AI sits inside how the company runs. The buyer asks where the income statement moved. PwC states the standard in its own words: measurable income statement impact. Article 5 explained why operating claims struggle to meet it. One deployment can read high on the operational lens and flat on the financial lens, because released capacity does not become savings until someone decides where it goes.
Employee self-reports show a similar gap at scale. Gallup surveyed 23,717 employed U.S. adults in February 2026. Within organizations implementing AI, 65 percent of employees said AI had improved their productivity and efficiency, and only about one in 10 strongly agreed that AI had transformed how work gets done in their organization. A CIM that cites the first number and a buyer who asks about the second are describing the same company.
The position claim. AI helps the business and does not threaten it. The buyer asks whether the revenue they are underwriting could be replicated by something AI-native. Grant Thornton's phrase for the risk is "assets that AI could fundamentally reshape." Bain's phrase for what buyers want is "underwriting confidence." The seller's evidence is a retention curve cut by the segments an AI-native product would reach first. A buyer who does not receive that cut has only the risk to underwrite.
"A buyer pays for cash flow. 'AI-enabled' names a technology and no cash flow."
One sentence, three claims, no unit on any of them. The CIM leaves the attachment to the buyer, and the buyer makes it on the buyer's terms.
FIGURE 1 · ONE SENTENCE, THREE CLAIMS
As written, "AI-enabled" attaches to no line in the model, no evidence and no owner. After the ledger, each claim lands on a named line and is proved by evidence a buyer can check.
WHY THE ANSWER DOESN'T EXIST YET
III. What Nobody Has Measured
We went looking for the number a seller could put in a model. We did not find it, and the reasons are more useful than the number would have been.
The premium has no public measurement for the companies in your portfolio. The multiples above describe "private AI deals" set against "traditional SaaS," and the outliers the advisory firm describes are businesses "with proprietary models, defensible data assets, and strong retention metrics." That is the vocabulary of companies that sell AI, and the page's own title says "AI-Native Businesses." EY's premium describes industrial and manufacturing assets. We found no published measurement of what a buyer paid for a financial services or growth-stage technology company because it used AI internally, and none of what a buyer deducted for one that did not. That is a statement about what we found. The absence may owe something to our search, and we would like to read the deal that proves it wrong.
The premium and the discount have never been measured against each other. They appear in separate documents about separate kinds of company. A seller needs to know which one applies to the business in front of them. Nobody publishes the dividing line. The buyer draws it deal by deal, from whatever evidence the seller supplies. Grant Thornton's phrasing for the seller who supplies it is modest. Sellers who can quantify their AI investments "with defined ROI, proven use cases and measurable outcomes are commanding attention from buyers." Attention, in that sentence, is the reward. It is not a multiple.
When a buyer cannot price a claim, the claim moves into the contract. Article 4 traced this path for AI governance. Osler published recommended AI-specific representations and warranties in January 2025, covering ownership of AI assets, training-data quality and legality, and governance practices. Skadden wrote in January 2026 that AI assets "often require tailored diligence, including by specialized third-party diligence firms," and that where AI is a critical value driver, buyers "may request that specific AI-related representations be categorized as 'fundamental.'" In August 2026 Fasken wrote that if the diligence record is incomplete, representation and warranty insurers "may push for broader language," including terms excluding losses arising from a target's AI systems generally.
Read the verbs: often, may. These are lawyers describing what to ask for. None of them counted what buyers do. A claim a seller cannot substantiate does not disappear. It returns as a representation, an escrow, an earnout, or a carve-out, and the seller pays for it in structure instead of in multiple. That is our inference. Nobody publishes how often it happens.
So the position is narrow. The premium is asserted. The discount shows up in software portfolio marks, where Bain counts an AI-driven rout among the causes. The proof standard is stated by advisers who work with buyers. Nobody can hand a seller a price, but a seller can meet a standard.
