Price AI by the cost per task, not the token

Here’s the mistake I’d steer you away from. Paying for AI by the token, a unit nobody runs a firm in. You run yours in finished work, like a quote sent or a schedule built. My view is that AI cost per task, measured by the finished job, is the number worth watching, and once you price AI that way the fog around the bill lifts.

That count matters more each quarter. Statistics Canada found 19.2% of Canadian businesses used AI to produce goods or deliver services in the year to spring 2026, triple the 2024 share, and cost was the second most common barrier owners named.

AI cost per task is the number that actually matters here, not the token count. A token is a small chunk of text, a few characters at a time. Vendors count them because counting them is easy. Your customers have never paid you for chunks of anything. They pay for a fence built, a return filed, a room ready. When the unit on the invoice has no shape in your head, you cannot tell a fair price from a padded one, and you cannot tell which jobs the AI is winning for you and which ones it is losing money on.

Why does the per-token price mislead owners?

It misleads because the token count is only one line of the real cost. A finished job also carries the runs that failed, the drafts nobody used, the minutes your people spent fixing the output, and the second tool you opened to check the first one.

Two models can sit side by side with token prices far apart and land within pennies of each other once the whole job is counted. The cheaper unit often needs more attempts. The dearer unit often finishes in one pass and hands your team less repair work. Per-token pricing hides all of that, which suits the seller fine. Whichever model wins on cost, I’d still treat it as a rental, which is the whole case for owning the layer a rival can’t buy.

There is a second trap under the first one. When the meter runs by the chunk, the team learns to write short, timid prompts, and short prompts make thin work. People start rationing the context the model needs to do the job well, then blame the model for the weak result they ordered.

How do you price a finished AI job?

You price it the way you would price a subcontractor’s piece work. Pick one job the AI does every week. Count what one finished, usable piece costs you, all in. The model fees for every run it took, including the discards.

The staff minutes spent prompting, checking, and fixing, at the rate you actually pay for those minutes. Any second tool in the chain. Then put that number beside the old way of doing the same job, a person’s hours start to finish, and beside what the finished piece is worth to the firm.

Run that count on three jobs and patterns show up fast. The summary job that looked nearly free stays nearly free. The proposal draft that looked cheap turns out to cost real money in senior time, because a partner still rewrites half of it. The research job everyone assumed was a luxury turns out to be the bargain of the three, because the old way burned a full afternoon. The same all-in habit works from the invoice side, where costs hide after the headline price.

Which jobs should move to the cheaper model?

The jobs with a clear right answer and a low cost for being wrong. Sorting, first drafts of routine replies, summaries of documents you already trust, turning notes into lists. A cheap model that finishes these in one or two passes costs you almost nothing per finished piece, and the per-token saving is real because the repair time is near zero.

Keep a person reading the output before it leaves the building, and count that reading inside the job’s price, the way this post counts everything.

The test is the finished piece, never the demo. A model that dazzles on a sample paragraph and then needs four runs and a rewrite on your real work is the expensive model, whatever its rate card says.

Which jobs earn the expensive model?

The jobs where judgment is the product. A recommendation a client will act on. A letter that carries your name into a dispute. The first pass at a decision you cannot easily undo.

On those, one strong run with the full context loaded beats five cheap runs and an hour of repair, and the finished piece costs less even at the higher token price. The same rule settles the switching question. A bigger model earns its place job by job, never across the whole firm at once. If I were setting the budget, I’d make every upgrade argue for itself one job at a time.

Price the finished job and the model menu stops being a fog. Each job gets the cheapest tool that finishes it well, and the expensive seats stop subsidising work a sorter could have done.

Here is how the count usually sorts the work.

Job	Model tier	Why
Sorting and filing	Cheaper	Clear right answer, almost no repair time
First drafts of routine replies	Cheaper	A person reads them before they leave
Summaries of documents you trust	Cheaper	Being wrong costs little
Recommendations a client acts on	Stronger	Judgment is the product
Letters in a dispute	Stronger	Your name carries the risk

My take: a rate card tells you nothing until you count the finished job.

Which three AI jobs should you cost this week?

Take the three AI jobs your firm runs most. Cost each one as a finished piece, all in, fees and minutes together. Move one job down a tier or up a tier based on what the count says, and write the count where your team can see it, so the next buying decision starts from your numbers instead of a rate card.

Setting up that count across a whole firm is usually week one of an AI Strategy Roadmap.

Common questions about AI cost per task

Cost your three most common AI jobs as finished pieces before the next bill arrives. I share more pricing calls like this one in the AI Newsletter with ZAK.