Hand over, promptly and without an argument. The agent confirms, passes the conversation on with a note so the customer does not repeat themselves, and gives a real wait time. If nobody is available, it books a callback. It may offer once to help in the meantime, but it must not stall, persuade or make the route harder.
Why not try to keep the customer with the agent?
Because a customer who asks for a person has already told you something important: they do not believe the agent will solve their problem, or they need something only a person can give, such as judgement, empathy or authority. Trying to talk them out of it confirms their doubt. Every extra step between the request and a person feels like an obstacle, and customers remember obstacles longer than they remember answers.
There is a measurement trap here too. Teams rewarded for keeping conversations away from people will, over time, make the human route harder to find. That improves one number and damages trust, complaints and, in regulated firms, conduct.
What should the agent do, step by step?
- Acknowledge at once. "Of course, I'll get you to someone."
- Offer once, briefly, to help while they wait. "While I connect you, is there anything I can check for you?" If the customer says no, stop.
- Route to the right person. Use the contact reason so the customer reaches the team that can help, not a general queue.
- Pass a note. Who the customer is, how they were verified, what they asked, what has been done.
- Give a real wait time. From the queue system, not a guess.
- Fall back to a callback if nobody is free, with a time the customer can rely on.
- Record the request and its reason, if the customer gave one.
What must the agent never do?
- Ask the customer to explain why they want a person before connecting them.
- Repeat the offer to help after the customer declines.
- Say a person "won't be able to do anything more", which is on the list of things an AI support agent should never say to a customer.
- Close the conversation if no one is available.
- Hide the request in a long menu of options.
Are there cases where the agent should hand over even without being asked?
Yes, and they matter as much as the explicit request. Distress, vulnerability, complaints, legal threats and fraud in progress all need a person. A customer who never types the word "human" may still need one most.
The difference between a person being available on request and a person actively overseeing what the agent does is worth knowing. The terms are explained in what human-on-the-loop means and what human oversight is. Under the EU AI Act, human oversight has a specific meaning for certain systems, distinct from simply being able to reach a person; the comparison is in human-in-the-loop vs human oversight under the EU AI Act.
An illustrative example
Take a telecoms provider's chat agent. A customer's first message is: "I want to speak to a person about my bill."
The agent replies: "Of course. I'll connect you to our billing team now. The current wait is about ten minutes. While you wait, would you like me to pull up your latest bill so it's ready for them?" The customer says yes. The agent verifies the customer, opens the bill and adds it to the note with the request. The billing specialist joins the chat with the bill already on screen.
A second customer, at 11pm, asks for a person. The billing team is closed. The agent says so, offers a callback the next morning between 9 and 11, and confirms the time in writing. It does not suggest the customer tries the agent instead.
What controls make it reliable?
- A trigger list of phrases meaning "I want a person", in every supported language, including indirect ones like "this isn't working" and "let me talk to someone".
- A single offer rule, so the agent never asks twice.
- Queue-backed wait times, never estimated by the model.
- Callback fallback when no team is open.
- Measured: time from request to a person, and whether the person needed to ask the customer to repeat anything.
What this means for staffing is the subject of whether an AI support agent will replace human support staff. The short answer is that you need people ready to answer.
Where Pryme fits
In Pryme, a request for a person is a platform guardrail: it holds the conversation for a person whatever the agent's depth or the customer's settings. The verdict and the handover note are filed together. More is in the AI support agent that knows when to say no.
Questions people ask
What if the customer asks for a person for something trivial?
Hand over anyway. You can study patterns later to improve the agent, but the individual customer gets what they asked for.
Should the agent tell the customer how many people are ahead of them?
A wait time is more useful than a position. Give whichever your queue system reports accurately.
Can the agent ask the customer to try a self-service link first?
It can mention it once alongside the handover, not instead of it.