THE METHOD
IV. Claim-to-Evidence Substantiation
The discipline we run at AWSM LABS for the exit is Claim-to-Evidence Substantiation. It treats every AI sentence in the equity story as a claim to prove, restate, or cut before the buyer's diligence team arrives. The instrument is a ledger with one row per claim. Six questions fill each row.
What is actually being claimed? Pull every AI sentence from the CIM draft, the management presentation, and the banker's model. Include the adjectives, such as "proprietary," "AI-native," and "AI-driven," and every figure that names AI as its cause. Each becomes a row. The claims that feel obvious go in first, because nobody has examined them.
Which claim is it? Product, operating, or position (Figure 1). A sentence that mixes types gets split. "AI has cut our cost to serve and made our product stickier" is two rows.
Where does it land? Name the line. For a product claim, that is a revenue line, a customer cohort, and gross margin after the cost of running the model. For an operating claim, a cost line in a named period. For a position claim, retention or concentration by segment. A claim with no line is the first finding.
What would a buyer verify without you in the room? Evidence that stands without a management explanation: a contract, a usage log, a ledger entry, a cohort table. For product claims, that means who owns the model, what the license says, and what share of outputs a person reviews or corrects. For operating claims, it means the Four-Lens Scorecard from Article 5 with its spread, plus the decision about where released capacity went. For position claims, it means retention by segment before and after the AI features shipped.
Who answers for it? Take the custody row from Article 4. A claim the company relies on and no one owns will surface as a request for a representation. Write down the named owner and the days a wrong output would run before anyone caught it.
Prove, restate, or cut? Grade each row as proven, partial, or asserted. Every asserted row gets one of three decisions. Build the proof before launch, if the calendar allows. Restate the claim at the level the evidence supports. Or cut it. Cut is a legitimate decision, because a claim that is not in the document cannot be found overstated in it.
The ledger applies Article 5's contradiction test to the sell side. Start with the claim that carries the most weight in the equity story and shows the widest spread across the four lenses. If the buyer reads a different lens than the seller did, that row is where the two readings part.
FIGURE 2 · THE LEDGER, ROW BY ROW
Every AI sentence becomes a ledger row. Six questions fill the row, drawing on the Four-Lens Scorecard and the AI Custody Register. The row ends in a decision: prove, restate, or cut.
WHAT IT SURFACES
V. A Composite, and Three Sentences
What follows is a composite. It describes no single client, and every figure in it is invented for illustration.
Picture a growth-stage B2B software company with $100 million to $150 million in revenue, five years into a sponsor's hold, preparing to go to market. The banker's first draft carries three AI sentences.
"Our proprietary AI engine classifies documents with industry-leading accuracy." A product claim, landing on the document-processing product's revenue and gross margin. The ledger finds that the extraction runs on a third-party model under a commercial license. The company's contribution is real: its fine-tuning data, its workflow, its review interface. Reviewers also correct a meaningful share of outputs before they reach customers, and no benchmark supports "industry-leading." Neither fact is a commercial problem. Both contradict the sentence. Status: partial. Decision: restate, as "built on a licensed model, with proprietary data and workflow." That version survives the two questions the Presto order turned on.
"AI has reduced our cost to serve by a quarter." An operating claim, landing on the support cost line. The Scorecard reads 5 on the operational lens: tickets handled per agent rose sharply. It reads 2 on the financial lens: support headcount did not change, and the released hours went into backlog and weekend coverage. Experiential reads 4, because agents like the tool. Custodial reads 2, because nobody owns the summarization tool by name. The spread is three. Status: asserted. Decision: restate as capacity ("support absorbed 25 percent more volume with no added headcount") if the volume data holds, and cut the cost figure. At flat headcount that volume is 20 percent less cost per ticket, so the banker's "quarter" overstated it.
"AI is a tailwind across our platform." A position claim, landing on net revenue retention by segment. One segment, the light-use customers who mostly run the basic reporting module, carries a meaningful share of revenue, and its retention has softened since AI-native alternatives appeared. Customers on the new AI features retain better, but that cohort is five quarters old. Status: partial. Decision: replace the sentence with the segment table and let the buyer read both curves.
