The Model Costs Pennies. The Rest Costs a Salary.

Here is a question I get asked in almost exactly the same words every time. What does the AI cost?

It is a reasonable question and the honest answer is annoying, because the AI is the part that costs almost nothing. Quoting the model price when someone asks what an AI system costs is like quoting the price of the paint when someone asks what it costs to paint a house. It is a real number. It is a true number. It is nowhere near the number they need.

There are four bills. Most people are shown one.

Bill one: the model, which really is pennies

Let me make the cheapness concrete, because it is more extreme than most people realise and it has changed fast.

Cheap capable model tiers now run at roughly a dollar per million tokens of input. A million tokens is somewhere in the region of three-quarters of a million words. So a workflow that reads two thousand documents a month, at a couple of pages each, is processing a few million tokens and costing tens of dollars. Not thousands. Tens.

Prices have fallen hard across the board. OpenAI’s cheapest current tier launched at a dollar per million input tokens, Anthropic launched Sonnet 5 at introductory pricing of two dollars, and the open-weight Chinese labs are well below that again. Meanwhile the cost of an hour of experienced engineering has done the opposite of collapse.

That divergence is the whole article. The intelligence got cheap. The work of putting the intelligence somewhere useful did not.

There is one caveat worth carrying, which is that cheap per unit becomes expensive at volume when the structure is wrong. An agent that loops, retries and re-sends its entire history on every step can turn fractions of a cent into a genuinely alarming invoice, which is a mistake I have made personally and written about. Structure decides the bill, not the sticker price. But with sane engineering, for most small-business workloads, inference is a rounding error.

Bill two: integration, which is most of the build

This is where the money actually goes, and it is worth being precise about why.

Almost no AI project is a model sitting by itself. The value comes from the model touching things: your booking system, your accounting package, your document store, the twelve-year-old line-of-business application nobody wants to open. The connecting work is ordinary software engineering, and it is where the hours accumulate.

It expands for reasons that are entirely predictable and consistently underestimated. The old system has no API, so you work around it. It has an API but the documentation describes a version from four years ago. Two systems disagree about what a customer is. Authentication has to be handled properly because you are now moving real records around. Somebody has to decide what happens when the connection fails at two in the morning.

None of that is AI work. All of it is the project. If a proposal you are reading spends most of its length on the model and a paragraph on integration, it has been estimated backwards, and the paragraph is the part that will overrun.

Bill three: the data, and the people

Two costs that get skipped for the same reason: neither looks like technology, so neither lands in a technology budget.

The data work. You cannot point a system at records that live in three inboxes, a spreadsheet on someone’s desktop, and the memory of your longest-serving employee. Getting to a structured, consistent, current record is frequently the largest single piece of a first AI project, and it is the piece clients least expect to pay for, because it feels like it should already be done. It is also the piece with the longest tail of value, since the record you build outlasts whatever model you point at it.

The people. Someone has to be trained. Someone has to review output while trust is being established, and that review labour is real payroll, often for months. Someone has to own the exceptions. And someone has to do the unglamorous work of changing how a team does a thing they have done the same way for nine years.

There is a well-worn consulting heuristic that puts roughly ten percent of the value of this kind of work in the algorithms, twenty percent in the technology and data, and seventy percent in people and process. Treat that as a rule of thumb rather than a measurement, because it is a characterisation of case work rather than a published dataset. But the shape of it matches what actually happens, and it is echoed in where the difficulty shows up in survey data.

Which brings me to the number in this article I find most telling. KPMG surveyed 753 Canadian business leaders in 2025 and asked where AI money was going. Forty-six percent were spending on hiring technical talent, forty-one percent on buying tools, and only thirty-three percent on change management and adoption. The least-funded line is the one that most reliably decides whether any of it works. In the same research, only two percent of those leaders said they were seeing a return. Those two findings are not unrelated, and the sample skews large, so read it as a signal about priorities rather than a verdict on small business.

Bill four: keeping it alive, every year, forever

This is the one that is genuinely missing from most quotes, and it is the one that recurs.

A running AI system has a standing annual cost, and it is not just hosting. It includes the search or vector index if you built one, monitoring so you find out when quality drifts before your customers do, the ongoing human review time in the workflow, periodic security and access review, and, unavoidably, migration.

Migration deserves particular attention because it is not hypothetical and it is not on your schedule. OpenAI retired a generation of models in early 2026 and Microsoft began auto-upgrading standard deployments in March of that year. Anyone whose system was tuned to a specific model had to move, on the vendor’s timetable, and the businesses that had the worst time were those pinned to a region where the replacement was not offered. Prompts calibrated to one model’s behaviour do not transfer cleanly. Evaluation sets have to be re-run. Budget for a model migration roughly every eighteen months and you will be about right; assume it will never happen and you will be wrong on somebody else’s timeline.

The practical rule: if a proposal quotes you a build price and no annual figure, it is not finished. Ask what year two costs. Ask what happens when the model you are built on is retired, and who pays for that.

Putting the four together

Take a plausible small-business project. A document intake workflow: things arrive by email, get read, get keyed into a system, get checked.

The build is mostly integration, so the majority of the invoice is connecting to the mail system and the record system, with the AI portion a modest slice. Before any of that, the existing records have to be made consistent enough to be useful, which is its own line. Then there is training, a period of full human review that tapers as trust builds, and someone named as owner of the exceptions. And every year afterwards: hosting, monitoring, the review time that never quite goes to zero, and a migration somewhere in the middle.

Inference, in that whole picture, is roughly the cost of a couple of coffees a week.

I have deliberately not put dollar figures on those lines, because they vary enormously and a made-up number is worse than none. What I will assert is the shape, and the shape is consistent: the AI is the cheapest thing in an AI project, and the further a line item sits from the model, the more it costs.

This is also why whether a project pays for itself turns on volume rather than cleverness. Bills two and three are largely fixed, paid once regardless of whether the workflow runs forty times a week or forty times a day. Bill one scales with use and stays trivial. So the same build, pointed at higher volume, spreads the same fixed cost over far more saved hours. That is the entire economics of it in two sentences.

What to ask before you sign

Five questions. Any competent firm answers them in a sentence each, and a firm that cannot answer them has not thought about your project properly.

What share of this build is integration rather than AI? What has to be true about our data before this can work, and who is doing that work? How many hours a week of our people’s time does this need, during rollout and afterwards? What does year two cost, in total? And what happens when the model this is built on is retired?

That last one, in particular, tends to produce a revealing pause.

The part worth carrying out the door

The reason this matters is not that AI is expensive. Compared to what it replaces it is often startlingly cheap, and it is getting cheaper every quarter as the model price falls.

It matters because a business that budgets only for bill one will conclude, twelve months later, that AI cost far more than promised and delivered less than expected. Both halves of that will feel true and neither will be quite right. The technology did what it said. The budget described a quarter of the project.

Four bills. The model, which is pennies. The integration, which is most of the build. The data and the people, which is where the difficulty actually lives. And the annual cost of keeping it alive, which never stops and which almost nobody quotes.

Ask for all four in writing. The number will be larger than the one you were shown, and it will be the first honest number anybody has given you.