AI Diligence & Unit Economics
Whether the target has AI capability or a wrapper — and whether the margin survives the growth plan
Every target is an AI company now. Most of them are not.
The AI premium is worth several turns of EBITDA, which means it is worth claiming whether or not it is earned. In a large share of the companies now described as AI-powered, the capability is a prompt against a commercial model — replicable by any competitor in a week, with a corporate card. That can still be a good investment. But the premium should attach to the distribution and the workflow, not to the API call.
The second problem is quieter and shows up later. Inference cost is variable and scales with usage, while SaaS pricing is generally per-seat or flat. Growth therefore amplifies the margin problem rather than fixing it. A company can look entirely healthy at current volume and be structurally unprofitable at three times the volume — and none of that is visible in the current-period financials you are underwriting against.
This is the same work as cloud economics, applied to the fastest-growing line item on the bill. Quantify it, attach a confidence level, and translate it into EBITDA and multiple impact. Drawing on years as a Principal Solutions Architect at AWS in a Field CTO capacity, and on hundreds of engagements with PE-backed companies, I answer the question the deck does not: is this real, and what does it cost at scale?
Three Ways In
Two are diligence-stage and answer a pricing question. The third is ownership, for when the answer is that something has to be fixed.
AI Diligence
2-3 weeksWhether the target has AI capability or an API call any competitor can make — and what the premium should actually attach to.
Every target is an AI company now, and the premium is worth several turns of EBITDA. Most of it is a wrapper. This engagement separates the two before you underwrite the multiple: what is proprietary, what is switchable, what is defensible, and what a competitor could replicate in a quarter with a corporate card.
What you get
- Capability assessment — proprietary model, fine-tune, retrieval layer, or vendor passthrough
- Data rights and provenance review, including what survives a change of control
- Model and vendor dependency map with switching cost estimated
- Inference cost per unit of revenue, benchmarked
- A written view on whether the AI premium is supportable, and at what level
- Findings translated to EBITDA and multiple impact for the investment committee
AI Unit Economics Review
2 weeksWhat AI actually costs this business per customer, per transaction, and per dollar of revenue — and what happens to gross margin at scale.
Inference cost scales with usage while SaaS pricing does not, so growth amplifies the problem rather than fixing it. A company can look healthy at current volume and structurally unprofitable at three times the volume. This review models the gross margin trajectory and quantifies the optimization available, with confidence levels attached rather than assumed.
What you get
- AI cost as a share of revenue, gross margin, and EBITDA today
- Margin trajectory modeled against the growth plan, not current volume
- Model tiering, caching, and routing savings quantified by confidence level
- Pricing and packaging exposure where usage is unbounded
- A prioritized optimization plan with cost, timeline, and confidence per item
Fractional AI CTO
1-2 days/weekOngoing ownership of the AI roadmap, the build-vs-buy decisions, and the spend — accountable to the value creation plan.
The follow-on engagement when the diligence or the review concludes that something needs to be owned rather than advised on. Most portfolio companies do not need a full-time AI executive. They need someone with the judgment to decide what to build, what to buy, which models to depend on, and what the whole thing should cost — and the standing to defend those decisions to a board.
What you get
- AI roadmap ownership tied to the value creation plan
- Build-vs-buy and model selection decisions with cost modeling
- Data readiness — the precondition most AI plans skip
- Inference spend governance tied to unit economics
- Evaluation discipline, so AI quality claims are measured rather than asserted
- Board and sponsor reporting in financial terms
Five Tests That Separate Capability From a Wrapper
Applied in order. A company that passes all five is defensible. One that fails the first two is a distribution business with an API call in it.
Could a competitor build this with a corporate card?
If the capability is a prompt against a commercial model, the answer is yes, in about a week. That is not worthless — distribution and workflow still matter — but the premium should attach to the distribution, not to the AI.
Does the company own anything the model does not?
Proprietary data, labeled outcomes, a feedback loop that improves with usage, or a workflow the customer will not rip out. Absent one of those, there is no moat, only a head start.
What happens if the vendor doubles the price or drops the model?
Single-model dependency with no abstraction layer is a concentration risk that belongs in the risk register. Ask what switching would cost in engineering weeks and in output quality.
Is quality measured or asserted?
Companies with genuine capability have an evaluation harness, regression tests on outputs, and numbers they can show you. Companies with a wrapper have demos. The absence of evaluation is itself a finding.
Does the margin survive the growth plan?
Inference cost is variable and scales with usage. If pricing is per-seat or flat and usage is unbounded, growth degrades gross margin. Model it at plan volume, not at today's.
The reasoning behind each test, and what to ask management to evidence it, is in AI Whitewashing: Telling Capability From a Wrapper.
The two-week AI assessment
Nobody should commit to an ongoing engagement with someone they have not worked with. So most engagements start with a fixed-scope, fixed-fee assessment instead.
Two weeks. I review what has shipped, what it costs per unit of revenue, what the company owns versus rents, and whether the margin survives the plan — and hand you a written assessment with a prioritized plan. It is a complete piece of work on its own.
The fee is scoped on the discovery call and fixed before any work begins. No hourly billing, no scope creep, no invoice you did not see coming.