Three sentences became three rows, and the rows produced three different decisions: one restated, one cut back to what the data supports, one replaced by a table. The CIM that went to market made fewer AI claims than the first draft, and each remaining claim had a document behind it.
We cannot say what the buyer paid for it. Nobody can, which is the premise of this article. What the seller gained was control over which questions arrived first.
THE FRAMEWORK
VI. Three Claims, One Ledger
FIGURE 3 · THE THREE CLAIMS INSIDE "AI-ENABLED"
| Claim | What the CIM says | What the buyer asks | Where it must land | Evidence a buyer can verify unaided | What the record says |
|---|---|---|---|---|---|
| Product | AI is inside what we sell | Whose technology is it, who is in the loop, and what does a dollar of AI revenue cost to serve? | A revenue line and cohort, plus gross margin after model cost | Model ownership and license terms; usage logs; share of outputs reviewed or corrected | SEC order against Presto: third-party model, human intervention. SEC actions against Delphia and Global Predictions: claims about AI use |
| Operating | AI is inside how we run | Where did the income statement move, and who decided where released capacity went? | A named cost line in a named period | Four-Lens Scorecard and its spread; the redeployment decision; the ledger entry | PwC: "measurable income statement impact." Gallup: 65 percent say AI improved their productivity and efficiency; about one in 10 strongly agree it transformed how work gets done |
| Position | AI helps us and does not threaten us | Could an AI-native alternative replace the revenue I am underwriting? | Retention and concentration by segment | Cohort retention before and after AI features; revenue concentration in workflows AI-native tools reach first | Grant Thornton: buyers "wary" of assets AI "could fundamentally reshape." Bain: capital rotating toward less AI-exposed businesses |
Read the second column against the fourth. A sentence in the CIM that cannot name where it lands is a sentence the buyer will land for you.
FIGURE 4 · THE SUBSTANTIATION LEDGER (COMPOSITE EXAMPLE)
| Claim as written | Type | Lands on | Evidence on file | Owner | Status | Decision |
|---|---|---|---|---|---|---|
| "Proprietary AI engine ... industry-leading accuracy" | Product | Document-processing revenue and gross margin | Third-party model under license; proprietary data and workflow; outputs corrected by reviewers; no accuracy benchmark | VP Product | Partial | Restate: licensed model, proprietary data and workflow |
| "AI has reduced our cost to serve by a quarter" | Operating | Support cost line | Scorecard 5 / 2 / 4 / 2, spread 3; headcount unchanged | None named | Asserted | Restate as capacity if volume data holds; cut the cost figure |
| "AI is a tailwind across our platform" | Position | Net revenue retention by segment | Light-use segment softening; AI-feature cohort retains better but is five quarters old | CRO | Partial | Replace the sentence with the segment table |
The ledger ends in decisions, and "cut" is a legitimate one. Finish it with no claim cut, then read it again.
FAILURE PATTERNS
VII. Four Ways Sellers Misread This
FIGURE 5 · FOUR EXIT-STORY TRAPS
| Trap | Observable symptom | The real problem |
|---|---|---|
| Pricing the sentence instead of the claims | "What multiple does AI-enabled get?" | No public measurement exists for companies like yours. The buyer has no market number to lean on and works from the claims it can verify |
| Borrowing an AI-native multiple | A model that cites 8x to 15x revenue for a company that uses AI internally | The cited range describes private AI deals, and its source discloses no sample, period, or selection rule |
| Substantiating after the banker writes it | "We will pull the support when diligence asks" | A seller who answers requests one at a time decides nothing about which claims are in the document. The cheapest edit is the one made before the CIM exists |
| Offering the operational lens as proof | "Tickets per agent are up" presented as savings | A buyer reads the financial lens. A real operational gain with no decision about released capacity appears nowhere in EBITDA |
The pattern under all four: the seller prices a story the buyer has not yet been given the evidence to believe.
