What an AI employee actually costs
Every few weeks a client asks us the same question, in different words: what does it actually cost to have AI running part of my business? Usually they are picturing a subscription — a seat, a monthly fee, something that scales the way software always has. That is the wrong model, and it is worth explaining why.
Priced per token, not per seat
Language models are not sold by the seat. They are priced by the token — roughly, by the amount of text they read and the amount they write. A short internal report might use a few thousand tokens. A long document, or a task that has to check its own work, uses more. The bill is a function of how much work happened, not how many people are on the plan.
That sounds abstract until you see a real number. One of the systems we run for a client composes a short daily operations brief — a summary of what happened across the business the day before, pulled from several systems into one page a manager can read in two minutes. That brief costs about $0.04 to generate. Four cents. Run every business day, that is under a dollar a month for that one task. It is not a hypothetical efficiency; it is a line item you can point to.
The number that matters isn’t the price
Because the per-run cost is so low, the conversation with clients moves quickly away from price and toward two questions that matter more: what happens if something goes wrong, and can we see what it did.
The spending limit and the written record matter more than which model you pick.
Every automated task we run for a client operates inside a spending limit — a ceiling on what it is allowed to cost in a given period, set by us and visible to the client, not something the software decides for itself. And every action the system takes is written to a record: what ran, when, what it changed, what it cost. That record is the thing you actually check when you want to know whether an automated process is behaving. Not the marketing material for whichever model is newest that month.
Which model sits behind a task matters less than people expect. Models change, prices change, and the underlying provider is, in practice, a swappable component. The spending limit and the record are not swappable — they are the part of the system that makes automated work trustworthy enough to leave running.
A person still checks the work
None of this removes a person from the loop, and it shouldn’t. Every task we automate has a human reviewing its output before it reaches a client, at least until we have enough history with that specific task to be confident in it — and for anything that touches money, a person stays in that loop indefinitely. The daily brief mentioned above is drafted automatically and read by a manager before anyone treats it as fact. That review step is not a limitation of the technology. It is the design.
An automation is not an agent
It is worth being precise about a distinction that gets blurred in most conversations about this. An automation follows a fixed set of steps: if this happens, do that. It is reliable and cheap, and most of the repetitive work inside a business — sending a reminder, updating a spreadsheet, moving a record from one status to another — is better served by a plain automation than by anything resembling intelligence.
An agent is different. It is given a goal and some tools, and it decides, within limits, what steps to take to get there — which system to check, what to pull, how to phrase a summary. The daily brief is an agent, not an automation: nobody wrote a script that says exactly which numbers to pull in which order. It composes the brief from what actually happened that day, inside boundaries we set.
The distinction matters for cost and for risk. Automations are close to free and close to predictable. Agents cost more, per run, and behave with a wider — though still bounded — range of outcomes, which is exactly why the spending limit and the written record exist. You use an agent where judgment is genuinely needed, and a plain automation everywhere else. Most of what looks like “AI” in a well-run business operation is actually the second kind, quietly doing repetitive work, with the first kind — the agent — reserved for the handful of tasks that need it.
When a client asks what an AI employee costs, the honest answer is: less than they expect, on a per-task basis — and the interesting number was never the price. It is whether the system is bounded, recorded, and reviewed. Everything else follows from that.