Warm transfers, explained: SIP REFER vs. bridged conferences

“Transfer me to a human.” Five words that sound trivial, but expose every shortcut a voice platform took. Receptic ships two transfer modes — SIP REFER and bridged conference — because there is no single right answer.

SIP REFER: the elegant handoff

SIP REFER tells the originating PBX to re-route the call directly to a third party. The agent steps out, the carrier handles the new leg, and the original call terminates cleanly. Lowest cost, lowest media path, no ongoing leg to pay for.

• Pros: One audio path. No ongoing minutes on our side. No conference bridge to manage.
• Cons: The receiving party must accept blind transfers. The agent can't whisper context — there's no overlap window.

Bridged conference: the warm handoff

The agent stays on the line, dials the human, plays hold music to the caller, briefs the human in a private leg, then merges the two legs into a three-way conference. The agent drops out once the human acknowledges, leaving caller and human talking.

• Pros: Human gets full context before the customer hears them. Caller never hears dead air.
• Cons: More complex media routing. Pays for the duplicate leg until handoff completes.

How we choose

The agent picks based on three signals: the destination's configured preference, whether the customer flagged the call as high-context (PI intake, billing dispute), and whether the receiving number has been set up to handle blind transfers.

function pickTransferMode(call, destination) {
  if (destination.prefersBridged) return "bridged";
  if (call.context.requiresBriefing) return "bridged";
  if (destination.acceptsBlind) return "refer";
  return "bridged"; // safe default
}

The hold music problem

A 4-second silence during a transfer feels like a dropped call. We loop a soft music bed — different per customer brand — under a brief spoken cue: “One second while I get someone on the line.” Latency to the bridge is 600-1200ms. The cue covers all of it.

Whisper briefing

On a bridged transfer, the agent speaks a 2-sentence summary into the human-only leg before merging. Format:

“You're getting Sarah Lee, calling about a leak under her sink. She tried the shutoff valve already.”

We render this from the structured call state — caller name, stated reason, attempted self-help, urgency signal — so the human picks up already in the conversation.

Failure modes

• Human doesn't answer — agent returns to caller, offers callback or voicemail.
• Human answers but rejects the call — agent recovers gracefully and routes to fallback.
• Carrier rejects REFER — fall back to bridged automatically.

What we'd do differently

If we were starting over, we'd build the bridged path first and only add SIP REFER as an optimization for high-volume customers. The cost savings of REFER are real, but the experience cliff when it fails is worse than the fixed cost of always bridging.