Blog Guides
AI-to-human handoff: when should an AI agent hand off to a human?
Six handoff triggers, the context a person needs to reply in one message, and what California's new right-to-a-human law says about speed.
Divine Tech AI 4 min read
An AI agent should hand off to a human when the customer asks for one, when the next step crosses a limit you set or needs judgment, when the answer is not in your own sources, or when the agent has failed twice on the same request. A good AI-to-human handoff gives the person the whole conversation and a record of what the agent already did, so the customer never has to start over.
The route to a person decides how customers feel about the agent. In Gartner's survey of 5,728 customers, 64% said they would prefer companies did not use AI for customer service, and the top worry was that AI would make it harder to reach a person (Gartner, July 2024).
What should trigger an AI-to-human handoff?
Write escalation triggers as rules the agent checks on every turn. Six cover most businesses:
| Trigger | Example | What the agent does |
|---|---|---|
| The customer asks for a person | "Can I talk to someone?", "agent", pressing 0 | Hands off at once, with no retry |
| A limit is crossed | A refund above your approval limit, a discount outside policy | Asks the named approver, or hands off |
| No source for the answer | The policy page is silent, or two pages disagree | Hands off instead of guessing |
| Repeated failure | The customer asks the same thing twice, or an action fails twice | Stops the loop and hands off |
| An action is not confirmed | The booking or refund never shows up in your system | Says so, and hands off |
| A sensitive topic | A complaint, a legal threat, a safety issue, a billing dispute | Hands off with the reason flagged |
Frustration and confidence scores can add early warnings, but a request for a person should never wait on a score. Treat it as a hard rule, in any wording and on any channel.
A limit is not always a full handoff. If a refund only needs a manager's yes, the agent can ask for that approval and finish the task itself. Hand over the conversation when the customer needs to talk to someone.
What should the person receive?
A transcript is not enough. The person taking over should be able to reply well in their first message, which means the handoff carries:
- who the customer is, matched to one record;
- the request in one sentence, with the full history from every channel;
- each action the agent took, and whether your system confirmed it;
- why the agent stopped, and the decision it needs;
- what the customer was told about timing.
On the customer's side, the agent should say that a person is coming and roughly when. Once a person joins, the agent stays quiet in that conversation. When the person is done, the agent can pick up the remaining tasks, with the decision on record.
How fast does a person need to respond?
California now puts a number on it. Governor Newsom signed AB 1609, the Right to Human Customer Service Act, on September 28, 2026 (Office of the Governor). According to the Senate Judiciary Committee analysis, it applies to businesses with more than $500 million in annual revenue, and requires them to:
- offer a clear way to ask for a person on every customer service channel, chatbots and phone lines included, during business hours;
- make a good-faith effort to connect the customer within 15 minutes, or offer a specific appointment within one business day;
- tell customers when they are talking to a chatbot.
The San Francisco Standard reports that the rules start next year. This is not legal advice, so check the final text with counsel. Smaller businesses are not covered, but the numbers are a sensible target anyway. In practice that means a named owner for every handoff, a queue someone actually watches, and a specific callback time to offer when nobody is free.
Frequently asked questions
Is a high handoff rate a bad sign?
Not on its own. Track the handoff rate together with the reason for each handoff. A rising share of "no source" handoffs points to gaps in your documents, which you can fix. If you push the rate down by making people harder to reach, customers notice.
What if nobody is available to take the handoff?
The agent should say so plainly, offer a specific callback time or the next opening, and record the promise as a task with an owner. A vague "someone will get back to you" with no owner is how requests get dropped.
Next steps
Write your six triggers down before you launch an agent, along with who owns each one. An AI-to-human handoff only works if someone is on the other end. In Aurora, your team can take over any conversation with the full history, requests that cross a limit go to the right person for approval, and every handoff is on record.
See how the full flow works in AI customer service agents: how they work end to end, explore the Aurora platform, or read how Aurora keeps you in control.