The unit economics of agents at enterprise scale

One number that survives contact with finance

Engineering teams tend to report cost per thousand tokens. Finance cannot use it. Operations cannot use it either, because nobody staffs a contact centre in tokens. The number that travels across all three rooms is cost per resolved interaction: everything spent to take one piece of work from arrival to a closed, acceptable outcome.

Resolved is the load-bearing word. An agent that answers cheaply and gets escalated forty per cent of the time has not made anything cheaper; it has added a step. We insist on defining resolution with the operations team before any modelling starts, and we write it down, because the definition tends to drift once the numbers get uncomfortable.

What actually shows up on the invoice

When we take apart a live deployment, inference is usually somewhere between a third and a half of the run cost. The rest sits in places the pilot never touched, because a pilot has ten users and no retention policy. This is the stack we now model from the start.

Retries are a budget line, not an incident

Agentic systems retry. They re-plan, they call a tool that times out, they take a second pass at a malformed output. In a well-behaved deployment we see fifteen to thirty per cent of total inference spend going to work the user never sees, and that is not a defect — it is how the reliability gets there.

The mistake is leaving it undeclared. If the model in the business case assumes one pass per interaction, the first month of real traffic looks like a forty per cent overrun and the conversation becomes about trust rather than tuning. Declare a retry budget, instrument it per step, and set an alert on the ratio rather than the absolute figure.

The human checkpoint is usually the expensive part

A senior underwriter reviewing an agent-drafted memo for four minutes costs more than the entire inference chain that produced it, often by an order of magnitude. This is not an argument against human review; in most regulated workflows it is the reason the system is allowed to exist. It is an argument for designing the checkpoint with the same rigour you apply to the prompt.

The lever is not removing review, it is narrowing it. Route only what is genuinely consequential or genuinely uncertain to a person, present it so the decision takes ninety seconds instead of four minutes, and let confidently routine work pass with sampled audit instead of full inspection. On one commerce deployment that single change moved cost per resolved interaction by thirty-one per cent without touching a model.

Model mix beats model choice

The frontier-versus-cheap-model argument is usually framed as a single decision, and it is not one. A typical interaction has four or five distinct jobs: classify, retrieve, reason, draft, check. Only one or two of them need the expensive model, and the classification step in front is often a fine-tuned small model that costs almost nothing to run.

What makes this real rather than theoretical is measuring quality per step instead of end to end. Once you can see that step three is carrying the accuracy, you can downgrade the other four with evidence rather than nerve. We have seen mixes take a blended cost down by half while the resolution rate moved less than a point.

The four things to instrument before you scale

Most cost surprises are measurement gaps that existed before launch. If these four are in place on day one, the second month of a rollout is a tuning exercise instead of an argument.

Before you sign the rollout