If the assessment concludes the AI premium is unsupportable, it will say so. If it concludes the AI is genuinely defensible and needs nothing from me, it will say that too. Either is a useful answer, and you will have paid a defined fee to get it.
2 weeks · fixed feeWhat it covers
- What has actually shipped versus what has been demoed
- Inference and training cost per unit of revenue, against benchmarks
- Data readiness, rights, and provenance — including what survives a change of control
- Model and vendor dependency, and what switching would cost
- Whether AI quality is measured, and whether the numbers hold up
- Whether the AI portion of the value creation plan is deliverable by this team
What you get
- A written assessment, in the language your investment committee uses
- Capability versus wrapper, stated plainly, with the evidence behind it
- Gross margin modeled at plan volume rather than current volume
- A prioritized optimization plan with cost, timeline, and confidence for each item
- A direct answer on whether the AI premium in the model is supportable
Who This Is For
PE Operating Partners
You are being asked to underwrite an AI premium on a target whose technical claims you cannot independently evaluate. You need someone who will tell you what is real, quantify what it costs at scale, and put it in terms the investment committee already uses.
Deal Teams Pre-LOI
The target's deck says AI on every page and the comparable transactions support a higher multiple. You need to know, before the LOI, whether the capability is proprietary or a vendor passthrough — and what the premium should attach to if it is the latter.
Portfolio Company CEOs
You have an AI roadmap you cannot personally evaluate and an inference bill that grows faster than revenue. You need a peer who can tell you which parts are load-bearing, which are theater, and what the margin looks like at three times current volume.
How an Engagement Works
Discovery call
Thirty minutes to understand the situation and the timeline, and to establish whether this is a diligence question, a margin question, or an ownership question. If it is none of them, I will say so.
Assessment
The two-week assessment described above. Most engagements start here, and it is a complete piece of work in its own right — plenty of firms take the assessment, act on it themselves, and need nothing further.
Engagement
Where ownership is the right answer, the agreed cadence begins against the plan. Standing time with the CEO and engineering leadership, a regular reporting rhythm with the sponsor, and clear ownership of the decisions in scope.
Handover
Every engagement is built to end. Documented architecture, model dependencies, evaluation harness, vendor relationships, and roadmap — handed to the permanent leader with onboarding support.
Common Questions
How do you tell a real AI company from a wrapper?
Five tests, applied in order. Could a competitor build the same capability with a corporate card and a week? Does the company own something the model does not — proprietary data, labeled outcomes, a feedback loop, or an entrenched workflow? What happens if the model vendor doubles the price or deprecates the model? Is output quality measured with an evaluation harness, or asserted with demos? And does gross margin survive the growth plan, given that inference cost is variable and scales with usage? A company that passes all five is defensible. A company that fails the first two is a distribution business with an API call in it, which may still be a good investment — but the premium should attach to the distribution.
What should AI cost as a share of revenue?
It depends on where the AI sits in the product, which is why the useful question is not the absolute number but the trajectory. Inference cost is variable and scales with usage, while most SaaS pricing is per-seat or flat, so growth widens the gap rather than closing it. The number to model is gross margin at plan volume rather than at current volume. A company can look entirely healthy today and be structurally unprofitable at three times the usage, and that is invisible in the current-period financials.
Is this different from ordinary technical due diligence?
It is a specialization within it, and it usually runs alongside. Conventional technical diligence covers architecture, the engineering organization, delivery record, security, and infrastructure spend. AI diligence adds the questions that determine whether the AI premium in the model is supportable: capability versus passthrough, data rights and provenance, model dependency and switching cost, evaluation rigor, and inference economics at scale. Both can be delivered together on a single engagement.
Do you build AI systems, or only assess them?
The core practice is assessment and ownership of the decisions — what to build, what to buy, which models to depend on, what it should cost, and how quality gets measured. On a fractional engagement that extends to owning the roadmap and holding the engineering organization to it. It does not extend to being the implementation team; where build capacity is the gap, that is a hiring or vendor decision, and I will say so rather than sell you the work.
Do I have to commit to a full engagement up front?
No. Most engagements begin with a fixed-scope, fixed-fee two-week assessment. It produces a written assessment and a prioritized plan, and it is a complete piece of work on its own — firms regularly take the assessment, execute against it internally, and never need anything further. There is no obligation to continue, and if the assessment concludes the AI premium is unsupportable or that no ongoing engagement is warranted, it will say so.
How quickly can this run inside a live deal?
Discovery calls happen within a week. Diligence work is scheduled around the deal timeline, and a focused AI assessment can typically start within a week to ten days — the constraint is usually data room access and management availability rather than my calendar.
How does this relate to your cloud economics work?
Inference is the fastest-growing line item in most portfolio companies' infrastructure bills, so AI economics is where cloud economics is going rather than a departure from it. The method is the same one in Cloud Economics for Private Equity: quantify the waste, attach confidence levels, and translate the finding into EBITDA and multiple impact. The cost structure is simply less well understood, which is exactly why the mispricing is larger.
Underwriting an AI premium?
Book a free 15-minute discovery call. We'll discuss the situation and I'll tell you honestly whether this is worth a formal assessment.
Schedule a Call