"What will this cost to run?" is a fair question, and most answers to it are either hand-waving ("pennies") or scary ("enterprise pricing"). Here's how I actually think about the cost of AI doing real work, based on running it every day.
The four costs
When AI takes over a business task, you pay in four places:
- Model usage. What the AI provider charges to process each task.
- Setup. Writing the context, the procedure and the checks, and connecting the AI to your systems.
- Review. The human time spent checking outputs and handling exceptions.
- Maintenance. Updating things when your process, prices or software change.
People fixate on the first one. In practice, it's usually the smallest.
Model usage: how it's measured
AI providers charge by the token. A token is roughly three quarters of a word. You pay for what you send in (the instructions, your context and the documents) and, at a higher rate, for what the model writes back.
The counter-intuitive part is that most of what you send is the same every time. Your context file, your procedure and your examples get read on every task. Providers now offer caching: content the model has seen recently is billed at a small fraction of the normal rate. In my own work, the large majority of tokens are this kind of cached context, re-read cheaply on every turn.
That's why a good context file pays for itself. You write it once, the model reads it thousands of times, and the cost of each read is tiny.
What that means in dollars
To make this concrete: in a test I ran this week, a set of twelve real-world tasks, the kind of bounded build and document work I do for clients, cost about $2.59 on a top-tier model and about $1.36 on a mid-tier one, measured at API prices. A small model did the extraction-style tasks for around ten cents.
Those are tasks that would take a person somewhere between a few minutes and an hour each. Even with generous room for retries, the model bill for a typical small business workflow lands in the range of cents to a few dollars per task. For most companies, a month of routine AI work costs less than a day of staff time.
Two things push costs up:
- Using the biggest model for everything. Top-tier models are worth it for writing and judgment calls. For extraction and checking, a smaller model with a good grader is often just as accurate at a fraction of the price.
- Long, unfocused sessions. An agent wandering through a huge folder of documents burns tokens. A well-scoped task with the right inputs doesn't.
Setup is where the real investment is
The work that makes AI useful is mostly not AI work. It's understanding how the job is done today, writing down the rules and exceptions, building connections to your booking system or accounting software, and testing against real examples until the output meets the bar.
For a single well-defined workflow, that's typically days of work, not months. It's also the part that decides whether the project succeeds, so it's not the place to cut corners.
Review time shrinks, but never to zero
Early on, someone should check every output. As the checks prove reliable, review shifts to exceptions only: the person sees the invoice where two reads disagreed, not the ninety that matched. That's where the time savings come from.
I'd budget for review on every workflow, permanently. It's cheap insurance and it's how you catch drift when something upstream changes.
A simple way to estimate
Before starting any project, I'd work out:
- Volume: how many times a month the task happens.
- Current time: minutes per task today, multiplied by the loaded hourly cost of whoever does it.
- Review time after: usually a small fraction of the original, more at first.
- Model cost: cents to a few dollars per task, depending on size and model.
- Setup: a one-time cost, spread over the first year.
If the monthly saving in staff time comfortably beats the model cost plus review time, and setup pays back within a few months, it's a good candidate. If not, pick a different task. There's always one with better numbers.
The bottom line
The cost of AI work isn't the token bill. It's the thinking up front and the checking afterward. Get those right and the running cost is small enough that the real question becomes which job to hand over next.