AI diligence is the assessment of whether an AI-centric company's advantage is durable or borrowed. It examines what the company actually owns — data, distribution, workflow integration, switching costs — versus what it rents from a model provider that could reprice, deprecate, or absorb the product outright. For growth-stage targets, it runs alongside an assessment of whether the acquisition engine is a system or a set of relationships.
A demo proves the product works today. It proves almost nothing about whether the company will be worth more in three years, because in AI the gap between "works" and "defensible" has never been wider.
The pattern that burns investors is consistent: a well-built product, real customers, genuine usage, and an underlying capability that becomes a checkbox feature in a general-purpose model the following quarter. The diligence question isn't "is this good." It's "what happens to this when the thing it sits on top of gets ten times better and free."
What we actually examine
Owned versus rented capability
Which parts of the value proposition depend on a third-party model, and what happens commercially and technically if that model's price, terms, or availability change. Vendor concentration in AI is a balance-sheet risk that rarely appears on the balance sheet.
Data position
Is there proprietary data, and does it actually compound? A lot of claimed data moats are collections of records the company has no exclusive right to and no mechanism to keep growing faster than a competitor's.
Workflow depth and switching cost
Products embedded in a customer's daily workflow survive model commoditization. Products that sit beside the workflow get replaced by whatever ships natively. This is usually more predictive of durability than model quality.
Acquisition engine quality
Whether growth comes from a repeatable system or from a founder's relationships and a few unrepeatable channels. This is where growth-stage valuations most often break after close.
Cost structure under load
Inference cost per unit of value delivered, and what gross margin looks like at ten times current volume. Many AI businesses have healthy-looking margins that invert as usage scales.
Data handling and compliance posture
What customer data flows where, under what contractual terms, and whether that posture survives an enterprise security review. For targets selling into regulated buyers, this is frequently the binding constraint on the growth plan.
Questions we bring to every AI target
| Question | Why it matters | Weak answer looks like |
|---|---|---|
| What breaks if your model provider doubles prices? | Tests vendor concentration and margin resilience | "We'd absorb it" with no modeling |
| What could a frontier lab ship that makes you a feature? | Tests self-awareness about commoditization risk | "Nothing — our fine-tuning is unique" |
| What data do you have that a competitor cannot buy or scrape? | Separates real data moats from collected records | Volume claims with no exclusivity |
| Where does the product sit in the user's day? | Predicts switching cost and retention durability | Adjacent to workflow, opened occasionally |
| Show me acquisition by channel for 24 months | Distinguishes a system from a founder's network | One channel, or a chart that starts six months ago |
| What does gross margin look like at 10x volume? | Surfaces inference cost inversion | Current margin assumed constant |
We're skeptical of architectures hard-coded to a single model or vendor, and that skepticism shows up in our diligence. We think model-agnostic design is a genuine durability signal and vendor lock-in is a genuine risk — a view we hold in our own build practice, not just in assessment. Where a target has deliberately gone deep on one provider for defensible reasons, we'll say so; we're not looking to confirm a prior.
Frequently asked questions
How is AI diligence different from standard technical diligence?
Standard technical diligence asks whether the software is well-built, maintainable, secure, and staffed. That still matters. AI diligence adds a durability layer: how much of the value depends on capabilities the company doesn't own, what happens when those capabilities improve or commoditize, whether the data position compounds, and how unit economics behave as inference volume scales. A codebase can be excellent and the business still be structurally fragile.
What's the most common overvaluation pattern you see?
Mistaking product quality for moat. A team ships something genuinely good, users like it, retention looks fine — and the entire differentiated capability is a prompt layer plus UI on a general-purpose model. That's a real business, sometimes a good one, but it should be valued as an application and distribution business, not as an AI technology business. The multiple difference between those two framings is very large.
Do you do diligence on non-AI targets?
Yes, where the question is growth. We assess acquisition engines, unit economics, and channel durability for growth-stage and performance-marketing-dependent targets, which is a common profile in lower-middle-market private equity. See lead economics for how that applies to lead generation and marketplace businesses specifically.
How fast can you turn around a diligence assessment?
It depends on data-room access and management availability, but a focused assessment typically runs one to three weeks. We'd rather scope narrowly and go deep on the two or three questions that actually determine the decision than produce a broad report that restates the CIM. If the timeline is compressed, tell us what the decision hinges on and we'll aim there.
Can you help after the deal closes?
Often that's the more valuable engagement. Diligence identifies where the growth engine is weak; the work of actually fixing it is a different project with a longer arc. We do both, and we're explicit about which one you're buying.
Do you sign NDAs and work inside a data room?
Yes, as a matter of course. We work with confidential deal material routinely and operate under whatever confidentiality and conflict terms the engagement requires.