There is a quieter version of the fourth trap. Each of the four lenses can produce a confident endorsement of a deployment that the other three would fail, and Article 5 argued that the disagreement between them is the finding. A buyer's team can read the lens the seller left out. The ledger is how the seller reads it first.
CONCLUSION
VIII. Write the Proof Before the Sentence
This series began with a measurement problem and ends with a pricing problem, and they are the same problem. The four lenses asked whether AI was working. Article 5 asked what the lenses say when they disagree. The exit asks what someone else will pay for the answer, and what they will accept as proof.
Nobody can give you that price today. In the documents we read, the premium is asserted with no transaction data behind it. The discount is visible in the marks. The one thing stated plainly, by advisers who work with buyers, is the standard: measurable income statement impact, defined ROI, proven use cases, measurable outcomes. You cannot choose the multiple. You can choose the order of operations.
Open your draft CIM, find the AI sentence that carries the most weight in the valuation, and write down the income statement line it lands on. If nobody in the room can name one, that sentence is row one of your ledger.
The Four-Lens Scorecard and the AI Custody Register supply the inputs, and subscribing brings the rest of the series.
Write the evidence first, then write the sentence the evidence will carry.
WORKS CITED
Bain & Company. Winning firms will focus on what they can control, weather the rest, as triple-shock brakes private equity's latest revival: Bain & Company 2026 Midyear PE Report. Press release, 8 June 2026. Practitioner research; the quoted valuation figure is a portfolio mark through March 31, 2026, and the release names an AI-driven rout in software valuations as one of three shocks.
EY. The AI valuation shift: a point of view for industrial CEOs and their deal teams. EY, 15 May 2026. Point of view on industrial and manufacturing assets; asserts a premium and a discount without transaction comparables, and its end-of-page reference list is not tied to specific claims.
Fasken. When AI Meets RWI in M&A: What Impact Will AI Have on Representation and Warranty Insurance? Fasken, 13 August 2026. Practitioner commentary; describes possible underwriting responses, not observed practice.
FE International. AI M&A Trends 2026: Why Acquirers Are Paying Premium Multiples for AI-Native Businesses. FE International, 21 April 2026. Advisory-firm commentary; the quoted multiple ranges carry no disclosed sample, period, or selection criteria.
Gallup. Rising AI Adoption Spurs Workforce Changes. Gallup, 12 April 2026. Probability-based panel; 23,717 employed U.S. adults, fielded 4 to 19 February 2026; margin of error plus or minus 0.9 points. Employee self-report, not financial results.
Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Gartner press release, 25 June 2025. Cited for its estimate of genuine agentic AI vendors and its definition of agent washing, not for the cancellation forecast.
Grant Thornton. AI is changing the M&A playbook. Grant Thornton, 17 June 2026. Practitioner commentary.
Osler, Hoskin & Harcourt. M&A transactions involving AI companies: representations and warranties. Osler, 21 January 2025. Practitioner commentary; recommended practice, not established market standard.
PwC. Private equity: US Deals 2026 midyear outlook. PwC, 17 June 2026, updated 16 July 2026. Practitioner commentary; the quoted sentences come from PwC's analysis and not from a survey of buyers.
Skadden, Arps, Slate, Meagher & Flom. M&A in the AI Era: What Buyers Can Do to Confirm and Protect Value. Skadden, 13 January 2026. Practitioner commentary.
U.S. Securities and Exchange Commission. In the Matter of Presto Automation Inc. Release No. 33-11352, 14 January 2025. Settled order; no civil penalty. The misstatements appeared in press releases, Forms 8-K and S-4, and registration statements.
U.S. Securities and Exchange Commission. SEC Charges Two Investment Advisers with Making False and Misleading Statements About Their Use of Artificial Intelligence. Press Release 2024-36, 18 March 2024. Penalties of $225,000 (Delphia) and $175,000 (Global Predictions).
\
AWSM DSPTCH
Get the dispatch.
Perspectives on AI activation — ROI, frameworks, and lessons from the frontier — sent when we publish. No noise.
