Does transfer QA grade the AI agent or the human?

The human leg. Reason codes, a resolution note, and a coaching note all describe what your person did after the handoff, not what the AI did before it. It is built to coach a colleague, not to build a case against one.

What is the 86% interruption number, and how was it measured?

It is the interruption-detection precision measured in production: how often the agent correctly reads that a caller meant to interrupt, rather than just registering a stray sound. It is a real, in-code metric, not a lab benchmark.

Can I get alerted the moment a call goes badly?

Write a plain-language flag rule and a severity, and every call is checked against it by a model reading the transcript. A call that trips the rule fires a webhook you can route anywhere your team already watches.

Where do I actually look at all of this?

The Calls Workbench. Filter by outcome, reason code, cost, or duration, open a call's Pod to see the transcript next to the tool trace and the transfer card, and build a playlist out of the calls worth reviewing with the team.

Does the recording keep going after a transfer?

Yes. The handoff bridges the human into the same room the AI was in, so the recording and a live listen both continue past the transfer. That is also what the transfer QA review is reading.

Call analytics and transfer QA

Call center analytics for the whole call, not just the AI's half.

Every call closes into one session report: what it cost, which provider served it, how long the queue held, and every tool the agent ran. When a call transfers, the human leg gets reviewed too, with a coaching note attached, not a compliance flag.

Cost, provider, and queue timing on every call
A full tool trace, not a partial one
Transfer QA on the human leg
Coaching notes for your team, not a scorecard on them

Book a demo See a call resolve

Broken out by component: Cost
Whichever one served the call: Provider
Reviewed, not just logged: Human leg

The session report

The call hangs up. The record keeps writing.

Not just call logs. Case outcomes. Every conversation is auto-scored Resolved or Unresolved, a typed, filterable signal, not a read taken off a transcript. Sentiment, escalation, and friction get the same treatment. Cost, provider, and queue timing fill in as the call closes, the tool trace resolves into checkmarks, and the transfer QA block completes once a supervisor reviews the human leg.

Session report

Call 4f21 · inbound

Escalated

Her pickup window closed before the driver arrived. She needs today's ride rebooked and the new time read back before she hangs up.

Outcome: Resolved
Sentiment: Positive
Escalation: Yes
Friction: Low

Cost: LLM, speech-to-text, text-to-speech, and telephony, broken out by component
Provider: openai · deepgram · elevenlabs, logged per component
Queue timing: Caller-joined, agent-first-audio, and hold time, timestamped to the call clock

verify_member
check_eligibility
reschedule_ride
warm_transfer

Transfer QA · human leg

Reason codeagent_could_not_resolve

ResolutionClean handoff, resolved

Coaching note

Confirmed the same-day override and read the new pickup time back before ending the call, so she did not have to call again to check.

Generated from the call itself. Nothing on this sheet was typed in by hand.

See what happened. Prove it was solved. None of this is retyped from a spreadsheet. It is what the call actually did.

What gets lost today

The call ends and the truth scatters

Cost sits in one invoice. The transcript sits in another tool. Nobody can say what actually happened after the transfer, because nobody was still watching.

Today

Where the record breaks apart

Cost lives in a billing export, not next to the call
The transcript is in one tool, the recording in another
A transfer is the last thing anyone sees of the call
QA time goes to a sample of calls, never the escalated ones

With Zumu

One record, from ring to resolution

Cost, provider, and queue timing sit on the same record as the transcript
The recording keeps running after the transfer, not just the AI leg
Every escalated call gets reviewed on the human leg, not skipped
A coaching note comes back with the review, not a spreadsheet cell

The call is not over when the AI stops talking. Neither is the record.

How it works

Four jobs, one call record

Each one traces to a real surface in the product. None of them are a marketing gloss over a spreadsheet export.

Session reportThe daily surface
Every call closes into one record
Cost, provider, queue timing, the full tool trace, and a typed Resolved or Unresolved outcome all land on the same call record as the transcript. Nothing about what happened on a call lives in a separate export you have to reconcile by hand.
- Every call closes Resolved or Unresolved, a typed signal filterable across the whole call log, not a read taken off a transcript
- Cost broken out by component: LLM, speech-to-text, text-to-speech, and telephony
- The provider that actually served the call, not just the one configured for it
- Queue timing: when the caller joined, when the agent's first audio played, and how long the hold ran
Transfer QA
The human leg gets reviewed too
Most call scoring stops at the bot. Every escalated call here gets a second review, on the part your own person handled: why it escalated, how it resolved, and what would have made the handoff cleaner.
- A reason code on why the call escalated in the first place
- A resolution note on how the human leg played out
- A coaching note for the person, not a verdict filed against them
The Workbench
One screen a supervisor actually opens
The Calls Workbench is where a call's full record lives: transcript, tools, cost, transfer card, and QA card in one Pod. Filters run on outcome, reason code, cost, and duration, so a supervisor can find the calls worth reviewing instead of scrolling past all of them.
- Exact footer totals on every filtered view, matched to the underlying numbers
- A recording player with moment pins, so a coaching conversation can start at the right second
- Playlists for building a training set out of the calls worth revisiting
Flag rules
Tell it what a bad call looks like
Write a plain-language rule and a severity, and every call gets checked against it automatically. A call that trips a rule fires a webhook instead of waiting for someone to notice it in a review queue.
- Rules are graded by a model reading the call, not a keyword match
- Severities from a note to a page: info, warn, and critical
- The same webhook system that reports on tool calls and handoffs reports on this
- A flagged call can post the recap, transcript link, and recording straight into your team's Slack or Telegram channel

Measured, not modeled

86%

Interruption-detection precision, measured in production. The agent reads whether the caller actually meant to cut in, not just that they made a sound.

What backs the number: Every session report lists the crash and struggle signals feeding the model that scores it, not just the score.
What you can check: Recording and live listen both survive a transfer, so a reviewer can hear the human leg the same report is scoring.

An analyst reviewing a stack of call reports at her desk, cross-referencing cost and transcript

The report used to take three tools to build. Now it is one page.

Where this connects

The handoff has its own page. So does the bill.

This page covers what the record shows once a call closes. The warm transfer itself, the hold music, the briefing, the human accept or decline, lives on its own page. The honest metering that keeps a transferred call from billing twice lives on pricing.

See the warm transfer See how metering works

Questions

What operators ask about the record

Still have questions?

Bring a call that escalated. We will show you the report and the QA note it produces.

Book a demo Talk to sales

Every call gets a full session report: a diarized transcript, the tool-call log with what ran and in what order, a cost breakdown by component, which provider actually served the call, and queue timing from the moment the caller joined to the moment the agent's first audio played.

Book a demo

Bring a call. Leave with its report.

Bring a recording of a call that escalated, good or bad. We will show you the session report it would have produced and the transfer QA note that would have gone to the person who took it.