Every conversation about AI agents eventually arrives at the same worry. "What if it sends the wrong email? What if it pays the wrong invoice? What if it deletes something?" The worry is reasonable. The usual responses are not: either "trust us, the model is very accurate" or "we'll keep a human in the loop," said vaguely, with no design behind it.
We build every agent around one rule: a person approves every consequential action. This article is about what that rule means in practice, how to decide which actions count, and how to design the approval step so it speeds work up rather than slowing it down.
Which actions need a checkpoint
Not every step needs a person. If it did, the agent would be useless. The test we use is simple: can this step be undone cheaply, and does anyone outside the company see it?
| Agent does it alone | A person approves first |
|---|---|
| Pull records from a system | Send an email or message to a customer, patient, or vendor |
| Match, classify, sort, and flag | Move money: pay, refund, invoice, adjust a balance |
| Draft a reply, a summary, or a document | Delete or overwrite a record of any kind |
| Recommend a next step | Change someone's access, status, or schedule |
| Write to its own scratch space or a review queue | Commit anything to a system of record that others rely on |
The left column is where most of the hours go. The right column is where most of the risk lives. Separating them is the whole design.
What a good checkpoint looks like
A checkpoint that just says "approve?" with a wall of text underneath will be rubber-stamped within a week. People approve what they can verify quickly. So the screen has to do the reviewer's work for them.
Show the batch, not the item
If the agent drafted seven follow-up emails, show all seven on one screen with the reason each one exists. A reviewer can scan seven drafts in two minutes. Seven separate approval requests would take fifteen and feel like nagging.
Surface the exceptions first
Out of 128 claims, 121 matched cleanly and 7 did not. The 7 are what the reviewer needs to see. Put them at the top, say why each one was flagged ("missing provider ID", "duplicate service date"), and let the clean 121 be collapsed and approved in one click.
Make the correction cheap
A reviewer who spots a bad draft should be able to fix it in place, not reject the whole batch and wait for a rerun. Edits made at the checkpoint are also the most valuable training signal you will ever collect, so log them.
Say what will happen on approve
"Approve 7 drafts" should mean exactly that: seven emails will be sent to these seven addresses, now. No side effects, no "and also update the CRM" that the reviewer did not see. If there are downstream effects, list them.
Record who, what, and when
Every approval is logged with the person, the time, the exact content approved, and any edits they made. This is what your compliance team will ask for, and it is what lets you answer "why did this go out?" six months later.
Design it in, don't bolt it on
The most common failure we see is a checkpoint added after the fact. The agent was built to run end to end, and then someone inserted a confirm dialog before the last step. It works, technically. But the agent's work was never shaped for review, so the reviewer sees a raw dump, can't change anything, and either approves blindly or turns the feature off.
Built in from the start, the checkpoint changes how the agent itself is designed. It batches work so it can be reviewed in batches. It explains its reasoning in a sentence because a person will read it. It separates "what I found" from "what I propose" because those are shown differently. The result is an agent that is easier to trust and, not coincidentally, easier to debug.
Doesn't this slow everything down?
Less than you'd think, and the comparison is not against a fully autonomous agent. It is against the current process, where a person does all of it. If a claims specialist spent three hours a day on follow-ups and now spends ten minutes reviewing a batch the agent prepared, the checkpoint cost ten minutes and saved nearly three hours.
Over time, as the logs show the agent's drafts are approved unchanged at a high rate for a given category, you can discuss widening what runs alone. But that is a decision you make with evidence, after launch, with the ability to tighten it again. It is not a decision you make on day one because the demo looked good.
The selling point nobody expects
Teams usually see the checkpoint as the price of safety. In practice, it is often the reason leadership says yes. "Nothing goes out the door unless one of our people says so, and we can see exactly who approved what" is an easy sentence to say to a board, a regulator, or a nervous department head. The agent gets built because the checkpoint exists, not in spite of it.
Want help picking and shipping your first one?
Discovery is free: a 45-minute call, then a short report of potential AI solutions with estimates and an honest recommendation, whether or not you build with us.
More from KloudDevs: Why most AI projects stall between the demo and the deploymentHow to pick the first AI use case worth building