In a period when the model race is accelerating, what makes an AI startup defensible? The evaluation list of our investment team.
Interest in AI startups has grown rapidly over the last two years, but the question we ask most as investors has not changed: “Will this company still be standing after the next release from the large model providers?” Below we share the seven criteria we use when evaluating AI startups.
1. Is the problem real, or does the model already solve it?
Startups solving a problem that a user could solve in a few minutes in a model interface (for example chat) can easily be covered by the model providers. We prioritize problems embedded in enterprise data, processes or regulation.
2. Proprietary data and feedback loop
The proprietary data a startup collects from users and how that data improves the product are decisive for defensibility. We look for a clear answer to “Does each user make the product a bit better?”
3. Depth of integration
A product deeply integrated into a customer's ERP, CRM or core banking systems is hard to replace. What matters is not the number of integrations but how embedded the product is in the workflow.
4. Unit economics
Model call costs directly affect gross margin per customer. At an early stage we want to see gross margin above 60 percent, a path to 75 percent at scale and a cost-reduction plan (smaller models, caching, distillation).
5. Security, privacy and regulation
KVKK, sector regulations and customers' data residency requirements are part of the enterprise sales cycle. It should be clear who keeps the data, where the model runs and how the output is audited. Solutions that can run on-premise offer a clear advantage in Türkiye.
6. Team: balance of research and product
Alongside a strong researcher, we look for a co-founder who can turn the product into a customer experience. We are more cautious about teams made up only of research, or those that only build an “interface on top of APIs.”
7. Proven customer usage
Impressive demos are no longer enough. The rate at which pilots convert to paid contracts, weekly active usage and depth of usage per customer are at the center of the valuation conversation.
An example from our portfolio
Our portfolio company Anlam Lab was chosen because it meets most of the criteria above: domain-adapted Turkish models for banking and the public sector, on-premise capability, 19 enterprise customers and gross margin above 70 percent.
If you have a startup that fits these criteria, you can fill in the application form or first take our readiness test.
When do we say “no”?
We usually do not proceed when the product's only feature could be added to a model provider's interface; when unit economics deteriorate as scale grows; when legal uncertainty about the use of customer data cannot be resolved; or when founders focus only on demos and press coverage rather than revenue. We give clear feedback on these points to every startup we decline.