What an AI feature actually costs to run, per user, per month.

A database does not get more expensive because one customer had a busy Tuesday. A language model does. That single difference is why AI budgeting defeats teams who are otherwise good at forecasting.

The number that matters is what one active user costs you in a month, because that is the figure you can multiply. A monthly total tells you what happened. A per user figure tells you what happens if you double.

Long inputs cost more than long outputs in most pricing models, so sending an entire document when a section would do is the most common avoidable expense. Repeated context that never changes should be cached rather than resent.

Voice is in a different cost bracket from text and is the line most often left unmodelled until it appears on an invoice. If your feature includes spoken conversation, model it separately and cap it per user.

Retries are the quiet one. Anything that runs unattended and retries on failure can multiply cost without changing behaviour, which is why the expensive work belongs inside a retry boundary rather than outside it.

You can state the cost per active user per month from a query rather than an estimate. You know which feature is the most expensive and why. Your spending ceiling has been triggered deliberately at least once so you know it works. And the usage that drives your invoicing and the usage that drives your costs come from the same record, so they cannot drift apart.

What is a normal cost per user for an AI feature?

It varies enormously by feature. Text based assistance for a typical business user is often well under a pound a month. Anything involving voice, long documents or heavy automated processing can be an order of magnitude higher, which is why per feature attribution matters more than a benchmark.

Can we cap what a customer costs us?

Yes, per user daily and monthly ceilings are straightforward to enforce and are the difference between a predictable line and an open ended one. The important part is testing that the cap actually stops work rather than assuming it would.

Does a cheaper model always mean worse results?

No. For classification, extraction and routine drafting, smaller models often perform indistinguishably at a fraction of the cost. The skill is knowing which work genuinely needs the expensive model, which is what routing decides.