When should AI hand over to a human? Handoff best practices

When your AI hands a conversation to a human, that moment makes or breaks your customer contact. A bot that keeps talking after the customer is already done with it burns more goodwill than a bot that lets go at the right time. This piece is about the practical side of that handoff. How you spot the signals that it is time, how you make sure your colleague does not start from zero, and where you set the threshold without burying your team under transferred conversations.

No theory about the perfect flow. Just the dials you actually turn when you set this up for your own customer contact.

Why most handoffs come too late

The instinct at a lot of companies: let the AI solve as much as it can on its own, because every handoff feels like a failure. That is the wrong lens. A handoff at the right moment is not a failure. It is the system doing exactly what it was built to do.

In practice you see something else happen. A bot tuned too aggressively gets stuck in a loop. The customer asks the same question three different ways, gets three almost-right answers, and gives up before a human ever comes into view. That customer does not call back with a complaint. They are simply gone. That is the expensive version of handing off too late, because you never even see it in your numbers.

The signals: when should AI hand off to a human?

There are roughly three kinds of moments where you want the AI to let go. They sometimes overlap, but when you are setting things up it helps to look at them one at a time.

1. The AI is unsure of itself

A good agent knows when it does not know. If the model scores low on confidence, or the question falls outside the knowledge you fed it, that should be a trigger. Do not guess and hope it lands well. Better to say it straight: let me pass you to a colleague who will sort this out for you. A wrong answer delivered with total confidence does more damage than a clean handoff.

2. Emotion from the customer

This is the most important trigger and the most underrated one at the same time. Anger, disappointment, someone raising the same problem for the second time: that is where a human adds value a bot cannot. Somebody who just received the wrong package for a birthday is not in the mood for a politely worded stock reply. Pick up on signals like swearing, a row of exclamation marks, or words like disappointed and ridiculous, and hand off before the frustration climbs any higher.

3. Complexity and risk

Some questions you just do not let a bot handle, no matter how good it is. Think of a disputed invoice, a legal question, a cancellation with financial consequences, or anything that deviates from the standard. These are the cases where a slip really costs you money or trust. Put a hard rule on those: certain topics always go to a human, no matter how confident the AI sounds.

• Uncertainty: the model scores low, or the answer falls outside the knowledge you provided.
• Emotion: frustration, the same complaint coming back, or clear dissatisfaction.
• Complexity: custom cases, exceptions, or anything that does not fit the script.
• Risk: money, contracts, complaints, or privacy-sensitive matters.
• Explicit request: the customer asks for a human. That is not up for debate, that goes straight through.

Passing on context: the handoff your colleague does not hate

A handoff without context is only half a handoff. You know it from being transferred at a call center. You tell your story, get connected, and have to do the whole thing over again. That is exactly what you want to avoid. When the AI hands off, your agent should see at a glance what it is about.

The bare minimum that has to come along:

1. A short summary of what the customer wants, in one or two sentences.
2. The full conversation so far, so nothing gets lost.
3. The details the AI already pulled up: order number, status, customer tier.
4. The reason for the handoff, so your colleague knows whether it is about emotion, complexity, or uncertainty.
5. The tone of the conversation. You approach an angry customer differently than someone with a simple question.

Systems that read live from your Shopify, CRM, or calendar have a head start here. The AI already looked up the order status before the handoff, so your agent starts by helping instead of searching. With Conveya that context travels along by default, so the customer does not have to repeat anything.

Setting the threshold: not too high, not too low

This is where the real trade-off sits. Set the threshold too low and you hand off everything, and you might as well not have used AI at all. Set it too high and the bot keeps floundering and you lose customers in the loop. There is no universally right setting, because it depends on your volume, your type of questions, and how many people you have to catch what comes through.

A workable approach: start cautiously and hand off one time too many rather than too few. In the first weeks you learn where the bot is strong and where it stumbles. Based on that you adjust the threshold per topic. You let the AI handle returns on its own once you see that running smoothly. Payment disputes you keep with a human for now.

Look at your staffing too. Outside office hours you might want to let the bot try more, because nobody is standing by anyway. During the day, with your team on hand, the threshold can go lower: a quick handoff is cheaper then than a customer who drops off. At a lot of web shops you see the share of conversations that end up going to a human land somewhere between twenty and forty percent, depending on how standard the questions are. Not a target to steer toward, but a sense of the order of magnitude.

What a good handoff gets you

Set this up well and the customer barely notices the switch. No repetition, no wait where you tell your story again, no bot stubbornly trying. Simple questions go through right away, and the cases that really matter reach someone who already has the context in front of them.

Start with a low threshold, watch for two weeks, and adjust per topic. The handoff is not the weak spot of your AI. Set up well, an AI handoff to a human actually shows that your system knows what it can and cannot handle.