The first AI bill is always small enough to ignore. The fourth one isn't, and by then nobody can explain it. Someone pulls up the usage dashboard, sees a wall of token counts, and the conversation stalls because token counts don't map to anything a business tracks. Did we spend that money well? Nobody knows.
The problem isn't the pricing page. It's the unit. Model providers sell tokens. You don't run a business on tokens — you run it on invoices coded, tickets resolved, quotes drafted, documents reviewed. Until you price your AI work in those terms, you're guessing.
Define the unit before you set the budget
Pick one workflow and name the finished thing it produces. Not "AI for support" — a customer email answered well enough to send. Not "AI for finance" — a vendor bill coded to the right account and ready for approval.
That unit becomes your denominator. Everything else is arithmetic: total spend on that workflow divided by the number of acceptable outputs it produced. Note the word acceptable. Runs that produced garbage still cost money. They belong in the numerator, not the denominator.
This one change reframes every conversation. "Our AI spend is up" becomes "our cost per coded invoice went from X to Y, and here's why." The second sentence is actionable. The first is anxiety.
What actually drives the bill
Once you measure per unit, the cost drivers stop being mysterious. In most workflows, a handful of things dominate:
- Context bloat. Stuffing an entire policy manual into every request when three paragraphs would do.
- Agent loops. A system that re-reads, re-plans, and re-checks itself five times to answer a question that needed one pass.
- Silent retries. A flaky integration fails, the job retries, and each retry pays full price. Nobody notices because it still works.
- Model overkill. Routing classification and formatting tasks to your most expensive model because that's what was configured on day one.
- Duplicate runs. The same document processed twice because two triggers fire on the same event.
None of these show up as a line item. They show up as a bigger number with no explanation attached.
Put the human minutes in the numerator
Here's where most AI budgets quietly lie. The model cost is visible. The review cost is not.
If a person spends six minutes checking and correcting every output, that time is part of the cost of the unit — and it's usually the larger part. An AI workflow that cuts model spend in half while doubling review time got more expensive, not cheaper.
Track three numbers per workflow: model and infrastructure spend, human minutes spent reviewing or fixing, and the number of acceptable outputs. You don't need a tool for this. A spreadsheet and two weeks of honest sampling will tell you more than any usage dashboard.
Compare against the status quo, not against zero
The question isn't whether AI costs money. It's whether the unit is cheaper than the way you do it now — including the way you do it now when someone's on vacation, when volume spikes, or when the work simply doesn't get done.
So price the current process the same way. What does it cost today, fully loaded, to code an invoice or answer a support email? Many operators have never calculated this, which is exactly why AI spend feels expensive while the manual alternative feels free. It isn't free. It's just already in payroll.
Where the savings usually live
Once you have a per-unit number, optimization gets boring in the best way.
Route by difficulty. Most tasks in any workflow are easy; send them to a smaller, cheaper model and reserve the expensive one for genuine edge cases. Trim retrieval so the system pulls the relevant section, not the whole document. Cache anything repeated — the same product description, the same policy text, the same boilerplate.
And fail fast. A workflow that knows when to stop and hand off to a human is cheaper than one that burns five attempts before producing something unusable anyway.
Set a ceiling and a kill switch
Every AI workflow in production should have a spend ceiling and a named owner who gets alerted when it's approached. Not because you expect a runaway — because runaways happen quietly, usually from a retry loop or a bad trigger, and the only thing standing between a bug and a painful invoice is a limit somebody set on purpose.
Pair that with a weekly glance at cost per unit. If the number drifts, something changed: volume, data quality, a prompt someone edited, an upstream system sending messier inputs. Drift is a signal, and it's one you only see if you're measuring the right thing.
The practical takeaway
AI gets expensive when nobody owns the unit economics. It gets predictable the moment someone does — and predictable spend is what turns a pilot into something the business will actually fund.
If you're running AI workflows you can't price, or you're trying to build a business case that survives the CFO's first question, we'd like to talk. Infraxio helps operators design AI systems with the measurement built in from the start — reach out and let's look at your numbers together.