Designing the Handoff: Where AI Agents Should Give Up

Most effort in deploying an AI agent goes into expanding what it can handle. The far more important design question is the opposite one: how does it know when to stop, and what happens then?

An agent that resolves 60% of contacts cleanly and hands over the rest with full context is better than one that resolves 80% and leaves the other 20% frustrated and repeating themselves. The second number looks better and performs worse.

The three failure patterns

Persisting past comprehension. The agent does not understand, asks the person to rephrase, still does not understand, asks again. Each loop increases frustration and the chance of abandonment. Two failed attempts is the practical limit.

Handing off with nothing. The agent transfers and the human starts from zero. The customer repeats their name, their account, their problem. From their point of view the first interaction was worse than useless — it cost them two minutes and produced nothing.

Confidently answering the wrong question. The most damaging. The agent matches on surface similarity, answers something adjacent, and the person acts on it. A wrong answer delivered confidently is worse than no answer, because it is acted upon.

Design the handoff triggers explicitly

Do not leave this to a confidence threshold alone. Enumerate the conditions:

Low confidence, twice. Two consecutive low-confidence interpretations. Not one — people phrase things oddly and recover.

Explicit request. "Speak to someone", "agent", "human", or pressing zero. Immediate, no persuasion, no "I can help with that". Attempting to retain someone who has asked for a person is the fastest way to make a complaint out of a query.

Emotional signal. Frustration or distress detected in wording or tone. Hand over early. The cost of an unnecessary transfer is small; the cost of making an upset person argue with software is not.

Out-of-scope intent. The agent recognises what is being asked and knows it does not handle it. This is a success — correctly identifying an unsupported intent is as valuable as handling a supported one, and should be measured that way.

High-stakes topics. Complaints, cancellations, anything involving money moving, anything with safety or medical implications. Route these to a person regardless of confidence. The agent may be capable; the exposure is not worth it.

Repeat contact. Someone contacting again about the same issue within a short window has already failed to get resolution. Do not put them through the same automation.

What must travel with the handoff

The minimum, and it is non-negotiable:

• the full transcript or summary of what was said
• any identity already established — do not ask for the account number twice
• what the agent believes the intent was, and its confidence
• what it already attempted or told them
• why it handed over

The last item matters for tuning. A handoff log with reasons is the most useful tuning dataset you have: it tells you precisely where the agent's coverage ends, ranked by frequency.

If your platform cannot pass context to the human, that is a significant limitation and worth knowing before you deploy, not after.

Set expectations in the agent's own words

The agent should say what it can do, early and specifically:

"I can help with appointment times, directions and account balances. For anything else I'll put you through to someone."

That one sentence reduces failed interactions substantially, because people ask for what is available rather than discovering the boundary by hitting it. It also sets up the handoff as the designed behaviour it is, rather than a failure.

Measuring it properly

Containment rate alone is a trap. It rewards holding onto contacts, which is exactly the behaviour that produces bad experiences. Always pair it:

• Containment rate — resolved without a human
• Post-handoff repeat rate — handed over and then needing another contact. This is your quality signal
• Handoff reasons, ranked — your tuning backlog
• Abandonment during AI handling — people giving up mid-interaction
• Time to handoff — long times before transfer mean the agent is persisting too far
• Satisfaction split by contained versus handed-off — if handed-off contacts score much lower, the handoff itself is the problem, not the agent's coverage

If containment rises while repeat contacts rise, you have made things worse with a better-looking dashboard.

The staffing question nobody asks

An agent that handles routine contacts changes the mix of what reaches humans. What is left is harder, longer and more emotionally demanding, because the easy contacts that provided variety have been removed.

That has real consequences: handle times rise (which looks like a regression and is not), and the role becomes more draining. Plan for it. Adjust targets, and recognise that "the simple ones" were also what made the job sustainable.

The principle

An AI agent is not trying to avoid human contact. It is trying to make sure that when a person is involved, they are involved in something that needs a person — with everything already known about the situation in front of them.

Judge the deployment on how well it does that, not on what fraction of contacts it kept away from your team.