When Should an AI Agent Escalate to a Human? Designing Handoff Rules

A good AI agent knows what it can handle. A great one also knows when to stop and hand the work to a person. Without that, an agent keeps going on a question it doesn't understand, apologizes in circles to an upset customer, or answers a legal threat with cheerful guesswork.

Escalation rules decide when the agent steps aside and how the person taking over picks up the thread. This guide covers how escalation differs from approval, the triggers that should hand work to a person, how to design the handoff itself, and how to tell whether your rules are working.

Escalation versus approval

Both put a person in the loop, but they answer different questions:

An agent usually needs both. Approval points control what the agent does; escalation rules control when it should stop trying.

When to escalate: the triggers

Most escalation rules fall into eight groups. Decide which apply to your agent, and make each one specific enough to test.

Write each trigger as something you can check: "after two replies without resolution", not "if the customer seems unhappy".

Designing the handoff itself

A bad handoff makes the customer start over. A good one feels like the same conversation continuing with someone more capable. Three parts make the difference:

1. A handoff note for the person taking over. The agent should pass along, in a few lines:

• what the customer wants, in one sentence;
• what the agent already checked, and what it found;
• the sources or records it used;
• why it escalated;
• anything it promised or proposed, so nobody contradicts it.

2. A clear message for the customer. Say that a person is taking over, what happens next, and roughly when. "I've passed this to our support team with the details; you won't need to repeat anything" beats a silent transfer.

3. No dead ends. Every escalation needs somewhere to go: a queue that someone watches, with an owner and a target response time. Outside working hours, say so honestly and give the customer a realistic expectation.

Once a person takes over, the agent should step back from that conversation unless the person brings it back in.

Where the rules live

Some triggers depend on judgment, like recognizing that a question is out of scope or that a customer is upset. Those belong in the system prompt, described clearly with examples.

Others shouldn't depend on the model noticing. Put these in the system around the agent, where they're enforced rather than requested:

• Amount and action thresholds: anything over a set value goes to a person.
• Loop limits: after a set number of turns or retries without resolution, hand over.
• Tool failures: when a required tool fails, escalate instead of guessing.
• Explicit requests: a "talk to a person" option the user can always reach, not only by asking the model.

The model can miss a trigger; a rule in your system won't.

Measure and tune

Escalation rules are never right on the first try. Track:

Missed escalations are the hardest to see, because nothing flags them. Reading a sample of conversations each week is often the only way to find them.

Common mistakes

• No way to reach a person. Hiding the human option frustrates people and doesn't reduce work; it moves it to complaints.
• Silent transfers. The customer is handed over with no explanation and has to start again.
• Escalating into a void. A queue nobody watches is worse than no escalation at all.
• Vague triggers. "If the customer seems unhappy" can't be tested; "after two replies without resolution" can.
• Relying on the model for hard limits. Thresholds and loop limits belong in your system.
• Never reviewing. Without sampling conversations, missed escalations stay invisible.

Checklist

• ☐ The user can always reach a person.
• ☐ Triggers are written down and specific enough to test.
• ☐ Hard limits (amounts, loops, tool failures) are enforced outside the prompt.
• ☐ Every handoff includes a short note: request, what was checked, sources, reason, any promises.
• ☐ The customer is told what happens next and roughly when.
• ☐ Every escalation lands in a queue with an owner and a response target.
• ☐ Escalation rates, reasons and time to a person are tracked.
• ☐ A sample of conversations is reviewed regularly for missed escalations.

Conclusion

Escalation is how an agent stays useful at the edge of what it can do. Define triggers you can test, enforce the hard ones in your system rather than the prompt, hand over with a note that saves the customer from repeating themselves, and make sure someone is waiting on the other side. Then measure both the escalations that happen and the ones that should have.

Explore LibraOS: libraos.com

Sources

• LibraOS Security & Private Deployment
• LibraOS homepage: Runtime

What’s Next?

Sign up and explore now.

🔍 Learn more: Visit our blog and documents for more insights or contact us for a demo of how MegaNova can power your AI products.

📬 Get in touch: Join our Discord community for help or Contact Us.