AI Startup Due Diligence
Most AI startup checks go wrong for one simple reason: people judge the demo, not the business. A slick prototype can hide weak data, thin margins, or risky customers. Good diligence asks what is truly hard to copy, what it costs to serve each user, and whether buyers keep coming back.
Start with the product, not the pitch
Ask what real job the product does. Then ask what happens if you remove the AI label. If the answer is “not much,” you may be looking at hype.
- What painful workflow does it fix?
- Who uses it every week?
- What result gets better: speed, quality, revenue, or cost?
- Can a buyer explain why they would pay for it?
A useful check is the Artificial Intelligence Risk Management Framework (AI RMF 1.0). It pushes you to look past the model and ask about reliability, safety, privacy, transparency, and who owns the risk when the system fails.
Find the real moat
Many AI startups look similar on the surface. The real question is whether the company owns something others do not.
- Weak moat: a nice interface on top of a public model.
- Stronger moat: unique workflow data, hard-to-build integrations, feedback loops, or deep customer process knowledge.
For example, a startup that helps insurers process claims may get stronger over time if it learns from private claim documents, adjuster feedback, and internal approval steps. A startup that only wraps a public model with a prompt library is much easier to copy.
This matters because the market moves fast. AI Index 2025: State of AI in 10 Charts from Stanford HAI shows that model capability is improving and usage costs have fallen sharply. That helps builders, but it also means thin wrappers can lose their edge fast.
Dive Deeper with BonsAI Chat
Check the cost story
Inference means running the model for each user request. Training means building or tuning a model ahead of time. A startup can grow revenue and still have a bad business if each customer action burns too much compute.
- What is the model cost per task?
- Does gross margin stay healthy after cloud, storage, and human review?
- Do costs spike at peak usage times?
- Can the product switch between expensive and cheaper models when needed?
The goal is not perfect precision. It is to see whether the founders understand the unit economics. If they cannot explain cost per workflow, they probably do not control the business yet. Stanford HAI’s AI Index 2025: State of AI in 10 Charts is a good reminder that lower model prices help, but they also make competition tougher.
Traction that matters
Do not stop at a logo slide. Ask whether customers are active, expanding, and staying.
- How quickly did the company reach first revenue?
- How much revenue comes from the top 3 customers?
- Are users active every week, or just during pilots?
- Are contracts renewing?
- Is usage growing inside existing accounts?
Stripe Atlas startups in 2025: Year in review shows that startups are getting to revenue faster than before. That is useful context, but fast early revenue is not enough. Pair it with basic operating metrics from SaaS metrics: A complete guide to tracking business growth, especially churn, net revenue retention, customer acquisition cost, and customer concentration.
Useful research tools
Questions that cut through hype
- If a bigger company copied this feature in 6 months, what would still be hard to copy?
- What private data makes the product better over time?
- How much human work is still hidden behind the AI?
- What happens if model prices rise, a key API changes, or a regulator asks for more transparency?
- Does the team treat risk as a real product issue, not just a policy slide?
If you want one simple rule, use this: back startups that own a painful workflow, collect useful feedback data, understand their cost base, and can show real customer pull. Be wary of startups that only own the prompt.