Per-minute AI pricing charges you for the thing working.
Nobody quotes a receptionist by the minute. The moment a supplier does, you have lost the ability to compare the two things you are actually choosing between.
Published 5 min read Sources checked 8 September 2026
Filed under AI pricing, Software procurement, Vendor contracts
The short answer
Usage pricing transfers volume risk to you and makes your best month your most expensive. It also breaks the only comparison that matters: against a salaried person, who costs the same whether busy or not. Ask for a fixed build fee plus a flat monthly, with a ceiling that fails loudly rather than bills silently.
On this page · 7 sections
A receptionist costs the same in a quiet February and a frantic September. That is not an accident of employment law. It is why you can budget.
Price the same job per minute and two things happen. Your costs now peak exactly when your cash is tightest, and you can no longer put the two options side by side on one page.
Why does the shape matter more than the rate?
Because you are not really choosing between two prices. You are choosing between two risk positions.
A flat monthly means the supplier carries the volume risk. They have to make the thing efficient, because a busy month costs them and not you. Usage pricing reverses that: their incentive is now for the agent to talk a little longer.
Ask yourself who is better placed to carry the risk that volume spikes. It is the person who built the system, not the person answering the phone at 7pm.
What should you ask for instead?
| Ask for | Because |
|---|---|
| A fixed fee for the build | Scope is the supplier’s job to estimate, not yours to discover |
| A flat monthly to run it | Comparable to a wage, which is your real alternative |
| A stated ceiling | Hitting it should say so plainly and offer a conversation, not bill you |
| Your data, exportable | The list, the transcripts, the logs. If you cannot leave, the price will move |
| A price that does not rise per seat | A seat implies a tool you log into. You wanted work done, not another login |
When is usage pricing actually fine?
Three cases, and it is worth being fair about them.
- Genuinely spiky, genuinely optional work: a one-off list build, a bulk document run. You would not want to pay a monthly for something you use twice a year.
- Where the underlying cost really is per unit and it is large: heavy inference on long documents, for example. Then a metered price is honest, and should be shown with the margin visible.
- Pilots. Paying per unit to find out whether the thing works is reasonable. It should convert to a flat price once it does.
What all three have in common: the meter is exposed and explicable. The problem is not metering. It is metering a thing whose volume you cannot control, and calling it a subscription.
The question that ends the conversation
Ask the supplier: “If this works twice as well as we expect, what happens to my bill?”
If the answer is “it goes up”, you are buying the wrong shape. A supplier confident in their system will price the outcome flat and keep the efficiency gain. That is the deal you want on both sides of the table.
What this does not tell you
What any vendor actually charges. This is an argument about the shape of a price, labelled as one in the sources, and we have not surveyed the market. It also assumes your volume grows. A shrinking or strongly seasonal business can come out ahead on usage pricing, because the thing that punishes growth rewards contraction.
A flat fee with a ceiling moves the volume risk back to the vendor, and a sensible vendor prices that risk in. Expect the flat monthly to be higher than your current usage bill in a quiet month; that premium is the insurance.
Questions people actually ask
Is per-minute pricing for AI voice agents a bad deal?
Usually, for predictable ongoing work. It transfers volume risk to the buyer, makes the busiest period the most expensive, and prevents fair comparison against a salaried person, which is the real alternative. It is reasonable for genuinely spiky, optional or pilot work.
What pricing should I ask for instead?
A fixed fee for the build plus a flat monthly to run it, with a stated ceiling that fails loudly rather than billing silently, and an export path for your own data.
What single question exposes the wrong pricing shape?
Ask: “If this works twice as well as we expect, what happens to my bill?” If it rises, the supplier has priced their inefficiency into your success.
Sources
- 01 No external statistics are used in this note. The argument is about contract shape, and the illustration is explicitly illustrative.
Weight: argument, not evidence: labelled as such Re-checked: never: it cannot go out of date
Checked 8 September 2026. Numbers that move (leaderboards, live indices) are re-checked every 30 days; annual datasets and rules in force every six months; dated research once a year. If something here has gone stale before we got to it, tell us and we will correct it and say what changed.
Cite this note
Noxia, “Per-minute AI pricing charges you for the thing working”, Field notes, 8 September 2026; sources checked 8 September 2026. https://www.noxia.co.uk/field-notes/per-minute-ai-pricing-punishes-growth
We quote a fixed build and a flat monthly. On purpose.
No per task, no per minute, no per row. If our system gets more efficient, that is our gain to keep and your price stays where it was.