Almost every AI company we meet says its product gets better with use. Far fewer can show it. The phrase has become so common that it barely means anything until you ask for evidence.
Here is how we test the claim at Series A and B, and what separates a real compounding loop from a slide.
Show the curve, not the story
The first thing we ask for is a chart: a quality metric over time, split by customer cohort. Automation rate, accuracy, first-pass resolution or time to complete, whichever fits the product. If the product truly learns, older customers should see better results than newer ones on comparable work, and each new cohort should start higher than the last.
A flywheel you can’t chart is a hope, not a moat.
Find where the learning lives
Improvement has to come from somewhere specific. We ask founders to point to it. Common sources are:
- Corrections. Human reviewers fix outputs, and those fixes are captured in a structured way and fed back into the system.
- Outcomes. The product sees what happened downstream: the claim was paid, the appeal succeeded, the alert was a true positive.
- Customer context. Each deployment builds a richer picture of the customer’s rules, exceptions and history.
- Cross-customer patterns. Learning from one customer improves results for others, within clear data boundaries.
If the answer is “we fine-tune on all the data,” we dig deeper. If the answer is “the next foundation model will make us better,” that improves every competitor too.
Separate your gains from the model’s gains
Foundation models improve every few months, which lifts every product built on them. To isolate a company’s own learning, we ask what happens when the same base model is run without the company’s accumulated corrections, context and evaluations. The gap between those two results is the moat. It should be large, and it should be getting larger.
Check that it shows up in the numbers
Real learning loops show up in the business: rising gross margins as less human review is needed, faster onboarding for new customers in the same vertical, and expansion as customers trust the product with more of the workflow. When the quality curve and the financial curves move together, we know we’re looking at something durable.