Yes. A voice agent that cannot transfer has no way to hand over, and an agent that cannot hand over should not be answering calls. The questions that matter are where it may transfer to, what context travels with the caller, what happens when nobody is free, and whether the agent may ever end a call instead.
Why is transfer a low-risk action?
Most actions an agent takes change something: a record, a booking, a balance. A transfer changes only who is talking to the customer. It is easy to undo, it moves the conversation towards a person, and it is the action every escalation rule depends on. On the depth scale it sits with routine, bounded work.
That makes transfer one of the first capabilities to allow, not the last. What it means for an action to be allowed, rather than held or refused, is set out in allowed: what it means for AI agents.
The risk is not in the transfer itself. It is in transferring badly: to the wrong team, without context, into a queue with no end, or as a way to get rid of a difficult caller.
What conditions decide how the agent may transfer?
Destinations are fixed. The agent can transfer only to queues or numbers on an approved list, each matched to contact reasons. It cannot dial a number the caller gives it. That closes off attempts to use your phone system to reach someone else.
Context goes with the caller. A warm transfer briefs the person before connecting, or puts the agent's note on their screen as the call arrives. The idea is explained in what a warm transfer is.
The caller is told. Where they are going, roughly how long the wait is, and that they will not need to repeat themselves.
There is always a destination. Out of hours, at capacity or during an outage, the agent offers a callback rather than a dead end.
What may the agent do, and what stays with a person?
| Situation | The agent may | Stays with a person |
|---|---|---|
| Caller asks for a person | Transfer at once to the right queue | The call |
| Request outside the agent's bounds | Transfer with a note and any draft | The decision |
| Caller distressed or vulnerable | Transfer on the priority route | Support |
| No one available | Book a callback; tell the caller when | The callback |
| Caller is abusive | Give the set warning; transfer or end the call by the written rule | Any block on future contact |
| Caller asks to be put through to a named employee | Transfer only through the approved directory | Exceptions |
| Caller gives a number to be connected to | Refuse | Nothing |
Ending calls deserves a note. An agent should not hang up on a caller because the conversation is hard. Your written abuse rule decides when a call may be ended, usually after a warning, and a person decides on anything further.
An illustrative example
Take a regional bank's phone agent. A verified caller asks about a payment they do not recognise, then says they think someone has been using their card.
The agent freezes the card within its bounds, tells the caller it has done so, and says it is transferring them to the fraud team, who will go through the transactions. The fraud team's queue shows a short wait. The agent puts the note on the receiving screen: caller verified by in-app approval; card frozen; three unrecognised payments listed; caller calm. It transfers. The fraud specialist greets the caller by name and starts with the payments.
Later that evening, a second caller with the same problem reaches the agent after the fraud team's live hours. The agent freezes the card and books a callback from the out-of-hours fraud line, which is staffed by a smaller on-call team. Knowing who is on call for agents and their escalations is covered in who is on call for AI agents.
What controls make it safe?
- An approved destination list, versioned, with each destination tied to contact reasons.
- No free dialling. The agent cannot connect callers to numbers they supply.
- Notes required. A transfer without a note is flagged.
- Callback fallback. If no destination is answering, the agent books a callback rather than leaving the caller in a loop.
- Transfer outcomes tracked. Did the person receiving it need to ask the caller to repeat themselves? Did the call reach the right team?
The step-by-step setup is in how to let an AI agent transfer calls to a person. The wider principle of controlling what each agent may do, action by action, is in how to control what an AI agent is allowed to do.
Where Pryme fits
In Pryme, transfer is a capability with its own bounds: the destination list is part of the customer's limits, and every transfer is filed with its reason and note. Freezing a card and transferring a call are certified separately, so one can be allowed while the other is still being proved. More is in the AI support agent that knows when to say no.
Questions people ask
Should the agent try to solve the problem before transferring?
Only if the caller has not asked for a person. If they have, transfer first.
Can the agent transfer to another AI agent?
Yes, if the destination agent is on the approved list and the caller is told. A person must remain reachable.
Should the agent stay on the line during a transfer?
On a warm transfer, it can stay until the person confirms they have the context, then drop off.