When teams adopt AI agents (AI that uses tools on its own to carry work forward), the most common worry is this: "What if the AI sends the wrong thing or deletes something?" But if you make a person press a confirm button at every step, much of the point of using an agent disappears.
The answer lies in where you put approval. This post covers how to break down an agent's work and place human approval only where it's truly needed.
Split the agent's work into three stages
Most of what an agent does falls into one of these three.
| Stage | Examples | Undo |
|---|---|---|
| Read | Finding emails or documents, querying data | No impact |
| Draft | Reply drafts, report drafts, schedule options | Just delete it |
| Execute | Sending email, payments, deletion, changing external systems | Hard or impossible |
Even when reading or drafting goes wrong, nothing leaves the building until a person looks at it. The problem is execution. So the default is to put human approval right before execution.
OpenAI's published guide to designing agents says much the same. It suggests rating each tool's risk as low, medium, or high, based on whether it only reads or also writes, whether it can be undone, what account permissions it needs, and whether money is involved. It also recommends having a person confirm sensitive or irreversible actions, such as canceling orders, large refunds, or payments, until the agent has proven trustworthy enough.
Show three things on the approval screen
If there's just a lone approve button, people end up pressing it without reading. An approval screen needs these three things.
- What: What will actually happen, such as the body of the email to be sent, before-and-after values for a change, or the list of items to delete
- Why: The reason the agent chose this action and the supporting material
- Undo: Whether it can be canceled after approval, and if not, saying so clearly
One more thing: don't let pending approvals pile up. If there's no set approver and no set wait time, the agent sits idle and a person ends up doing the work by hand.
Want a written plan for where to start?AI Readiness Assessment, KRW 490K · priorities and expected scope in a reportRequestKeep records, and grant narrow permissions
There are two more things to take care of alongside human approval.
Records. Log, in order, what the agent read, what drafts it made, and who approved what. When something goes wrong, you need to be able to trace where it went wrong.
Least privilege. Give the agent only the account permissions that job requires. If an agent only writes email drafts, leave out the permission to send. Anthropic's post on agent design also explains that agents can pause at set checkpoints for human review, and recommends extensive testing in sandboxed environments along with appropriate guardrails.
Rollout order: start with drafts
Rather than handing over execution from day one, it's safer to expand in stages.
- Read and draft only: The agent only produces drafts, and a person sends them.
- Execute + approval: When a person presses approve, the agent executes.
- Partial auto-execution: Start handing over tasks that are easy to undo (for example, scheduling internal meetings) without approval.
- Have you split the agent's tools into read, draft, and execute?
- Does the pre-execution approval screen show what, why, and undo?
- Are the approver and the wait time defined?
- Does the agent's account have only the permissions it needs?
Summary
When adopting AI agents, more human approval isn't better; it needs to sit precisely in front of executions that are hard to undo. Split the work into read, draft, and execute, show what, why, and undo on the approval screen, and back it up with records and narrow permissions.
References
- OpenAI, A practical guide to building agents: https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
- Anthropic, Building effective agents: https://www.anthropic.com/research/building-effective-agents
In a one-hour AX consultation we map AI candidates and similar cases for you.
We build web apps, mobile apps and business systems, then add and run the AI they need. We build and operate our own meeting notes, contract and company analysis AI.
See our projects