When the agent needs a person, it says so

An agent that never asks for help is either doing very little or making things up. The useful question isn't whether an agent escalates. It's what happens next: does the request reach someone, does that person have what they need to answer, and does the answer get back to the customer?

In IDP, anything an agent can't finish lands in Needs a person, with the customer's question in their own words and the reason the agent stopped. From there your team has three ways to act, and the agent does the legwork for each one.

What lands in Needs a person

Each request shows who it's for, what they asked, and why the agent needs a person: a loaner only an advisor can approve, a price it isn't allowed to quote, a complaint that needs a manager. It also shows who owns the request, if the customer asked for someone, and everyone who was told about it.

A hang-up doesn't clear a request your team still owes, and a request stays in Needs a person when a campaign finishes with that customer. When a request is marked done, your team chooses whether the agent carries on with the customer or the conversation closes, and can leave a note.

Tell the agent the answer

Most requests need a sentence from someone who knows. So you can tell the agent the answer in your own words, even "yes, Thursday, loaner's reserved", and the agent writes the reply to the customer in its own voice and sends it. The request is marked answered once the reply is delivered.

Busy people don't always open the dashboard, so the owner can answer from their phone, by replying to the alert, or on a call from the agent itself. When only a person knows, the agent calls them covers both.

Give the agent a job for one customer

Sometimes the right move is for the agent to reach out: to rebook someone whose reschedule didn't go through, or tell a customer their part arrived. You type what you want done.

Before anything is sent, you see the plan: how and when the agent will reach the customer, what it may say and offer, what it won't do, and the exact first message. The agent keeps working until the job is done: a booking, a reply or a "no" from the customer closes it. If it can't finish, it stops after the time you chose and tells you how far it got. You get a notification either way.

Everything it sends respects the customer's contact choices and your quiet hours. If it can't reach the customer, the plan says why before you send.

Take over, then hand back

For the conversations that need a person all the way through, take over. The agent pauses while you reply, and knows that what you wrote was said by a person, not by itself.

Texts and emails the customer sends while you're in charge get an answer when you hand back, or when the takeover ends on its own. Campaign messages wait while you're in the conversation and continue afterwards.

Teach the agent, once

The best escalation is the one that doesn't happen twice. From any conversation, you can teach the agent what it should have known, a fact about your store or how to handle a kind of request, and it applies that from then on.

Why it matters

Escalation is where trust in an AI agent is won or lost. Customers forgive "let me check with the team" when the answer comes back. They don't forgive silence. Needs a person makes sure every request has a reason, an owner and an answer, and that the answer reaches the customer in the agent's voice, from your team's words.