Who Pays When the Agent Runs.

For twenty years business software was priced by the seat. A login, a monthly price, the same whether the person used it every hour or never. Agents break that, and the companies approving their first agent budget this quarter are about to find out how. An agent costs something each time it runs, it runs on behalf of a workflow rather than a person, and the bill arrives from three places at once. This is the arithmetic, so the number you approve is the number you pay. Everything on screen is a mock demo with sample figures, and we quote no prices, because they move monthly.

The structure is what holds. Here is the difference on one card. On the left, a seat: one person, one price a month, flat. On the right, an agent: it ran four hundred times last week for the order workflow, each run read three systems and wrote one note, and each run cost a little. Multiply, and the week has a cost that goes up when the business is busy and down when it is quiet. That sounds like a problem and it is actually the opportunity, because for the first time the cost of the work tracks the work.

The trouble starts only when nobody is counting. A seat bill is boring and predictable. An agent bill is a meter, and a meter nobody reads produces the one outcome we see most often: the agent is switched off in month four, not because it failed, but because the invoice surprised someone. Line one is usage. Every run spends a little on the model that does the reading and reasoning, a little on each system it reaches, and a little on storage for what it keeps. Usage is the line the vendors talk about, and it is the smallest of the three for most companies our size, as long as the agent is scoped to a workflow and not left to wander.

The decision on this line is scope. An agent that reads the whole customer history every time it answers a simple question burns ten times the usage of one that reads the three fields the question needs. Scope is set when the workflow is designed, and it is why an agent designed by someone who knows the data costs less to run than one designed by someone who knows only the model. Line two is operations, and it is the line the demo never shows. Somebody writes the rules the agent follows.

Somebody reads what it did last week and finds the rule that was wrong. Somebody keeps the five doors into the five systems working when a system is updated. Somebody sits at the gate when the agent proposes something that spends money. That is people's time, every week, for as long as the agent runs. For a company of fifty to five hundred people this is usually the largest line, larger than usage, and it is the line that decides whether the agent keeps working or quietly degrades. An agent with no one operating it is a flow nobody owns, and we have a video about what happens to those.

Budget operations as a line, not as somebody's spare afternoon. Line three is the outcome: what the workflow produced that it did not produce before. Orders released in two minutes instead of a day. Invoices chased without a person chasing. Leads answered within the hour. It is the only line that justifies the other two, and it is the one most companies never write down, which is why they cannot tell in month four whether to keep paying. So here is the meter. One row per workflow: runs this month, usage, the operations hours it consumed, and the outcome it kept.

Cost per run, cost per outcome. Not cost per seat, which measures nothing, and not one blended number for all agents, which hides the one workflow that is eating the budget. The meter is the single most important thing to agree before the first agent is switched on, and it is cheap: it is a report. Who pays is now answerable. The department whose workflow it is. Sales pays for the lead-response agent out of the number it is already measured on. Finance pays for the invoice chase. Not an AI budget owned by nobody, which is the budget that gets cut first, but a line in the budget that already owns the outcome, metered by the report on the last card.

And the price model that survives this arithmetic is the one the large managed-service firms have already moved to for the enterprise, and the one we sell to companies of fifty to a thousand: a subscription, with a usage tier that moves when the runs move, one accountable party operating the layer, and the outcome written into the agreement. We run our own marketing on exactly that shape, count its cost by workflow every week, and bring that number to the cost review. It is the only way I know to keep an agent that is paying back from being cancelled by its own invoice.

Usage, operations, outcome. Metered per workflow, paid by the department that owns the result, priced as a subscription with a usage tier. Approve an agent budget in that shape and the invoice in month four is a number you already understand. If you want the arithmetic run on your company, with our own numbers beside it, that is the thirty-minute all-in cost review, no obligation. We are CodeStringers. We design, build and run business systems for companies of fifty to a thousand people, and the agent layer over them. The link is below.