The true cost of AI agent work: count the review, not just the tokens

Why tokens aren't the cost

In a 2026 field study of 802 developers, AI doubled the number of changes per engineer, and the work didn't disappear: review load per reviewer doubled too, and AI-written changes took about 20 % longer from first human review to merge. In an earlier randomized study, experienced developers believed AI made them 20 % faster while they were measured 19 % slower. Feeling productive and being productive are different numbers.

So the useful question isn't "what did the model cost?" but "what did it cost to get this task done, people included?"

The four parts of an agent's task

• Agent cost: model and tool spend the agent reports, converted to your currency
• Briefing: people's time on the task before the agent delivered (scoping, explaining, preparing)
• Review and rework: people's time after the agent delivered (checking, fixing, sending it back)
• Approval: who finally said it was done, a person or the agent itself

Measure it per task, then per agent

1. Put agent work on tasks, not in loose chats, so time and cost have somewhere to land.
2. Mark the hand-off: the moment the agent delivers (a finished request, a move to review, the time or cost it logs).
3. Split people's time on the task into before (briefing) and after (review) that moment.
4. Cost people's time at their hourly cost rate, not their billing rate.
5. Add it up: full cost = agent cost + briefing + review. Divide by finished tasks for a cost per outcome.
6. Watch two ratios: minutes of review per agent hour, and the share of tasks sent back.

What the numbers tell you

• High review per agent hour on one task type: the agent isn't ready for it, or the brief is too thin
• Many tasks sent back: tighten instructions or acceptance criteria before raising autonomy
• Agent closes its own tasks without a person: fine for low-stakes work, a risk for anything client-facing
• Negative margin per task: the client's price or the agent's scope is wrong

How Hourtick handles it

Reports → Agent outcomes lists every task with agent work that was done in the period: the agent's time and reported cost, people's briefing and review time, how often it was sent back, and who approved it. Admins also see the full cost per task at people's cost rates, the revenue and the margin. The same numbers are in the API (GET /api/v1/reports/agent-outcomes) and over MCP (get_agent_outcomes), so an agent can read its own track record.

To make sure the agent's part is right, connect Anthropic, OpenAI or OpenRouter in Settings → AI provider costs. Hourtick compares what each agent reported with what the provider billed, every month, fills in what's missing and flags double reporting.

Hourtick doesn't score people. It measures agents and tasks: the point is to decide which work to hand to an agent, not to rank colleagues.