What does an agent cost? Start with the ceiling

What does an agent cost? Start with the ceiling.

Most questions about AI in an organisation are about what it can do. The question that comes with the first real rollout is a different one: what does it cost, and what happens if it gets out of hand?

An agent that works uses tokens. Every time it reads, thinks or writes, that costs money. An agent caught in a loop can use more in one night than you expected for a month. That is why every role I run gets a budget before it is allowed to do anything.

The budget has three levels:

1. Per task. One ticket, one turn, a fixed upper limit. When it is reached, the agent stops and says where it got to.

1. Per day. All the tasks of a role together may not exceed a daily amount.

1. A ceiling. A hard limit nothing goes through, not even when there seems to be a good reason.

That feels strict, but it gives you two things. You know in advance the most it can cost, which makes the decision to start much smaller. And you get a sum: what does this role cost per week, and what would the same task otherwise cost in time?

That sum is not hard, but you do have to make it before you scale up:

• how many tasks per week does the agent do;
• what does that use in tokens, in euros;
• how much time did the same work take before, and who checked it;
• how often did something have to be fixed.

Without those four numbers, "it feels faster" is the only evidence, and that convinces nobody who has to decide whether the agent may stay.

A stop condition belongs to it too. My roles stop when there is doubt whether something fits the rules, and when a source sits behind a paywall or a login. Stopping costs nothing. Carrying on without knowing what it costs sometimes does.

What is easily overlooked: the costs are not only tokens. An agent that often gets something wrong also costs the time of the person who fixes it. So measure, next to the usage, how often something had to be rolled back. A cheap agent that causes a quarter of an hour of repair work every day is not cheap.

That is why I start small. One role, one task, one week of measuring. Only when the sum works out does the agent get more work. And every month I look again: is the task still the same, has the usage stayed level, and is the result still as good as in the first week?

Do you already have an upper limit for what an agent or an automation may cost, before you switch it on?

#AIagents #AI #Workprocesses #ProcessImprovement

All pieces