Founders arrive with a build budget. They have thought about design, development and a launch. What almost nobody has modelled is what the product costs to run once people are using it, and with AI features that number is not a rounding error the way hosting usually is.
Traditional software has a cost curve that flattens. You build it, you host it, and the tenth thousand user costs you barely more than the first thousand. AI inverts that. Every single use of an AI feature costs real money, and that cost scales linearly with usage. If you have not modelled it, your most successful feature becomes your biggest expense.
We run AI features in our own products and build them for clients, so we have had to get specific about this. Here is how we think about it.
Model the cost per action before you write the spec
The useful unit is not tokens or API calls. It is the action a user takes: one document summarised, one enquiry drafted, one record classified.
Work out what a single action costs by estimating input size, output size, and how many model calls sit behind it. That last part catches people out. A feature that looks like one call is often three, because you are retrieving context, generating a draft, then checking the output. Multiply the per action cost by realistic monthly usage per user, and compare it to what you charge that user.
We do this during scoping, before anything is built. It has changed feature designs more than once, and occasionally it has told a client that the version they imagined does not work at the price they want to charge. Better to learn that in week two than month eight.
Not every task needs your best model
The most common source of unnecessary cost is running everything through the largest available model because it worked in the prototype.
Plenty of steps in an AI workflow are simple: classifying an enquiry, extracting a date, deciding whether text needs review. A smaller, cheaper, faster model handles those well, and you save the expensive model for the step where quality is actually visible to the user. On a multi step workflow this routinely cuts running cost substantially without any noticeable drop in output quality.
The way to find out is to test it rather than assume. Run the same evaluation set through both and look at where the smaller model genuinely fails.
Cache the work you are repeating
A surprising share of AI spend goes on generating the same thing twice. Two users asking the same question of the same document. A summary regenerated every time a page loads. A system prompt of several thousand tokens sent with every request.
Caching results at the application level, and using prompt caching where the provider supports it, is unglamorous work with an immediate return. It also makes the product feel faster, which users notice more than they notice your margin.
Design the feature so cost tracks value
The cleanest position is one where the expensive action only happens when the user gets something worth paying for. That usually means putting AI behind an explicit action rather than running it automatically on everything.
A feature that summarises every record on import costs you money for records nobody ever opens. The same feature triggered by a button costs you money only when someone wanted the summary. Same capability, very different cost curve, and often a better experience because the user knows when the AI ran.
If your pricing is per seat and usage varies wildly between seats, consider whether a fair use limit belongs in the plan. Most users never hit it, and it protects you from the small number who would otherwise consume more than they pay.
Instrument it from day one
You cannot manage what you cannot see. Log cost per feature, per customer, and per action from launch, not after the first alarming invoice. We treat this as part of shipping, the same as error monitoring.
Two things to watch: which customers cost more to serve than they pay, and whether cost per action is drifting up as prompts grow over time. Both are fixable early and painful late.
What to plan for
For most products we build, AI running cost lands somewhere between a few percent and the high teens of revenue, depending on how central the AI is and how well the workflow is designed. It is manageable. It just needs to be a number in your model rather than a surprise.
If you are working through what an AI feature would cost to build and to run, that is a conversation we have most weeks. Talk to us about your product.