The clean handoff: when your AI agent brings in a human

Your agent handles 80% of messages without breaking a sweat. The problem is almost never that 80% — it's the other 20%, the moment a case gets too big for it and has to go to a human. Get that handoff wrong and it undoes all the goodwill the agent had built. The customer retells their whole story, nobody's sure whose job it is, and a two-minute reply ends up taking two days. Let's build it properly.

Why the handoff is where things break

Automating the easy path is easy: repeat questions, confirmations, opening hours. The hard part — the part that actually decides whether the system works — is the exit route the agent takes when something is beyond it. And that's exactly where most people get sloppy.

There are two classic ways to ruin it. The first is the black hole: the agent says "I'll pass you to a person," and that person either doesn't exist or never finds out. The message lands in an inbox nobody checks until Monday. The second is the parrot problem: by the time someone picks up the case, the customer has to explain all over again who they are, what they asked, and what they'd already been told. Both leave the same impression — that nobody cares.

There's a third, subtler one: ping-pong. The agent passes the case to the team, the team decides "the bot handles that" and bounces it back, and the customer sits in the middle watching it get volleyed around. A clean handoff avoids all three, and you don't need anything fancy to pull it off.

Four signals to let go of a case

An agent shouldn't try to win every conversation. It should know when to step back. These are the clear signals:

• Emotion. An angry or disappointed customer, or a complaint. A technically correct but cold reply does more damage here than "someone will call you right now."
• Money or commitments off-script. Applying your price list, fine; negotiating a discount, accepting a late return, or closing a big quote, no.
• A gap in its playbook. It's been asked something it wasn't briefed on. The right move isn't to invent an answer — it's to admit the gap and pass the case up.
• An explicit request. "I want to talk to a person." No debate. An agent that keeps trying to solve it itself after that only annoys people.

Add a quiet fifth: repetition. If a customer rephrases the same question twice and still isn't satisfied, the agent isn't getting it right — even if it "thinks" it is. Letting one failed attempt count as an escalation signal saves you a lot of burned customers.

Who takes what (so nothing floats)

The golden rule: every handoff has an owner with a name, never "the team." "The team" is no-man's-land; "Marta's got this" is a person who replies.

You don't need an org chart. Routing by case type is plenty:

1. Complaints and upset customers → you or the manager.
2. Big quotes or negotiation → whoever runs sales.
3. Specific technical questions → whoever's on the floor or closest to it.
4. Anything else that doesn't fit → a default owner, so nothing falls through.

And decide what happens after hours, because that's when the cracks show most. An honest agent doesn't promise the impossible: it says "there's no one here right now, but a person will reply by 10am tomorrow" — and then delivers. A clear expectation beats a fast, false answer every time.

A good handoff carries three things

When the agent lets go of a case, the handoff should always include the following.

Context, not just the case. Whoever picks it up needs a three-line summary: who the customer is, what they asked, and what they've already been told. If they start from scratch, you've turned your customer into a parrot and thrown away half the benefit of having an agent.

A deadline said out loud. "Marta will reply by 6pm" is a concrete promise someone has to be able to keep. With no deadline, the customer stares at their phone with no idea whether to wait ten minutes or two days.

A closed loop. When the person resolves it, it gets logged, and if needed the agent picks the thread back up for follow-up ("did that sort everything out?"). What can never happen is two parallel conversations with the customer unsure which one to trust.

A quick test for whether your handoff is well built: time how long an escalated case takes to get a human reply, and count how many times the customer had to repeat themselves. If you don't like either number, there's your work.

When NOT to build an elaborate handoff

A couple of honest exceptions, because not everything needs process.

If nearly every message ends in a handoff, the problem isn't the handoff — it's the agent's playbook. Before you build routes, teach it to handle those cases. The handoff is the escape hatch, not the main road; if it becomes the main road, you've automated very little.

If you work alone, don't build a four-branch routing tree. A single destination — you — and a decent queue where nothing gets lost is more than enough. And if we're talking three escalations a month, same thing: have the agent ping your phone and reply when you can. Building process for that little volume is spending time you won't get back.

Start with one rule

You don't need to redesign anything this week. Write down two things: the exact escalation line the agent uses when it lets go of a case ("a person will look at this and reply by 6pm") and the default owner for anything that doesn't fit a category. That alone plugs the two biggest holes: the silence and the no-man's-land.

That split — the agent on the front line, filtering and summarizing; your team stepping in exactly when they add value — is precisely what a good AI receptionist does: it answers what it can and escalates what matters, with context and a deadline attached. The rest is yours — you're the one who puts a face to it, after all.