Skip to content

What it costs to keep an agent in production

Models, infrastructure, checking time, exceptions and Presidio: what makes up the cost of an agent in production and what drives it up.

JoinX · Last updated

The project has been approved, the agent that prepares claim settlement proposals has been working for a few weeks, and in committee the finance director asks how much it costs to run each month, on top of what it cost to build. It is the right question, because the build cost is paid once while the running cost comes back with every file. In this article we give no prices, which depend on the process and the volumes; we explain instead which items make up that cost, what drives it up and who can act on each one.

The items that make up the cost

The most visible item is model usage, which is usually paid per run according to the amount of text the model receives and produces, measured in tokens. Next to it are infrastructure and third-party licences, which in our proposal we show separately from our own fee. Then there is the time of the people who check the agent's work, and the time spent on the exceptions the agent passes to a person. Lastly, if the company chooses Presidio, there is its recurring fee.

Without amounts, which depend on the process and the volumes.

ItemWhat drives it upWho can act on it
Model usageWhat drives it upContext reloaded on every call, a model larger than necessaryWho can act on itWhoever designs the agent and whoever sets the spending limits
Infrastructure and third-party licencesWhat drives it upThe volume of runs and the environment the agent works inWho can act on itIT, with the suppliers
People's checking timeWhat drives it upA low level of autonomy, checking every file instead of a sampleWho can act on itWhoever is accountable for the process
Exceptions passed to a personWhat drives it upIncomplete rules, missing data, cases the rules do not coverWho can act on itThe owner of the rules and the owner of the data
PresidioWhat drives it upThe number of agents covered and the service level requiredWho can act on itManagement

Why context weighs on token spend

A model remembers nothing between one call and the next, so on every call it receives everything it needs to work. When rules, access rights and memory sit inside the agent, the agent reloads the whole context on every run and the tokens grow with the number of cases it has to cover. In X Lab we study how to keep context, rules and memory outside the individual agent, so that each agent receives only the part that concerns it and the model comes into play only when needed.

It is the same reason why, on the Company Brain page, we write that with organised rules and processes agents produce better results at a fraction of the token spend. The choice of model matters too. We work with those best suited to the task in terms of accuracy, cost and confidentiality. With Security all requests to the models pass through a single point, where it is decided which models are allowed, with which data and within which spending limits.

The time of the people who check

In Deployment we measure the duration, errors and rework of the process before building, and after release we track the same aspects together with the time spent checking the agents. This item depends above all on the level of autonomy. When the agent prepares a complete piece of work, the analyst checks and approves it file by file, whereas when it acts within set limits the person approves the critical steps and checks a sample. The move from one level to the next is decided by whoever is accountable for the process, as we explain in relation to the levels of autonomy. The cost always includes the time for the approvals that no level takes away from people, namely actions that cannot be undone and those addressed to the end customer.

The exceptions that pass to a person

An agent stops when a threshold is exceeded, when data is missing or when it meets a case the rules do not cover, and at that moment the cost of the run is added to the time of the person who picks the file up. When a person corrects a proposal, we record the correction and analyse it to understand whether to act on a source, a rule or the agent. Reusable procedures, that is, activities written once and used by several agents, such as verifying a document, then avoid paying for the same work several times.

Outcomes and costs tracked by Presidio

With Presidio we continuously observe the outcomes, exceptions and cost of each run, on a dashboard the company reads together with us. When a new model comes out we test it on the company's cases and adopt it only if it improves quality or costs without weakening the controls. Presidio is optional, and a company that brings the running in-house shifts this item onto its own people's time.

What is measured, and what the company decides

We measure and show the cost per run, while the financial result depends on choices that are the company's to make, starting with how to use the time the agent frees up, and that is where our measurement stops. That is why our price stays tied to the work done, while the measurement before and after release gives management the numbers to decide. And when the second agent arrives, it finds the rules, definitions and access rights left by the first already in place, as we explain on the Compounding Intelligence page.

Read the cost of an agent in production

If you have an agent in production, or one on the way, we can look together at what makes up its cost per run.