13 posts
Pricing and running AI products
What a product with a model inside it costs, how to price it, and how to keep the bill under the price.
- Flat, usage-based or allowance pricing for an AI product?
An allowance. Set it at the heaviest sensible month, with a charge above it that pays. Customers can predict their bill, and nobody can cost you a fortune.
- How do retries affect AI costs?
A failure retried 3 times costs 4 times. Cap the attempts on every item, never retry a bad input, retry a wrong answer once at most, and read the pile every week.
- How do you cut LLM costs without making the product worse?
Move every job you can write as a rule into ordinary code first. 2 of my document product's 5 model calls were doing arithmetic and date checks. Moving them roughly halved the model spend.
- How do you price an AI SaaS?
Measure what one unit of work costs on real inputs, multiply by your heaviest sensible customer's month, and keep that under a tenth of the price.
- How do you raise prices without losing existing customers?
Raise it for new customers first, and tell existing ones in a proper letter with the reason. Never shrink the product quietly instead.
- How do you stop a surprise AI bill?
A hard cap on every key, watched until it fires, and a daily ceiling in your own code that stops, holds the work and tells you. Set the ceiling from the shape of the month.
- How do you work out what an AI feature costs per customer?
Run 30 real inputs through the model before you build, then read a page every week listing what each customer cost you. One row many times the others is either keen or broken.
- Should an AI product have a free tier?
No. A free tier on anything with a meter reads documents for people who haven't paid, every month, with no price to divide by. Show saved examples, run a sample by hand, or promise the money back.
- Should you call the model live or queue the work?
Queue it. Put the work in a queue, check it, store it, then tell the customer. Call the model live only when somebody is genuinely waiting.
- What do you do when your AI model is retired?
Rerun your own past work on the new model before any customer meets it. One door in the code, 3 months of stored inputs and a second option you know make it a Sunday job.
- What does a wrong AI answer really cost?
The model fee is a fraction of a penny. The real cost is your evening, an apology, and a customer's trust that the product says true things. Right, then cheap, then fast.
- What should you never hand to an AI model?
The judgement. Give the model labour you can check, and keep the price, who it's for, what you say when it goes wrong, and what to switch off.
- When is a cheaper AI model good enough?
When it agrees with the strong model on your own past work and its failures are the kind your check catches. Test it on 50 real inputs, read every disagreement, then route.
Questions
- What does "Pricing and running AI products" cover?
- What a product with a model inside it costs, how to price it, and how to keep the bill under the price.
- Which book goes deeper?
- Who Pays for the AI?: The bill: pricing and running a product with a model inside it, so every customer costs less than they pay.
- How long are the posts?
- A few minutes each. Every one answers its question in the first two sentences, then gives the decision and what it costs.