The estimate is the document the CFO sees before the project. The bill is the document the CFO sees every month after it, and the two are written by different people with different incentives. We wrote the estimate up as four lines in an earlier piece. This one is the bill, for a company of about a hundred people running one agent that does a real job, and it has three lines on it. Two of them are usually missing from the budget.
Tokens are the metered part. Every time the agent reads a record, reads the records around it, reasons, and writes, it consumes model capacity that is billed by volume. The number depends on three things the company controls: how many records the agent handles a month, how much context each one needs, and which model does the reading.
For a hundred-person company, the volumes are modest. A sales agent that drafts follow-ups and quotes handles a few hundred to a few thousand records a month. A support agent that proposes resolutions handles whatever the ticket volume is. A finance agent that reconciles counts runs a few times a day. At those volumes, with a capable model and sensible context, the token line is in the low hundreds of dollars a month, and it is the line the budget was built on.
It is also the line most likely to surprise in the other direction. A token bill that doubles month over month is not a pricing problem; it is a sign the agent is reading too much per record, usually because the records are messy and it has to fetch the history to make sense of the present. Clean records are cheaper to reason over. That is the first place the second line pays for itself.
Operations is the line the licence pitch leaves off, and it is the largest line on the bill for the first year.
An agent is a layer over the systems the company runs. The connections to those systems break when a vendor changes an API, when a field is renamed, when a permission is revoked by someone tidying up users. The permission boundary has to be reviewed when the agent's job changes or a person leaves. The trail has to be read, because a trail nobody reads is decoration. The exceptions the agent raises, the records it could not act on, have to go to a person who fixes the record rather than the agent. And the agent's outputs have to be checked against the outcome, continuously, because the day a quote goes out wrong is the day the agent is switched off.
That is a part-time job for someone who understands both the business process and the systems, every month, for as long as the agent runs. Priced honestly it is between a quarter and a half of a skilled person for one agent in the first year, falling as the records clean up and the exceptions thin out. Whether it is done by an internal person, a consultancy on hours, or a managed subscription with an accountable party, it is on the bill. We set out how the human side of this work gets priced in how Zoho consulting services are priced; the same three models apply to running an agent, and two of them pay the supplier for the problem to persist.
A budget approved on line one and silent on line two is approved on the smallest number on the bill. It is blown by month three, and the agent is blamed.
The third line is the one that justifies the other two, and it is the hardest to put on a bill because it is a saving rather than a spend. It has to be measured, or it is a story.
The outcome is what the agent's job was for. Quotes out in an hour instead of two days, measured as the days between request and quote. Tickets resolved on the first reply, measured as a rate. Three inventory counts that agree, measured as the number of corrections at month end. Each of those has a cost in the business today, in people's hours or in lost orders or in write-offs, and that cost is the third line.
The arithmetic the CFO should insist on is this: line three, measured, must exceed line one plus line two within the first year, or the agent is a cost centre with a good demo. For a hundred-person company that usually means the agent has to save the equivalent of one person's time, or recover one deal a month, or stop one class of write-off. Those are not heroic numbers, and that is the point. An agent with a clear job in a clean part of the business clears them. An agent asked to do everything clears none of them, because nothing can be measured.
We priced our own operation this way, line by line, in what it costs to run a marketing function as a system. The tokens were the smallest line there too.
Put the three lines together for a single agent with a real job at a hundred people, in the first year. Tokens in the low hundreds a month. Operations the equivalent of a quarter to half a person, which is the largest line and the one that shrinks. The outcome, measured against a baseline taken before the agent started, which must clear the sum of the first two.
That shape is why we think an agent is an operating cost with a usage tier rather than a licence. A licence is priced per seat, forever, regardless of whether the agent did anything useful. A usage tier is priced on what the agent handled, with the operations line included, and it moves when the volume moves. It is the same reason a company buys electricity by the unit and not by the number of sockets.
Before approving an agent budget, ask for all three lines. If the proposal has only the token line, it is a licence pitch. If it has tokens and build but no monthly operations, it is an estimate, not a bill, and the operations line will arrive uninvited. If it has no measured baseline for the outcome, there is nothing to hold the agent to in month six.
Then ask who is accountable for line two, by name, and how line three will be measured, by number. Those two answers are the budget. The model is a detail.
Our managed subscription prices the three lines together by usage tier, with one party accountable for operations and measured on the outcome. Discovery is no-risk: we take the baseline for the job the agent would do, put the three lines against it, and you pay only if the arithmetic says to go ahead.
In a no-risk discovery we look at your CRM, finance and operations systems and who owns each part, and show what a connected system would do differently. You pay only if you proceed. Or see how we approach it.
More on the same problem:
"Something isn't working." "We're outgrowing our tools." "We're afraid to make the wrong move." No-Risk Discovery is a short, practical conversation that gets you clarity before you commit to anything big. We'll tell you if we're a fit. If we're not, we'll tell you that too.
Say who you are and ask. A person replies within one business day.
