Every Company Claims AI. Here Is How to Tell When It Is Real.

The deal team has seen this presentation before. A founder walks through a vertical AI platform built for financial services. The demo is good, really good. It surfaces insights in real time and flags anomalies a human analyst would miss. The CIM calls it "proprietary AI" with a "self-reinforcing data advantage." Everyone nods, because everyone has sat through a version of this a dozen times this year. Different logos, same three words.

What the team doesn't have is a reliable way to know, sitting right there, whether any of it is real. And that gap has a price. On a business doing $10M in ARR, the difference between an AI-native multiple and the one the company earns once its architecture is examined runs $30M to $50M. The premium is big enough to reorganize the deal, and the question of whether it applies usually goes unasked.

The Premium Is Real. So Is the Gap.

Start with what the market pays. In the private lower middle market, traditional SaaS clears 4 to 6x revenue. A genuinely AI-native business, one with data it owns and models it has trained sitting at the center of the product, trades at 8 to 15x, and the strongest assets push past 20x.¹ At the top of the public market, the highest-flying AI names have traded at 25 to 30x while conventional software sits in single digits.² Buy-side research pegs the clean premium at 1 to 3x revenue over a comparable non-AI peer.³

This is where the capital is going. Bain found that almost half of all strategic technology deal value above $500M in 2025 came from AI-native companies or deals that explicitly cited AI benefits, a sharp jump from the year before, inside a global M&A market that grew 40% to $4.9 trillion.⁴ AI M&A hit 266 closed deals in the first quarter of 2026 alone, up 90% year over year,⁵ against worldwide AI spending forecast at $2.59 trillion for the year.⁶ The premium is documented, and it is growing.

Now the gap. Two-thirds of buyers say they see only limited AI adoption in the companies they evaluate, even though nearly all of those companies claim AI in their marketing.³ McKinsey's read is starker. Only about 6% of organizations qualify as AI high performers, the ones that trace 5% or more of their EBIT to AI and report real value from it. The factor that separates them most cleanly is fundamental workflow redesign, and only about a fifth of companies have bothered to do it.⁷ So the premium and the reality have come apart, and they have come apart worst in the middle market, where technical diligence gets the least funding. Both errors cost real money. Overpay, and you have bought a premium on an asset that will miss its thesis. Pass when you shouldn't have, and you have walked away from a compounding business because you could not confirm the moat.

The Three Tiers

The distinction a deal team needs is architectural. It sorts into three tiers.

A Tier 1 business is AI-native. The product was built around AI from the first sprint. It runs on data the company controls and on models it trained or fine-tuned and owns outright, and the result is a flywheel: usage generates data, data improves the model, the better model improves the product, and the better product pulls in more usage. That is what earns 8 to 15x, sometimes more.

A Tier 2 business is AI-integrated. It has adopted third-party AI in a way that genuinely improves the product and shows up in customer outcomes, but the models underneath are available to any well-funded competitor. The AI is real. The moat is thinner. A modest premium is fair, and diligence should go toward whether the edge is durable or just early.

A Tier 3 business is AI-washed. Legacy architecture, an LLM wrapped around it, and no real change to the customer's outcomes or the unit economics. Look closely and it is worth whatever its non-AI numbers support. The danger is that these companies get priced as Tier 1.

Here is why the line sits at the architecture. The model is becoming a commodity. Gartner now files foundation models under "strategic commodities," which is to say performance alone won't hold one company apart from the next for long.⁸ What holds is the data that trains the model and the machine learning know-how sitting inside the team. A wrapper around someone else's API gives you neither.

The pattern is easy to spot once you have seen it. A company presents an "AI-driven risk engine" as its core differentiator and prices off it. Diligence opens the system and finds a hand-maintained rules table, the same architecture the company has run since the 1990s, with a general-purpose chatbot bolted on top that accounts for a low single-digit slice of what the product actually does.⁹ The demo worked. The valuation rested on a capability that was barely there. And it happens often enough to have produced real consequences: SEC enforcement against the public company Presto Automation, charged in January 2025 for overstating an AI product that leaned on a third party and heavy human intervention; criminal fraud charges against the founder of the shopping app Nate, who raised more than $40M on AI claims while hundreds of workers in an overseas call center actually processed the orders; and the collapse of Builder.ai, once valued at $1.5 billion and now insolvent and under federal investigation, with allegations that human engineers did much of the work its AI was credited for.¹⁰¹¹¹²

So here are the four markers of the real thing: training data the company owns with clean provenance, models it owns rather than rents, customer outcomes you can trace to the AI in the actual numbers, and machine learning capability that shows up in a team interview and a codebase review.

The Questions, and Where the Answers Are Not

The framework earns its keep only if it turns into questions you can ask in a management meeting. Four do most of the work.

What data does the company own, and what edge does it create that a competitor with API access to the same model can't replicate? If nobody can spell that out concretely, the edge probably isn't there.

If the company's AI vendor killed its model tomorrow, what breaks, and how fast? An AI-native business has an answer ready. An API-dependent one finds out it just has a vendor.

Where in the retention and expansion data can you actually see the AI working? Real improvement shows up at the cohort level, in net revenue retention driven by expansion rather than one more feature nobody leans on. Ask for the cohort data, not the headline NRR number.

Can it survive the outside-in test? Bain's diligence teams build quick prototypes out of off-the-shelf tools to see whether a target's AI can be rebuilt in a couple of weeks.¹³ If it can, it was never a moat. That is one test among several, but it is a fast way to tell architecture from story.

Then there is the part that has shifted most recently. AI claims now carry legal weight. The SEC set the precedent with its first AI-washing actions in March 2024 and pushed it onto a public company in 2025, and the DOJ has since brought criminal charges.¹⁰¹⁴ In private M&A, a claim that does not survive technical diligence hands the seller indemnity exposure and the buyer a repricing risk it took on blind. Deal structures are adjusting to match: earnouts that pay out only when the AI hits real benchmarks, escrows that release only after the architecture checks out.¹⁵ The technical review that prices the AI claim is turning into standard deal protection.

The founder is still at the front of the room. The demo is still good. The CIM language is still aspirational. What is different, for a team that has learned to sort the tiers, is that they know which questions pull the architecture out from under the narrative. The multiple the AI claim commands is large, and it is real. The work to confirm whether it belongs to this company is specific, and it is worth doing before the check clears.

Prepared June 2026. All figures validated against the live sources cited below.