AI features have a unit cost, and design decides it

By on

Dials on a mechanical electricity meter

Every other interface decision you make is free to run. This one is not. A single default setting can multiply what a feature costs to operate, and the person choosing it is usually not looking at the bill.

For twenty years, design decisions were free to operate. Moving a button, adding a step, changing a default: all of it cost time to build and nothing to run. Serving the millionth user cost about the same as the thousandth.

A feature backed by a model breaks that. Every interaction has a marginal cost, and a surprising amount of it is set by interface decisions rather than engineering ones.

Decisions that move the number

  • Autorun versus a button. A summary that generates when the page opens costs on every view. The same summary behind a button costs on the fraction of views where someone wants it.
  • How much context you send. Attaching the whole document because it was simpler than deciding what matters is a design choice with a line item.
  • Regeneration affordances. A prominent "try again" invites three attempts where one would have done.
  • Streaming a response the user abandons. You pay for the whole generation regardless of whether they read past the first line.

None of those look like cost decisions in a design review. All of them are.

The autorun trap

Automatic generation almost always tests better. It feels effortless and it demos beautifully, so it wins the room.

Then it ships and the ratio bites. If a fifth of visitors actually wanted the output, four fifths of your spend produced something nobody read, and that ratio holds forever. The version behind a button feels marginally worse in a usability test and costs a fraction to run at scale.

This is a real trade, not an obvious win either way. The point is to make it deliberately instead of discovering it in a finance review.

Where it changes your pricing

Flat-rate subscriptions assume usage does not carry marginal cost. Once it does, your heaviest users can become your least profitable, and the pricing page becomes a product decision rather than a marketing one.

That is not an argument for metering everything, which users hate and which makes them ration a feature you want them using. It is an argument for knowing the number before you set the price.

Where to start

  • Work out the cost of one interaction with your main AI feature. One number. Most teams cannot produce it on request.
  • Multiply it by monthly active usage and put that figure next to the feature in your roadmap.
  • Find anything that generates automatically and measure what fraction of users engaged with the output. Below half, try the button.
  • Ask what changes at ten times current volume. If the answer is nothing, you have not done the arithmetic.

Tags: ai, product, design