Skip to content

Bananaflow MCP is live.

Read the docs

Call QA

Call QA that reviews the calls your agent takes

When a call ends, Bananaflow reads the transcript step by step and returns a 1 to 10 score, tags that quote the moment, the caller's sentiment and a short summary. Use our checks or write your own scorecard.

Isometric diagram of a call review in white lines over a soft, out-of-focus carnation in muted aqua, with sky blue and mint petals dissolving into fine veils around a calm teal centre. A finished call, drawn as a long glass bar with its sound wave along the side and labelled finished call, runs on a bold arrow into one lit glass cube marked with a check: the review Bananaflow runs once the call ends. A pill beside the cube carries its result, score 8 of 10.

What is AI call quality assurance?

AI call quality assurance is the automatic review of a call's transcript by a language model after the call ends, against a set of checks, to score it and flag what went wrong.

Bananaflow's Call QA does it inside the agent itself, so there is no second tool to connect.

How does Call QA work?

You add a QA node to your agent's workflow, publish, and each eligible call is reviewed after it ends. Setup takes five steps.

  1. Add a QA node

    Drop it anywhere on the workflow canvas. It needs no connections, because it never takes part in the live call.

  2. Keep our checks or write your own

    Use the default checks, or write your scorecard in plain words and name the fields you want back.

  3. Choose which calls

    Set a minimum call length, switch voicemail calls on or off, and set how large a share of calls to review.

  4. Publish the agent

    Each call is reviewed after it ends, one step of the workflow at a time.

  5. Read the result

    Open it on the call's page, or receive it in your webhook and through the API.

Have a QA form your team already uses?

What does Call QA check on each call?

By default it scores the call from 1 to 10, flags ten common problems such as dead air or a looping agent, rates the caller's sentiment and writes a one or two sentence summary. Four things make that useful.

  • A score for each step of the call

    See which step went wrong, not only which call.

    Greeting 9, Policy lookup 4, Renewal offer 8.

    • Each workflow step is reviewed on its own, with what that step is for and what was said before it.
    • Fix the step that scored 4 instead of rewriting the whole agent.
    • When steps cannot be told apart, Call QA reviews the whole call at once.
  • Problems flagged, with the line that proves it

    Every tag comes with a short reason that quotes the transcript.

    Dead air: 9 seconds of silence after "let me check your policy" at 41s.

    • Ten default tags: dead air, a looping agent, a frustrated caller, hearing problems, a caller who spots the AI, and more.
    • The reviewer sees a timestamp on every line and the response times, so a long silence shows.
    • Read the quoted line and decide in seconds whether the tag is fair.
  • Your QA form, written in plain words

    Replace our checks with the ones your team already scores.

    identity_verified: true, options_presented: false, tone_score: 4

    • Write your checks as instructions, such as "did the agent confirm the caller's identity first?".
    • Name the exact fields you want back: a yes or no, a 1 to 5 tone score, a short note.
    • Add a second QA node for a separate review, such as compliance.
  • You choose which calls and which model

    Review all eligible calls, or only a share of them.

    • By default it reviews every call except those under 15 seconds and voicemail calls. You can change both.
    • Lower the sample rate from its starting point of 100% to review only a share.
    • Use the agent's own language model, or give QA its own provider, model and API key.

Call QA reads the transcript, with its timestamps and response times. It does not listen to the audio, so it does not judge tone of voice or sound quality.

How does a call center use Call QA?

A QA lead stops sampling a few calls a day and reviews the lowest scores instead. This example is made up: the names and numbers are illustrative, not a customer.

Diagram in white lines over a soft, out-of-focus calla lily in muted sky blue, with a blush pink glow along the left edge. A list of all reviewed calls feeds a Bananaflow Call QA review, which sorts them into three results: the lowest scores first, a step with dead air, and the fix published as version 8.
  1. 9:00 am

    The scorecard goes in

    A call center in Pune runs a motor insurance renewal campaign: 2,400 calls a day. QA lead Priya Kulkarni and two analysts used to hear about 60 calls a day, roughly 2.5%. Developer Arjun Mehta adds a QA node with the client's scorecard: identity checked before policy details, renewal price read out, tone from 1 to 5. A webhook sends each result and the recording link to the team's own sheet.

  2. End of day 1

    2,180 calls are reviewed

    The rest were voicemail or under 15 seconds. Priya sorts her sheet by score. 37 calls score 4 or below, and 31 of them carry a dead air tag on the policy lookup step.

  3. Day 2

    Three calls, one cause

    She opens three of them on the call page. The agent goes quiet for 8 to 10 seconds while the policy system answers. Arjun adds a short line to say before the lookup, tests it in the browser, publishes version 8 and checks it against version 7 side by side.

Why review all of them? Google Cloud's documentation says manual review often covers less than 1% of call volume. A 2016 Call Centre Helper poll found most centers evaluated one percent of interactions or less.

Support teams use the same review to check the calls a voice agent takes. See how support teams check each call.

Where do Call QA results go?

Into the call's page next to the recording and transcript, and out to your own tools through a webhook or the API. Three outputs sit beside each other in the call record.

  • ScoreTagsSummaryYour table

    A webhook with the QA fields

    Call QA runs before your webhook is sent, so the payload can carry the score, tags and summary with the rest of the call data. Your sheet, CRM or data tool receives them as the call ends.

  • Callback requestedRenewal dateQuoted priceAgreed to pay

    Outcome codes and extracted variables

    Each call also ends with one outcome code from your own list, and each step can pull named values, such as a renewal date, from the conversation. They are separate from QA and sit in the same call record.

  • AgentCallerQA results

    The call's page

    Open any call to see the recording on a timeline beside the transcript, with the QA results in the same view. Read the quoted line, then play that moment.

To see the numbers across all your calls, open analytics and reports.

FAQ

Call QA questions, answered

  • What is AI call quality assurance?

    AI call quality assurance is the automatic review of a call's transcript by a language model after the call ends, against a set of checks, to score it and flag what went wrong. Bananaflow's Call QA returns a 1 to 10 score, tags with evidence, the caller's sentiment and a summary, or the fields of your own scorecard.

  • Does Call QA review every call?

    By default it reviews every call except those under 15 seconds and voicemail calls. The sample rate starts at 100%, and you can change the minimum length, switch voicemail calls on, or lower the sample rate to review only a share.

  • Can we use our own QA scorecard?

    Yes. Replace the default checks with your own instructions and list the fields you want back, and each call returns those fields. You can add a second QA node for a separate review, such as compliance.

  • Does Call QA change what the agent says during a call?

    No. It runs only after the call ends and reads the finished transcript, so it measures calls without touching them.

  • Does Call QA listen to the audio?

    No. It reads the transcript, with a timestamp on every line and the response times for each step. It does not judge tone of voice or sound quality.

  • Where do we see the results?

    On each call's page, beside the recording and transcript, and in your own tools through a webhook or the API.

  • Which AI model reviews the calls?

    The agent's own language model by default, or a provider, model and API key you choose for QA alone.

  • Does Call QA review calls handled by our human agents?

    No. It reviews the calls your Bananaflow voice agents take, not calls handled by your own human agents.

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot