How to bill for AI agents that run on your harness

Three things you're selling

• The build: discovery, instructions and checklists, tool connections, evaluations. A project with a clear scope.
• The upkeep: fixing failures, updating instructions as the client's business changes, moving to better models, monitoring. Ongoing.
• The runs: the work agents do, what it costs in model usage, and the time people spend reviewing it.

Pricing models that work

Most teams combine several of these: a setup fee, then a base retainer plus usage or per-outcome pricing for the runs. Some agencies choose not to charge for the orchestration layer at all and fold it into outcome prices, which works too as long as you know what it costs you. The options:

1. A setup fee for the build: a fixed price or time and materials, like any implementation project.
2. A monthly retainer for upkeep: improvements, monitoring and an agreed amount of human review.
3. Usage passed through with a markup: model and tool costs, shown openly, plus a margin. Published agency guides range from 15–20 % to 30–60 %.
4. Per outcome: a price per accepted deliverable, such as a report, a resolved ticket or a draft. Best once you know your acceptance rate and cost per result.
5. Agent time at an agent rate: the time agents logged on tasks, usually below the human rate, with the reviewer's time billed on its own line.

Usage or outcomes?

Usage pricing is safest while you're learning, because you don't carry the cost of failed runs. Its weakness: every efficiency gain you make lowers your own revenue.

Outcome pricing is where a good harness pays you back. The client pays for results, not attempts, so a better acceptance rate means a better margin. Move from usage to outcomes once your numbers are stable, and agree an error budget: an acceptable failure rate, and what happens when you exceed it.

What to record for every run

This record is what lets you price per outcome, show a client what the retainer bought, and see which harness changes paid off. For every run, keep:

• The request: who asked, and which client, project and task it's for
• Which agent ran it, on which model
• What it did: progress notes, questions it asked, the result
• Time worked, tokens and cost
• Who reviewed it, how long that took, and whether it was accepted

Showing it on the invoice

Use separate lines, for example: "Harness setup (fixed)", "Harness upkeep, September", "Agent work: 42 client reports, reviewed by Anna", "Model usage at cost + 15 %". Link each line to its record. Clients rarely dispute what they can inspect.

Settle these in the contract

• Who owns the harness, and who owns the client-specific parts
• Usage caps, and what happens when a cap is reached
• What counts as an accepted outcome, and the error budget
• Which work a person reviews before it reaches the client
• Which model providers may see the client's data

How Hourtick handles it

In Hourtick each agent is a workspace member with its own token, model and skills. It takes tasks from people or other agents, asks in the thread when it's stuck, and logs its time and reports its tokens and cost against the task. Billability follows the task type, the same as for people.

Reports show agent time and cost per client, project and agent, next to people's hours, and entries lock once invoiced. Track the harness build and upkeep as their own projects or task types, so the setup fee and the retainer have evidence behind them too.