Underwriting the AI-Native Target: Six Questions the Model Does Not Ask

A target arrives on the desk describing itself as AI-native. Revenue is growing at ninety percent. Gross margin is sixty-two percent, which is lower than software but higher than services. The founder is technical, the deck has a section on the proprietary model, and there are eleven logos on slide four. The banker is guiding to a revenue multiple because EBITDA is negative and that is the convention in the category.

The standard underwriting model does not know what to do with this. It was built for businesses whose competitive position is legible: contracts, switching costs, route density, brand, regulatory licence, installed base. This company's stated moat is a model, and models are the fastest depreciating asset class in the modern economy.

What follows is a structure for the conversation. It applies to a company built entirely on AI and, increasingly, to any target claiming AI as a differentiator.

Question one: which layer does the company actually own?

Nearly every disagreement about AI valuation dissolves once you locate the company precisely in the stack. There are four layers, and the economics of each are different enough that treating them as one category is the root error.

The practical test is blunt. If the company's product were rebuilt from scratch by a competent team with access to the same public models, how long would it take, and what would they still be missing at the end. If the answer is four months and nothing, you are looking at the bottom row wearing the language of the middle two.

Question two: does the data advantage compound, or merely exist?

Almost every AI target claims a data advantage. The claim is usually true and usually irrelevant, because having data is not the same as having data that compounds.

A compounding data asset has three properties. It grows through normal product use rather than through purchase or manual labelling. It improves the product in a way the customer can feel. And the improvement makes the customer use the product more, which generates more data. That is a loop, and loops are what create defensibility.

A static data asset is a corpus the company assembled once. It has value, it took work, and a competitor can assemble a comparable one with the same work. It is inventory, not a moat.

The diligence question is not how much data do you have. It is: what specifically got better in the product over the last twelve months because of data generated by customers, and can you show me the before and the after. Companies with a real loop answer this immediately and with pride. Companies without one answer with volume statistics.

Question three: what does gross margin do at ten times scale?

This is where AI-native businesses diverge most sharply from software, and where models built on software intuition break.

In classic software, marginal cost is near zero, so gross margin expands with scale. In an AI-native business, every unit of usage consumes inference, and inference is a real cost paid to a third party. Margin behaviour therefore depends on the relationship between three curves: falling inference prices, rising usage intensity per customer, and the company's ability to route cheaper models to easier tasks.

Three outcomes are possible and they are not equally likely.

The single most useful artefact to request in diligence on an AI-native target is cost of inference per account by cohort over time. It is rarely in the data room and the company always has it. If they will not produce it, that is itself the finding, and it belongs in the category of absences discussed in what machine review cannot see.

Question four: how much of the moat is workflow rather than model?

This question is worth more than the other five combined, and it is the one most often skipped because it sounds unglamorous.

The model is the part everyone discusses. The workflow is the part that makes the product hard to remove: the integrations into systems of record, the permission structures mapped to how the customer's organisation actually approves things, the audit trail a regulator will accept, the exported artefacts that other people in the customer's business now depend on, the habits of two hundred users.

Models get matched. Workflow entrenchment does not, because ripping it out requires the customer to re-plumb their operations. A company with a mediocre model and deep workflow entrenchment is a considerably better asset than a company with an excellent model sitting beside the workflow rather than inside it.

The test is a question put to three customers: if a competitor offered you the same output quality for forty percent less, what would it take to switch. If the answer is a procurement conversation, there is no moat. If the answer involves re-doing integrations, retraining staff, and re-certifying a process, there is one.

Question five: what happens when the foundation model improves?

Every AI-native company sits in one of two relationships with the underlying model layer. Improvements in the foundation model either make the company better or make the company unnecessary.

A company whose product is a thin layer over a general capability is in the second relationship. Each foundation model release absorbs some of what the company sold. This has already happened repeatedly to document summarisation, transcription, and basic code assistance businesses.

A company whose product is domain data plus workflow plus evaluation is in the first relationship. A better underlying model makes its output better at no engineering cost, and its own layers remain the differentiator.

Ask the founder directly: describe the last major foundation model release and what it did to your product. The answer separates the two relationships in about ninety seconds. Founders in the first relationship describe an upgrade. Founders in the second describe a scramble.

Question six: is the customer buying an outcome or running an experiment?

Revenue growth of ninety percent means something different depending on which budget it came from.

Innovation budget is discretionary, annual, and disappears when the sponsor changes role or the economy tightens. Operating budget is defended, because someone's function now depends on the thing.

The diagnostic questions are practical. Which cost centre funds this. Who signs the renewal, and is it the same person who signed the original. What breaks in the customer's operation if the product is switched off on Monday. If nothing breaks, the revenue is a pilot with an invoice attached, and pilot revenue does not survive a downturn or a sponsor change.

Cohort retention past the second renewal is the number that settles this. First renewals are frequently sentimental. Second renewals are budgeted.

What a good answer set looks like

A target worth underwriting at a premium owns the domain and workflow layers, has a data loop that visibly improved the product in the last year, can show inference cost per account falling or stable by cohort, has customers who would need to re-plumb operations to leave, gets better rather than threatened when foundation models improve, and is funded from operating budget with clean second renewals.

Most targets answer two of six well. That is not a reason to pass. It is a reason to price it, and to be specific in the value creation plan about which of the remaining four the hold period is meant to build. An AI company bought at a workflow-layer price and improved into a data-loop asset is a real thesis. An AI company bought at a data-loop price that turns out to be a wrapper is a write-down that took three years to become visible.

The discipline is the same one that applies to any technology claim inside a business: separate what the vendor says it does from what the customer cannot do without. That distinction is the whole of the analysis, and it is set out for portfolio technology spend in the digital EBITDA bridge.

VCI Institute in collaboration with Mohamad Chahine
Published 26 October 2026

About the VCI Institute

The VCI Institute is a nonprofit dedicated to building practical capability and shared standards for value creation in private equity. The Institute publishes operator-grade frameworks and runs certification programmes for operating partners, portfolio company executives, and value creation analysts. You can see what each programme actually covers before deciding. Analysis published here draws on the Institute's certification curricula and on structured review of mid-market transaction patterns rather than on any single proprietary dataset. Where a figure is directional rather than measured, it is described as such.

Further material is available in the Institute's Insights library and its free resource library of templates, checklists, and case snapshots.

© 2026 VCI Institute. All rights reserved. The frameworks, terminology, and analysis presented in this article are the intellectual property of the VCI Institute. Reproduction or derivative use without written permission is prohibited. Citation with proper attribution is welcomed.