What an AI agent actually costs per completed task
Token pricing is not the interesting number. Cost per successfully completed task, including retries and human escalation, is the one that decides whether an agent ships.
Key takeaways
- Price the agent per completed task, not per thousand tokens.
- Retries, failed tool calls and human escalation belong inside the unit cost.
- A cheaper model with a higher escalation rate is often the more expensive option.
How should AI agent cost be measured?
Measure cost per successfully completed task. Add up model calls, tool calls, retries and the human minutes spent on escalations over a period, then divide by the number of tasks the agent finished to an acceptable standard. Token price alone hides the retries and the humans.
Token pricing flatters every agent
The first cost model on every AI proposal is tokens in, tokens out, multiplied by volume. It always looks excellent, because it silently assumes every run succeeds on the first attempt and nobody ever picks up the phone afterwards.
Real agent traffic does not behave that way. A meaningful share of runs retry after a malformed tool call. Some loop until a step limit stops them. Some produce an answer that a human then has to redo — which is the most expensive outcome available, because you paid for the model and the person.
Two models, same task, different unit economics
| Metric | Cheaper model | Stronger model |
|---|---|---|
| Cost per run | $0.019 | $0.052 |
| First-pass success | 71% | 93% |
| Escalations per 100 tasks | 29 | 7 |
| Human cost per escalation | $4.10 | $4.10 |
| Cost per completed task | $1.22 | $0.34 |
Build the escalation path before the happy path
Agents that can hand off cleanly cost less than agents that cannot, because a fast escalation is cheaper than a confident wrong answer that gets discovered three days later.
We instrument three counters from day one: completed without help, escalated with useful context attached, and abandoned. The third number is the one that predicts whether the team will still trust the agent in six months.
Set the benchmark against the manual cost
An agent is worth building when the cost per completed task is meaningfully below the fully loaded cost of a person doing the same work — including their tooling, their supervision and the queue time customers experience while they wait.
We look for a 60% reduction or better before recommending the build. Below that, normal variance in volume and model pricing eats the case, and the project becomes a maintenance obligation rather than a saving.
FAQFAQ
Frequently asked questions
About the author
Ravi builds the applied AI systems we ship: retrieval pipelines, evaluation harnesses and the guardrails that keep a model useful once real users reach it. He writes about the engineering that decides whether an AI feature survives production.
- LLM evaluation
- Retrieval-augmented generation
- Prompt engineering
- AI cost modelling
