Sector and stage playbooks
How to build an AI startup pitch deck
Short answer
An AI startup deck must show a specific buyer problem, what is proprietary (data, workflow or distribution, rarely the model alone), what it costs to serve each customer, and evidence that it works on real customer data. Investors assume models are widely available, so the defensibility question arrives early.
What investors look for
- A problem a buyer pays to solve, not a capability looking for a use
- Something proprietary that grows with use
- Margins that survive the cost of inference
- Evidence in production
- A team that can build and sell
Slides that are specific to AI companies
| Slide | Show |
|---|---|
| Result for the buyer | Time or cost saved, measured at a customer |
| What is proprietary | Data you collect, workflow you own, distribution you have |
| Evidence | Accuracy or quality on a named task against a baseline |
| Cost to serve | Inference and infrastructure cost per customer; gross margin |
| Risk | Dependence on providers, data rights, regulation |
Answer defensibility directly
| Claim | Stronger version |
|---|---|
| "Proprietary AI" | "Trained on 2 million labelled invoices from our customers, which improves extraction accuracy on Indian formats" |
| "We use the latest models" | "We can switch providers in a day; our advantage is the workflow and the review data" |
| "First mover" | "Integrated into the accounting systems of 300 firms" |
The examples are illustrative.
Show the economics
Usage-based costs can erase margin. Show revenue per customer, cost of inference and gross margin, and how each moves as usage grows. If margins are low today, show the path: caching, smaller models, pricing.
Evidence over demos
A polished demo proves little. A customer using it daily on their own data, with a measured result, proves a lot. Lead with that.
Data and trust
State what data you use, who owns it, and how it is protected. Enterprise buyers ask before they sign, and investors ask because buyers do.
Mistakes
- Leading with technology
- A market defined as "all knowledge work"
- No cost-to-serve figure
- Benchmarks with no baseline
- Pilots with no paid conversion
Rehearse the moat question in the Q&A simulator.
Quick answers
Should I explain the model architecture?
Briefly, in an appendix. The main deck is about the buyer and the result.
Is using a third-party model a weakness?
Not by itself. Show what you add on top and why switching models would not hurt you.
How do I show accuracy?
On a named task, against a baseline, on customer data, with the sample size.