How to tell whether an AI startup's moat is real
Four kinds of defensibility an AI company can claim, which ones survive a competitor with the same model, and the test that separates them in one question.
Most AI startups describe their moat as proprietary data, a fine-tuned model or domain expertise. Only one of those is usually defensible, and it is rarely the one named in the deck. The useful test is to ask what a well-funded competitor with the same model access would have to do to reach parity, and how long it would take. If the honest answer is measured in weeks, the moat is distribution or workflow depth rather than technology, which is fine as long as everyone is clear that is what is being backed.
Every AI deck has a defensibility slide and most of them say the same four things. Proprietary data. A fine-tuned model. Domain expertise. Network effects coming later.
The problem is not that these are lies. It is that three of the four describe an advantage that a competitor with the same model access and a year of funding can close, and the deck rarely says which one is doing the work.
The test, in one question
What would a well-funded competitor with the same model access have to do to reach parity, and how long would it take?
Ask the founder. Then ask your own technical reader. The gap between those two answers is usually the finding, and it is a more useful finding than anything on the slide.
If the honest answer is weeks, the technology is not the moat. That is not a problem on its own. Plenty of durable companies have no technical moat at all. It becomes a problem when the valuation was set as though they did.
The four claims, and what each is worth
Proprietary data
Sometimes the real thing. Usually not.
The question is whether the data is proprietary or merely private. Private means nobody else happens to have it. Proprietary means nobody else can get it, because it is generated by something only this company does.
A company that accumulated ten thousand labelled examples by paying contractors has private data, and a competitor can pay the same contractors. A company whose data is produced as a byproduct of customers using the product has proprietary data, and the gap widens on its own. Ask which one it is and ask how the volume has grown over the last two quarters.
A fine-tuned model
Worth less every year, and the founders usually know it.
Fine-tuning used to be where the work was. Base models have improved fast enough that a fine-tune built against a model from eighteen months ago is frequently beaten by careful prompting of a current one, and the company that built it has to decide whether to redo the work or abandon it.
Ask when the fine-tune was done, against what base, and what happened the last time the provider shipped a new model. The answer tells you whether this is an asset or a maintenance liability.
Domain expertise
Real, and almost never defensible by itself.
A founder who spent eleven years in claims processing knows things a competitor does not. That knowledge is in their head, and heads can be hired. What makes domain expertise durable is when it has been turned into something structural: an evaluation set that encodes what good output looks like in that domain, a taxonomy, a set of rules nobody could derive without the experience.
Ask what has been written down. If the expertise exists only in the founder's judgement, you are backing the founder, which is a reasonable thing to back as long as the terms reflect it.
Workflow depth
The one that usually turns out to be the actual moat, and it is rarely on the slide.
A product wired into four of a customer's systems, holding six months of their history, with three people's daily work built around it, is hard to displace regardless of what model is underneath. The switching cost is the moat. A competitor with a better model still has to persuade somebody to migrate.
This is unglamorous and it is the thing to look for. Ask how many systems the product touches at the median customer, and what a customer would have to do to leave.
What this means for the number
A company whose defensibility is workflow depth and distribution is a good company with a specific risk profile. It must get deep quickly, because the window before a competitor arrives is measured in quarters rather than years.
A company that believes its moat is a fine-tune is carrying a risk nobody has priced, and the first model release after you invest will surface it.
Neither is a reason to pass. Both are reasons to write the real thesis into the memo, so that the next round is raised on something true.
One limit worth stating
This is a framework for a conversation rather than a scoring system, and frameworks like it have a failure mode worth naming. Applied strictly, it talks you out of good companies whose moat has not formed yet, which at seed is most of them. At that stage the question is not whether a moat exists but whether the team understands which one they are building, and whether the plan would produce it. A founder who says "we have no moat yet, here is the one we are going for" is giving you a better answer than one who claims four.
Where this sits in the work
The moat question is judgement and sits on top of the things that can be checked directly. Those are in the technical due diligence checklist for an AI startup, and what it costs to have somebody else do the checking is in what technical due diligence on an AI startup costs.
Our first memo for a fund is free, one per fund, inside 48 hours. It states its own gaps on the first page. Book the call.
If you are evaluating a company and want a technical read before the next partner meeting, that is the conversation to have.
Request the memoWritten by
Radwan Altaf
Radwan runs AISynq. Before that he delivered software inside enterprise programmes at DHL, AT&T, DirecTV and Accenture, which is where the habit of measuring a result against its baseline came from. More about the firm.
Read next
- The AI due diligence question bank: what to put to the founderThirty questions for the technical founder, grouped by what each one is really testing, and the follow-up that catches a rehearsed answer.
- What technical due diligence on an AI startup costsWhat firms charge for a technical read on an AI company, what moves the number, and the three things a cheap memo leaves out.
- Technical due diligence checklist for an AI startupThe twelve questions worth asking before you back an AI company, what a good answer sounds like, and the two that most founders cannot answer at all.
Work with us
Looking at a company right now and want a technical read before the partner meeting?
Request the free memoGet the next one
New writing in Technical due diligence as it goes up, roughly twice a month. One article per email and nothing else in it.