Every agent we deploy ends in the same place: a person reads what it produced and decides. The question is never whether to keep a human in the loop. It is where, and how much of their time it takes.
Approve the output, not every step
Early pilots often ask people to confirm each step of a run. It feels safe and it is exhausting: after a week, approvals become clicks. We move the check to the end, where a person can judge the result as a whole.
- One approval per deliverable, not per action
- The draft shows its sources, so checking is reading, not searching
- Sending back with a note teaches the agent, the same way it would teach a junior
Decide who approves what
Approval follows ownership. The account lead approves what goes to the client; the finance lead approves invoices above a threshold. Everything else runs and is logged.
| Deliverable | Approver | When |
|---|---|---|
| Client report | Account lead | Before sending |
| Invoice over €5,000 | Finance lead | Before payment |
| Internal summary | Nobody | Logged only |
Thresholds are configuration
Thresholds live in the agent's settings, for example approval.minAmount = 5000, so changing a rule is a review, not a rebuild.
approval:
required_when:
amount_over: 5000
vat_mismatch: true
approver: finance-lead
timeout: 48hMeasure the time it takes
If approving takes longer than doing, the agent failed. We track review time per item in the monthly review.
- Time to review a draft
- Share of drafts kept as written
- Share sent back, and why
The best approval flow is the one people stop noticing because the drafts are right.
Read how we set the baseline in measuring the hours an agent gives back.