A good handoff passes a transcript, a sentiment read, and whatever the caller already gave (name, account, reason for calling) to the human before the caller hears a new voice. A bad one drops the call into a queue with nothing attached, and the first thing the human says is "sorry, can you start over." The difference is a design choice made before launch, not something you can bolt on after callers complain.
What actually happens during a voice agent handoff?
There are three patterns, and most vendors only build one of them well.
A cold transfer ends the AI's part of the call and dials the human directly, sometimes with a written summary attached to the ticket, but no verbal briefing. It is fast and cheap, and it's fine for a simple routing decision like "billing versus support," where nothing complicated happened yet.
A private warm transfer puts the human on a brief hold, plays them a spoken or text summary of the call, sentiment included, then merges the caller in once the human has the context. This is the pattern that actually stops a caller from repeating themselves, and it's what people mean when they say a handoff was "smooth."
A conferenced warm transfer keeps the AI on the line as a third participant after the human joins, which is useful when the AI already has data pulled up (an order status, a calendar) that the human would otherwise have to look up live.
| Transfer type | What the human hears before joining | Best for |
|---|---|---|
| Cold | Nothing, or a written ticket only | Simple routing, low call volume |
| Private warm | A spoken or text summary and sentiment read | Most business calls |
| Conferenced | Nothing extra, but the AI stays on the line | Calls where the AI already pulled data |
What should trigger a handoff to a human?
Four things, and a well-built agent watches for all of them at once rather than waiting for the caller to say "agent."
An explicit request is the obvious one. A request that falls outside what the agent is built to do, a policy question it has no data for, a competitor comparison, is the second. The third is a confidence signal from the agent itself: it doesn't recognize the intent, or it's about to guess. The fourth is sentiment, specifically rising volume, repetition, or a caller correcting the same detail twice. Waiting for a caller to lose patience and ask for a person is the weakest possible trigger, because by then the call has already gone badly.
What information has to travel with the call?
Five things, and missing any one of them is what makes a handoff feel like starting over: the full conversation transcript up to that point, structured data the caller already gave (name, account or order number, reason for calling), a short AI-generated summary of intent and sentiment, a note on what the AI already tried, and any authentication the caller already completed. A caller who gave their account number to the AI thirty seconds earlier should never be asked for it again by the person who picks up.
Voice adds a wrinkle text-based handoffs don't have: dead air. A caller on a text chat doesn't notice a two-second gap while a summary loads. A caller on a phone call does, and it reads as the call dropping. The fix is a short hold message, not silence, timed to the actual transfer latency your setup produces, not a generic "please hold."
How much does a warm transfer actually cost?
More than a cold one, but usually not by much, and the difference shows up in the meter, not the build. Per-minute AI voice agent pricing runs roughly $0.05 to $1.00 a minute depending on the platform, and warm transfers add a separate line item on some platforms. Bland charges $0.03 to $0.04 per minute during a transfer on its own phone numbers, and nothing extra if the transfer routes through your own Twilio account, which is the kind of detail that doesn't show up until you ask a vendor directly which meter keeps running once the human joins.
That's the real cost question to ask before signing anything: does the AI fee stop the moment the human picks up, does telephony billing continue, and does the platform bill both legs of the call simultaneously during the handoff window. Vendors rarely publish this, and it can double the effective cost of every escalated call if you don't ask.
How many calls should even reach a handoff?
Fewer than you'd think, if the agent is built for your actual call types rather than a generic script. Containment rate, the share of calls the AI resolves without a human, is the metric to watch. Basic voice bots running fixed scripts typically land around 10 to 20 percent containment, while agents built against real intent recognition and your actual data can reach 60 to 90 percent on repetitive call types like appointment scheduling or order status, according to the same vendor comparison. Your number will land somewhere between those depending on how many of your calls are genuinely simple versus how many need judgment a machine shouldn't be making.
A low containment rate isn't automatically a failure. A dental office where every third call is "can you check if my insurance covers this" should expect more handoffs than a pizza shop taking orders. The mistake is not tracking the number at all, because without it you can't tell whether a bad month is call volume or a broken handoff pushing calls to humans that the agent should have resolved.
How do you test a handoff before it goes live?
Call your own number and get transferred on purpose. Have the human side of the transfer confirm, out loud, what context they received: did they hear a summary, do they know why you're calling, did they ask you for anything you already said. Then do it again on a bad connection, with background noise, and with a caller who's clearly frustrated from the first sentence, since sentiment-triggered handoffs behave differently than ones triggered by an explicit request. Agent inefficiency during a handoff traces directly back to context that didn't transfer: 83 percent of contact center leaders in a 2024 CCW Digital Market Study said their agents spend too much time on interactions that should have been simple, and a broken handoff is one of the most common reasons a simple call turns into a long one.
Getting the handoff right the first time
Suvysoft builds voice agents with the handoff designed in from the start, not patched on after launch: which trigger fires a transfer, what data travels with it, and a warm transfer as the default rather than a cold one that makes a caller repeat their name and reason for calling to a second person in the same call. It's part of a broader agentic AI setup built against real call recordings from your business. Recent builds are in our case studies. If you're evaluating a voice agent, or your current one hands calls off badly, get in touch and we'll walk through what a real transfer from your business should sound like.
Frequently asked questions
What's the difference between a cold transfer and a warm transfer?
A cold transfer connects the caller to a human with no verbal briefing, sometimes with a written note attached to a ticket the human has to read separately. A warm transfer briefs the human first, with a spoken or text summary and a sentiment read, before merging the caller in. Warm transfers take a few extra seconds but are what actually stops a caller from repeating themselves.
What should trigger an AI voice agent to hand off to a human?
Four things: an explicit request from the caller, a question that falls outside what the agent is built to answer, the agent's own low confidence in the intent it detected, and sentiment signals like rising volume or repeated corrections. Waiting only for the caller to ask for a person is the weakest trigger, since by then the call has usually already gone badly.
Does a warm transfer cost more than a cold one?
Usually, but not by much on most platforms, and the bigger cost question is which meter keeps running once the human joins. Some vendors charge a separate per-minute rate during the transfer window, roughly $0.03 to $0.04 a minute in one published example, while others stop the AI fee entirely once a human picks up. Ask a vendor directly before signing.
How many of my calls should actually need a human?
Depends heavily on what the calls are about. Basic scripted voice bots typically resolve 10 to 20 percent of calls without a human. Agents built against real intent recognition and your actual business data can resolve 60 to 90 percent of repetitive call types, like scheduling or order status. A business with mostly simple, repeatable calls should expect a high containment rate; one where every call needs judgment should expect more handoffs, and that's not a failure.
Can the AI stay on the call after a human joins?
Yes, that's a conferenced warm transfer, and it's useful when the AI already has data pulled up, like an order or a calendar slot, that would otherwise make the human look the same thing up live while the caller waits. Most platforms support it, but it has to be configured on purpose; it isn't the default behavior.
How do I know if my current voice agent's handoff is actually working?
Call it yourself and get transferred on purpose, then ask whoever picks up what context they received. If they ask you anything you already told the AI, name, account number, reason for calling, the handoff is dropping data. Test it again on a bad connection and with a deliberately frustrated tone, since sentiment-triggered transfers often behave differently than ones triggered by a direct request.