Agents
·
·
6 min read
human in the loop
Approval steps that protect you without turning every run into a bottleneck.
By
An agent that asks permission for everything is just a slower inbox. An agent that asks for nothing is a risk you will eventually regret. The craft is in choosing the few moments where a person's judgement really matters, and making those moments quick and easy.
This is how we decide where approval steps go, and how we keep them from turning into a bottleneck.
Why approvals matter at all
Agents are good at reading, sorting and drafting. They are also, occasionally, confidently wrong. That is not a reason to avoid them. It is a reason to design for it. A well-placed approval step turns a possible mistake into a ten-second check, and it gives the team the confidence to let the system do more over time.
Approve what is hard to undo
The simplest rule we use: if you can't take it back, a person says yes first. In practice that means:
Payments and refunds.
Emails and messages to customers, suppliers or anyone outside the company.
Changes to contracts, prices or customer records.
Deleting anything.
Internal drafts, tags, summaries and suggestions usually don't need approval. If they are wrong, someone fixes them and moves on.
Review in batches, not one by one
Approval steps fail when they interrupt people all day. Ten notifications an hour means none of them get read properly. Instead, let the agent prepare its work and queue it up. Someone reviews twenty drafted replies in one sitting, approves most of them with one click, edits two and rejects one. That takes minutes rather than hours, and the quality of the review is far higher.
Watch the edges
Not every case deserves the same treatment. Route the unusual ones to a person automatically, even if the normal ones go straight through:
A supplier, customer or address the system has never seen before.
Amounts above an agreed limit.
Cases where the agent reports low confidence, or where two sources disagree.
Anything mentioning a complaint, a cancellation or legal action.
Most of the risk lives at the edges. If you catch those, you can be relaxed about the middle.
Make approving fast
A good approval screen shows the person everything they need to decide, in one place. That means what the agent wants to do, why it thinks so, the source it read and a clear way to approve, edit or reject. If someone has to open three tabs to check one decision, they will start approving without looking, and the step stops protecting you.
An approval step is only as good as the thirty seconds a person spends on it.
Loosen it over time
Start strict. In the first weeks, more things should need approval than strictly necessary, so the team can see how the system behaves. Keep a simple record of what was approved, edited and rejected. When a step has run cleanly for a month, move it from “approve first” to “review after”, where a person spot-checks a sample instead of every case.
A simple approval policy to copy
Anything external or irreversible needs approval before it happens.
Anything unusual is routed to a person automatically.
Everything else runs, and a sample is reviewed every week.
Every rule is written down where the team can see and change it.
Four lines are enough for most teams. Write yours down before you build anything, and the build itself becomes much simpler.
Sample article for the Relay template. Replace it with your own writing.
Field notes from building agents and automations for real teams. Short, practical and free of hype, with a new post every fortnight.
We are always glad to meet operators and engineers who hate busywork.