Pricing and running AI products
How do you price an AI SaaS?
Cal Renshaw, , 2 minute read
Price an AI SaaS from the bill upwards. Work out what one unit of work costs you on real inputs, multiply it by the month your heaviest sensible customer would have, and set a price where that comes to less than a tenth of what they pay.
That's the whole method. The rest of this post is why each part matters, because each part is where I've seen products lose money, including mine.
Start with the unit, measured on real inputs
A unit is whatever one press of the button causes: one quote drafted, one document read, one reply written. Its cost is everything the model is paid for to produce it, including the check and the retries.
Measure it on 30 real inputs from the people you're building for. Not the example you demo with. The demo is short and tidy, and real inputs are neither. (Nobody demos the messy one.) Then take a figure near the expensive end, not the middle, because the long, ugly inputs are usually the jobs that matter most to the customer.
Price for the heaviest sensible customer
The average customer is a fiction. In every product I've run with a meter inside it, a handful of customers do most of the using, and on a flat price they're the ones who cost you the most.
So ask the people you're building for how much of the thing they do in their busiest week. Multiply by 4. That's the month you're pricing for. I learned this on a product that drafted replies for people who let holiday cottages: I priced it on the average, and by the end of the second month the 3 keenest customers each cost me more in model fees than they paid.
Keep model fees under a tenth
A tenth sounds cautious. It's meant to. The sum leaves things out: support time when the model is wrong, the afternoon a provider changes its prices, the customer whose scanner sends the same folder every night. A tenth leaves room for all of that and still pays you.
It also leaves room for success. Model fees grow with use and your price doesn't, so the month a product goes well is the month the ratio gets worse.
Choose the shape of the price
Once the sum works, choose how to charge. There are 3 shapes:
- Flat. One price a month. Simple, and it makes your keenest customers your least profitable.
- Metered. A price per unit. Fair, but a small business can't budget for a bill that changes every month.
- Allowance. A monthly price that includes a fixed amount of use, and a known charge above it.
For a small product sold to small businesses I almost always end up at the allowance. It gives the customer a bill they can predict and gives me a ceiling on what any one customer can cost before they start paying their way.
If the sum doesn't work
If the heaviest sensible customer comes out above a tenth, don't write any code yet. There are 4 ways out, in this order: move the arithmetic and the comparisons into ordinary code, test a cheaper model on your own past work, change the shape of the price, and only then charge more.
The cost sheet is the one-page version of all of this. Who Pays for the AI? is the long one.
Questions
- How much should an AI SaaS charge per month?
- Enough that your heaviest sensible customer's model fees come to under a tenth of the price. Work the number out from a unit cost measured on real inputs, not from what competitors charge.
- Should AI products use usage-based pricing?
- Usually not on its own, for small customers. An allowance, a monthly price with a fixed amount included and a known charge above it, gives them a predictable bill and caps what a keen customer can cost you.
- What is a unit cost for an AI feature?
- What the model is paid for to produce one unit of work, such as one document read or one reply drafted, including the check and any retries. Measure it on 30 real inputs and use a figure near the expensive end.