In this article
An AI startup can create real value using another company’s model. Plenty of useful businesses depend on infrastructure they did not invent. The due-diligence question is what the company contributes, controls and improves when access to a capable model is available to competitors too.
My focus here is operating and product diligence. This is not a valuation formula or an investment recommendation. The questions are useful to founders preparing for scrutiny, product leaders making a strategy and investors trying to understand the business behind the presentation.
Ask what customers would miss tomorrow
Start with the customer’s work. Which recurring task becomes meaningfully easier? What does the customer currently do instead? Which user depends on the result, and which budget owner agrees it is worth paying for? A polished answer from an unfamiliar interface is an experience. A repeatable improvement in a necessary workflow can become a business.
Ask for evidence that distinguishes curiosity from dependence. A paid pilot is useful information, but so are the reasons it did not expand. A champion who loves the product may still be unable to get it approved. A good founder should be able to explain both the enthusiasm and the friction without treating every hesitation as customer ignorance.
Separate data possession from data advantage
“We have proprietary data” needs a follow-up. What is the origin, what use is authorised and how does it improve the customer outcome? A pile of scraped material, an exclusive commercial agreement and a well-labelled record of customer-approved corrections are different assets.
Hugging Face’s dataset-card guidance offers a useful documentation starting point: describe what the data contains and the context for its use. For diligence, I would go further and ask the company to show the relevant rights, quality checks and update process. That proposed evidence request is my own operating judgement, not a claim that a card certifies ownership.
Read the guide to official data sources if the team is still using “public” as its entire explanation. The commercial question is whether the advantage can be maintained and explained, not just whether a large file exists.
Stress-test the dependency on the model supplier
Ask the team to name its model, endpoint and critical features. What happens if the price, supported version or service availability changes? OpenAI’s official deprecation notices provide a concrete reminder that product dependencies have lifecycles. A startup does not need to predict every change; it needs a credible way to discover and respond to one.
A useful exercise is a supplier-change tabletop. Give the team a hypothetical deadline to replace its current model. Ask which behaviours would need retesting, which customer promises might be affected and who would own the work. A provider-neutral diagram is less convincing than a small working comparison and a list of known differences.
Follow one completed task through the economics
Request the cost of a useful outcome. Include model calls, retrieval, retries, infrastructure, human checking and support. Keep one-time onboarding effort separate from recurring delivery effort. Otherwise a heavily assisted pilot can appear to have the economics of a self-service product.
For illustration only, suppose an accepted task uses two initial calls, one retry and a human review. The cost of the first call cannot stand in for the cost of the accepted task. This is an accounting structure for investigation, not a claimed benchmark or a forecast. Ask the team to populate it from its own records.
Then test the busy day and the awkward customer. Which costs grow with usage? Which grow with customer complexity? Can pricing support the service the buyer actually expects? My article on AI integration costs develops that conversation from the buyer’s side.
Look for an advantage that survives a better base model
Imagine the underlying model improves substantially. Does that make the startup’s product more valuable, easier to copy, or both? A company built around correcting one temporary model weakness needs a different plan from a company embedded in a difficult, trusted workflow.
Possible strengths to investigate include access to a specific customer group, a hard-won integration, a repeatable implementation process, authorised feedback and a reliable operating relationship. None is automatically a moat. Ask how each affects acquisition, retention, quality or the cost of switching in this particular business.
Use a compact evidence pack
- A real workflow map with the customer and decision owner identified.
- A cohort view separating pilots, paying use and renewals.
- A cost-per-outcome calculation with human effort visible.
- A dependency register and a tested replacement path.
- A data provenance record and unresolved rights questions.
- An evaluation report that includes failures and limitations.
This is my suggested discussion pack, not a universal investment standard. Its value is that different functions can interrogate the same evidence. Product, engineering, commercial and security teams should not each be describing a different company.
I can help a founder or leadership team turn the broad claim “we have an AI advantage” into a clearer product and operating story. The useful work is often asking the awkward question early, documenting what is true and deciding what evidence to build next. The cardboard castle can stay on the desk. It just should not be the whole strategy.
Questions people ask
Can an AI startup have a moat while using a third-party API?
Yes, potentially. Investigate the customer workflow, distribution, authorised data, integrations and operating capability that the company contributes. API use alone neither proves nor disproves defensibility.
What should AI startup product due diligence include?
Examine customer need, repeat use and renewals, data rights, cost per accepted outcome, evaluation evidence, supplier dependencies and ownership of operational failures.
Is a proprietary dataset automatically a competitive advantage?
No. Its rights, quality, relevance, update process and demonstrated contribution to a customer outcome all need examination.
Something here sounds like your workplace?
Product strategy, hands-on AI delivery and experience across startups and enterprise teams.