How Should AI Hand a Task to a Human? Designing the Handoff

A good handoff tells the right person, in the tool they already use, exactly what the AI has done, the one thing it needs from them, and by when, and then picks the task back up the moment they answer. Without that system, the AI's unfinished work collects in a queue nobody checks, and the automation fails quietly.

This is the final part of the Digital Employee Playbook. Part four sorted each step of a workflow into lanes, and accepted that some steps need a person. This part is about making sure that person actually finishes them.

Why do human-in-the-loop systems fail?

Usually not because the AI is wrong, but because the handoff goes nowhere. Five failures account for most of it:

• The handoff goes to a shared inbox. Everyone assumes someone else has it.
• It has no deadline. A request without a due time loses to everything that has one.
• It has no context. The person has to reopen the case, find the documents, and work out what the AI already did. At that point they might as well have done the task themselves.
• It never comes back. The person answers, but the answer does not flow back into the workflow, so the last steps still wait.
• People stop reading. When most handoffs are routine, approvals become reflexive. Researchers call it automation bias: a 2012 systematic review in JAMIA found that incorrect advice from clinical decision support software raised the risk of a wrong decision by 26% compared with working without it. Showing the facts behind a suggestion, not only the suggestion, was one of the ways the review found to reduce it.

Each of these is a design problem, and each has a fix.

What should trigger a handoff to a person?

Write the triggers down with the workflow, so a handoff is a planned path, not an error. Most fall into six types:

1. Policy: the decision table says a person decides, such as an authorization or a refund over a set amount.
2. Missing information: the AI asked and did not get what it needs.
3. Low confidence: the AI could not read the document, or two sources disagree.
4. Exception: the case matches none of the written rules.
5. Repeated failure: a portal timed out three times, or a patient did not answer three attempts.
6. Request: the customer asked for a person. Always honor that one.

Two of these mirror the triggers OpenAI recommends in its guide to building agents: failure thresholds, such as repeated attempts that do not succeed, and high-risk actions, which stay with a person until the system has earned trust.

Who should get the handoff, and where?

Route by role, not by name, and deliver it where that role already works.

• Route to a role with a backup. "Billing specialist on duty," not "Maria." People take vacations.
• Use the system of record when possible. A task in the EHR or CRM keeps the handoff next to the case and creates a record of who did what.
• Match the channel to the urgency. A task for today, a team chat message for the next hour, a text for minutes. Email is for things that can wait.

What should the handoff message say?

Everything the person needs to act without opening anything else. Five parts:

1. What happened, in one line.
2. What the AI already did, so nobody repeats it.
3. The one decision or action needed, phrased as a question.
4. The options, ideally as buttons: approve, edit, or reject.
5. The deadline, and a link to the record.

For the referral workflow this series has followed since part two, a handoff reads like this:

New referral from an orthopedic office. Insurance is active, but the plan is out of network. I have prepared the self-pay estimate. Send it to the patient, or close the referral? Needed by 2:00 pm today. Open the referral in the EHR.

The coordinator reads it in ten seconds and answers in one click. The AI does the rest.

What happens when nobody answers?

Every handoff gets a timer and an escalation path, decided before launch:

1. Reminder at the halfway point of the deadline.
2. Escalate to the backup when the deadline passes.
3. Escalate to the manager if the backup does not respond.
4. Tell the customer if the delay will be noticeable, so silence is never the experience.

How does the AI finish the task after a person decides?

The person's answer should be an input, not the end of the case. When the coordinator taps "send estimate," the workflow sends it, logs who approved it and when, waits for the patient's reply, and schedules the appointment when they agree. Nobody reopens the case.

Then review the handoffs weekly. Count them by trigger. A trigger that fires constantly for the same reason is a missing rule, and adding it shrinks the queue. A step people approve unchanged every time is ready to run alone.

How do you measure a human-in-the-loop system?

Five numbers, tracked weekly:

• Handoff rate: the share of cases that need a person.
• Time to response: how long a handoff waits.
• Resolved on time: the share finished before the deadline.
• Override rate: how often the person changes what the AI proposed.
• Top reasons: the three triggers that fire most.

Someone has to own that queue. In Cream Digital's Back Office Operations, trained assistants work the handoffs while AI handles the volume. If you want this designed into your own workflows, the free AI audit is where it starts.