Founder guides
Creating a Business Model for an AI Startup: Who Pays
A business model is the answer to three questions: who pays, what it costs you to serve them, and whether the first covers the second. AI products change the second answer more than most founders expect.
Creating a business model means deciding who pays you, for what, on what unit, and what it costs you to deliver it. For an AI startup, pay particular attention to costs that grow with usage, such as model inference and human review, because they can make a product that sells well lose money on every customer.
Steve Blank made the point years before AI products existed: "a product is just a part of a startup, but understanding customers, channel, pricing, etc. are what make it a business" (Steve Blank).
Key takeaways
- Start with the buyer and the budget, not the product.
- Use the nine-block canvas to see every part of the model on one page.
- Model your cost to serve per customer, including inference and human review.
- Match the pricing unit to your costs where you can, so heavy users do not lose you money.
- Treat every block as an assumption until a buyer confirms it.
What is a business model, in plain terms?
It is how a company creates value for a customer, delivers it and gets paid more than it costs. If any of the three breaks, the product may still be useful but the business does not work.
Strategyzer, which publishes the Business Model Canvas, describes it as a way to "describe how your business creates, delivers, and captures value on one page" (Strategyzer).
For a first-time founder, the useful version is three questions: who pays, what do they pay for, and what does it cost you to deliver it to them. Everything else is detail on those three.
How do you use the Business Model Canvas for an AI startup?
Fill in all nine blocks, mark each as evidence or assumption, and test the riskiest assumptions first. For AI products, the cost structure block deserves more attention than usual.
The canvas's nine building blocks were first proposed by Alexander Osterwalder in 2005 (Wikipedia):
| Block | AI-specific question |
|---|---|
| Customer segments | Who has the problem and a budget, not just the problem? |
| Value propositions | What does the buyer stop doing, and what does that save? |
| Channels | How will you reach the first 50 buyers? |
| Customer relationships | Self-serve, assisted, or done with the customer? |
| Revenue streams | Subscription, usage, per outcome, or a mix? |
| Key resources | Data, models, domain expertise, integrations |
| Key activities | Building, evaluating output quality, supporting customers |
| Key partnerships | Model providers, data sources, systems you integrate with |
| Cost structure | What does each customer cost you each month, including inference and review? |
Why do AI startups have different cost structures?
Because serving each customer costs money in a way that traditional software mostly does not. Computing costs grow with usage, and many AI products need people to check or correct output.
Martin Casado and Matt Bornstein of Andreessen Horowitz found that AI companies had "gross margins often in the 50-60% range – well below the 60-80%+ benchmark for comparable SaaS businesses" (a16z, 2020). They pointed to two causes: cloud infrastructure for training and running models, and ongoing human involvement, from labeling data to reviewing output.
Model prices have changed a great deal since 2020, so treat the percentages as a historical benchmark rather than a current one. The structural point still holds: in an AI product, your cost of delivering the service can rise with every customer and every task, and your model has to account for it.
How should an AI startup choose its revenue model?
Choose the unit that tracks the value the buyer gets, and check that it also tracks your costs. When the two diverge, your heaviest users become your least profitable ones.
| Model | Works when | Watch out for |
|---|---|---|
| Per seat | Value grows with the number of people using it | Heavy users on a flat seat price |
| Per usage (task, document, call) | Costs and value both grow with volume | Unpredictable bills that buyers resist |
| Per outcome | You can measure a result the buyer values | Disputes about what counts as the outcome |
| Flat platform fee plus usage | Buyers want predictability and you need cost cover | Complexity in the first sales conversations |
Test the model in the first sales conversations. Say a price and a unit, and note which part the buyer questions.
How do you check whether an AI business model works?
Estimate the monthly cost to serve one typical customer, compare it with the monthly price, and see what is left to pay for acquiring and supporting them.
A simple per-customer check, with your own numbers:
- Price per customer per month.
- Minus cost to serve: model and computing costs for a typical month's usage, plus any human review time.
- Equals gross profit per customer.
- Compare with the cost to acquire a customer. If it takes many months of gross profit to recover what you spent winning the customer, and they may leave before then, the model needs work.
Run the check for a light user and a heavy user, not only an average one.
How can NELL help you test a business model?
DeepValidate scores the monetization side of an idea, including pricing, unit economics and the revenue model, using sourced competitor pricing, so you can see which assumptions look weakest before you commit.
DeepValidate's "Can we monetize?" pillar covers pricing and willingness to pay, unit economics and the revenue model, as three of its 11 scored dimensions. Start with a free Quick Validate for a first read, or see a full example in the sample report.
Frequently asked questions
What are the main parts of a business model?
Who you serve, what value you offer them, how you reach and serve them, how you earn money, and what it costs you. The Business Model Canvas breaks these into nine blocks.
Why do AI startups have lower gross margins than SaaS?
Serving each customer uses computing for running models, and many AI products need human review of output. Both costs grow with usage, unlike most traditional software costs.
What is the best revenue model for an AI startup?
There is no single best one. Choose a pricing unit that tracks the value the buyer gets and, where possible, your own costs, then test it with buyers before you commit.
Should a business model be written before building a product?
A rough one, yes. Knowing who pays and what it costs you to serve them changes what you build, especially for AI products with usage-based costs.
How often should a startup change its business model?
Change a block when evidence from buyers consistently contradicts it. Early on that can happen often; the canvas is meant to be updated as you learn.
Start where you are
Every block on the canvas is an assumption until a buyer confirms it.
Sources
- Martin Casado and Matt Bornstein, The New Business of AI (and How It's Different From Traditional Software), Andreessen Horowitz (2020)
- Strategyzer, The Business Model Canvas
- Wikipedia, Business Model Canvas
- Steve Blank, The Lean LaunchPad: Teaching Entrepreneurship as a Management Science (2010)
Gross margin figures are from a 2020 analysis and are a historical benchmark. Model and computing prices change quickly; use your own current costs.