Human-in-the-loop: where an AI agent should stop and call a person

Three ways to put a person in the loop

Not all human involvement is the same. It helps to separate three formats:

• Approval before action. The agent prepares everything and waits for a “go ahead”. Example: a draft reply to a formal complaint that only goes out after someone in customer service reads it.
• Escalation. The agent handles normal cases on its own and passes to a person the ones that fall outside the norm or where it isn’t confident. Example: an order with a discount above the allowed limit goes to the sales manager.
• Review after action. The agent acts and a person checks a sample or a report afterwards. Example: invoice classification reviewed once a week.

Most well-designed agents use all three, at different points in the same flow.

Four questions to decide

For each action the agent can take, answer:

• Can it be undone? Sending an email, paying a bill or deleting a record are hard or impossible to reverse. Changing a ticket’s internal status is easy. The harder it is to undo, the stronger the case for approval beforehand.
• How much damage if it’s wrong? Think about money, customer relationships, legal exposure and personal data. A mistake on an internal label is annoying. A mistake in a reply to a patient or on a court deadline can be costly.
• How often does it happen? Very frequent, low-risk actions call for sample review. Requiring approval for each one creates queues and fatigue, and fatigue leads to approvals nobody reads.
• Does the agent know when it doesn’t know? If you can clearly identify the doubtful cases, like missing data, an amount out of range or an unusual request, escalation works well. If you can’t, approval beforehand is safer.

The harder something is to undo and the bigger the possible damage, the earlier the person comes into the loop.

Case-by-case examples

The examples below are illustrative and show how the questions lead to different choices, even within the same department.

Imagine an accounting firm with an agent that receives client documents by email. Classifying and filing each document is frequent, easy to fix and low risk: the agent does it alone and someone reviews a weekly sample. Telling a client a document is missing is external communication, but standardized: the agent sends it alone using approved templates. Calculating and sending a tax payment involves money and deadlines: the agent prepares it, and a person approves before it goes out.

Now think of a clinic with an agent on WhatsApp. Confirming and rescheduling appointments is routine: the agent does it alone. Answering questions about test preparation based on guidance already approved by the medical team, too. Any message describing a symptom, a worsening condition or an emergency goes straight to a person, because clinical guidance belongs to licensed health professionals and follows their professional rules.

On a sales team, the agent can qualify leads, log them in the CRM and book meetings on its own. A proposal with commercial terms outside the price list goes for approval. A contact who asks not to be contacted again is handled immediately, without depending on anyone, and recorded.

How to help the person decide quickly

A poorly designed approval point becomes a bottleneck. A few things make a difference:

• Arrive with everything ready. The person should receive a summary of the case, the proposed action, why it stopped and the data that matters. They shouldn’t have to open three systems to decide.
• Use the channel the person already uses. If the team lives in Slack, Teams or WhatsApp, the approval arrives there, with simple approve, adjust or reject buttons.
• Set a deadline and a plan B. What happens if nobody replies within two hours? Does the case go to someone else, does the customer get a notice that the answer will take longer, or does the agent take a safe path?
• Record the decision. Who approved, when and with what adjustment. That serves for audits and for teaching the agent.

The loop changes over time

The initial design is usually more cautious, and that’s good. After a few weeks of operation, the logs show where people approve almost everything without changing anything, and where they correct things often.

Points where the approval is always the same can become sample review. Points with frequent corrections call for adjusting the instructions or the data, or even reinstating mandatory approval. This review should be periodic and agreed with the team, not a silent change made on a busy day.

Autonomy is earned through track record. Start cautious and loosen the reins where the data shows the agent gets it right.

What regulation already says

Some rules already help define these points. Privacy laws like GDPR and Brazil’s LGPD give people rights over decisions based solely on automated processing that affect them, including asking for human review. In practice, that means having a clear path for a person to look at the case when asked. The EU AI Act goes further and requires human oversight for systems it classifies as high-risk.

Regulated professions have their own limits too. In healthcare, clinical guidance belongs to licensed professionals. In law, bar associations have issued guidance on using AI under the supervision of the responsible lawyer. In finance and insurance, there are industry rules on customer service and complaints. Check with whoever handles legal and compliance at your company before you finalize the design.

When in doubt, the safest path is simple: the agent prepares, the person decides.