Evaluating an AI moat when the models are everyone’s

The starting point for any honest discussion of AI moat and defensibility: the models belong to everyone. Your target calls the same APIs its competitors can call, at the same prices, and the providers improve that layer for the whole market at once. So when we evaluate an AI moat in due diligence, the first move is subtraction. Strip out everything the model provides, because that part is rented.

What remains after the subtraction is the moat. It is usually smaller than the deck claims, and rarely zero. The job is to size it.

What does not survive model commoditization

Benchmark wins. A benchmark lead is a snapshot of a moving race, and the next base model release reshuffles it. If the pitch leans on a leaderboard, ask what happened to the product’s edge across the last two major model releases. Teams with a real edge can answer; teams whose edge was the model cannot.

Prompt libraries and pipeline cleverness. Real work, worth doing, and replicable by a competent team in weeks. The same goes for being three months early: a head start is only a moat if it compounds into something else, data, contracts or distribution. Otherwise it is a countdown.

"Our fine-tuned model." Ask what it was fine-tuned on, whether that data source is durable, and whether the advantage survived the newest base models. Plenty of fine-tunes we meet get beaten by a newer base model with a paragraph of prompting. Fine-tuning is a technique, not a moat, unless the data behind it is one.

None of this makes a company worthless. A thin technical moat with strong distribution can be a good business; it should simply be priced as one, which is the same honesty we apply to the wrapper question.

Data feedback loops, the real kind

Most claimed data moats are storage, not loops. Data counts as a moat when four things are true. It is proprietary: not scrapeable, not purchasable by a competitor. It feeds the product through a working mechanism: eval sets, fine-tuning, retrieval corpora, learned rules. The improvement is measurable. And more usage strengthens it, so the lead widens instead of merely existing.

The test is to ask for the loop, not the warehouse. Show the last three product improvements that came from usage data, and the metric each one moved. A team with a real loop answers in specifics. A team with a data story answers in terabytes.

Corrections are the highest-grade ore: users fixing the system’s output, with the fix captured, labeled and fed back. Products that sit inside a workflow collect corrections naturally; products that sit beside one have to beg for feedback.

Workflow depth and switching costs

The durable moats we find in AI deals mostly look like classic software moats wearing new clothes. The product is wired into the customer’s systems through integrations that took months of joint work. It holds state the customer cannot cheaply export: history, configuration, accumulated corrections the system has learned from. It owns a workflow rather than decorating one.

The question we put to the target: what exactly does a churning customer lose, in weeks of effort and in accumulated context. If the answer amounts to "they would have to re-enter their prompts", the switching cost is an afternoon.

Distribution deserves its own line, because it decides more AI markets than technology does. An incumbent with the customer’s ear and a mediocre model beats a startup with a better model and no ear. If the target is a startup, ask what it has that incumbents cannot copy faster than the startup can build distribution.

How we test AI moat claims in a deal window

The questions that do the work:

• Which part of the product survives if the top model providers ship this capability natively next quarter
• Show three improvements driven by usage data, with the metric each moved
• What does a switching customer lose, concretely, and how long would onboarding a replacement take
• What happened to your quality edge across the last two major model releases
• Which customers came through channels a competitor cannot access

Two adjacent checks belong in the same session. A moat that only exists at unsustainable unit economics is not a moat, so we pair this analysis with the margin question. And dependence on a single model provider can quietly cap what the moat is worth, which is why we also look at model provider dependency.

The output of a moat evaluation should be one sentence a partner can defend in committee: what specifically survives commoditization, and why. When the AI thesis carries the valuation, we build that sentence as part of a full technical and AI due diligence, with the evidence attached.