Skip to content

Bananaflow MCP is live.

Read the docs

Call Audit in BPO: Parameters and Scorecard

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Updated
Cover for "Call Audit in BPO: Parameters and Scorecard": a white line diagram of one recorded call being audited: a sound wave cut by four short ticks into five stretches, labelled Opening, Listening, Resolution, Tone and Compliance, each with a ring under it holding its mark, a check, a half bar, a check, a cross and a check, and the wave running on into a lit node over a lit pill reading 85 of 100, drawn over a soft, out-of-focus lotus in muted rose, its broad petals blurred into veils, with a pale blush petal glow and a soft sky blue haze at the edges.

A call audit is a structured review of a recorded call against set parameters, ending in a score and written evidence. In a BPO the person who does it is a quality analyst (QA), sometimes called a call auditor. This guide shows the parameters they check, a full weighted scorecard, how many calls to audit, how to calibrate, and which Indian rules the card should test.

Here is why the details matter. Meena is a QA at a collections BPO in Jaipur. She scores 10 calls for an agent named Rohit, and he averages 72. Another QA scores the same 10 and gets 84. Rohit's bonus sits between those two numbers. Neither QA is lazy. They simply read the card differently, and nobody has measured how much a small sample can swing. Fixing that is the whole job of a good audit programme.

The Jaipur centre, its numbers and its people are an example we made up for this post. They are not customer data.

What is a call audit?

A call audit is the formal, scored review of one recorded call. The reviewer listens, scores each parameter, writes down the evidence, and passes the result to coaching and compliance reports. Call auditing is the same thing as a process.

How is a call audit different from call monitoring?

Call monitoring is the wider habit of listening in on calls, often live, and a call audit is the scored, documented part of it. A supervisor who sits beside an agent and gives a tip is monitoring, and so is a supervisor who uses call barging and whispering to listen to, coach or join a live call. A QA who scores the recording and files the evidence is auditing.

Which kinds of call audit are there?

There are four common types, and most centres mix them.

  • Random audits: calls drawn by chance, which tell you how the floor is doing overall.
  • Targeted audits: calls picked for a reason, such as complaints, very long or very short calls, transfers, or a low customer score.
  • Compliance audits: calls checked only for the rules, such as the recording notice and required disclosures.
  • Side-by-side audits: a QA listens live next to the agent, which is closer to monitoring.

What does a quality analyst in a BPO do?

A quality analyst audits calls, writes the evidence, coaches with team leads, runs calibration sessions and reports trends. "QA in BPO", "call quality analyst" and "call auditor" are names for the same working day.

A typical week looks like this:

  1. Pull the calls for the week: a random draw plus the targeted ones.
  2. Score each call and write one line of evidence per low mark, with the time in the recording.
  3. Meet each team lead to agree which two or three behaviours to coach.
  4. Join the monthly calibration session.
  5. Send the trend report: average score, top lost lines, fatal errors by type.

The analyst is not the agent's manager. The job is to make the standard clear and fair, so a score means the same thing whoever gives it. Our call center metrics guide shows the numbers that sit beside the audit score, and the guide to AHT explains why a faster call is not always a better one.

What parameters does a call audit check?

A call audit checks the behaviours that decide whether a call went well and stayed within the rules. There is no fixed list of "7 quality parameters" in BPO. Each team chooses its own, and the choice should come from what your customers and your rules need.

We group the parameters into five areas, with about twelve lines in all:

  • Opening and verification: greeting, name, purpose, and the identity check before any account detail is shared. Our guide to call center scripts has opening and closing lines that this part of the card tests.
  • Listening and accuracy: discovery questions, correct information, and clean handling of any hold or call transfer.
  • Resolution: the issue was solved or sent to the right place, and the caller knows the next step and its date.
  • Tone and empathy: courtesy, patience and calm, especially with upset callers. See how to handle angry customers on calls for what good looks like.
  • Compliance and records: the recording notice, required disclosures, and the call logged with the right outcome code.

What should a good scorecard measure?

It should measure what the caller experiences, not only what the agent did. SQM Group, a customer-experience research firm, warns that an agent "can receive a 90% QA score while the customer remains dissatisfied" when the card rewards process over outcome (SQM Group on quality assurance). Its suggested check is to compare QA scores with post-call customer satisfaction scores, look for a high score with an unhappy caller or the reverse, and then change the weights.

What does a sample call audit scorecard look like?

A sample card has five groups, twelve lines and 100 points, with a separate list of fatal errors that fails the call. Each line is scored yes (full points), partly (half) or no (zero).

  • Line: Greeting, name and purpose

    Area (area points)
    Opening and verification (15)
    Points
    5
    Yes, partly or no means
    Yes: all three, in the first ten seconds. Partly: one missing
  • Line: Identity check before account detail

    Area (area points)
    Opening and verification (15)
    Points
    10
    Yes, partly or no means
    Yes: done first. No: detail shared before the check
  • Line: Discovery: right questions, no interrupting

    Area (area points)
    Listening and accuracy (30)
    Points
    10
    Yes, partly or no means
    Partly: asked, but talked over the caller
  • Line: Information is correct

    Area (area points)
    Listening and accuracy (30)
    Points
    15
    Yes, partly or no means
    No: any wrong price, date or policy fact
  • Line: Hold and transfer handled well

    Area (area points)
    Listening and accuracy (30)
    Points
    5
    Yes, partly or no means
    Yes: asked first, short wait, summary given. Mark as yes if there was no hold
  • Line: Issue solved or correctly escalated

    Area (area points)
    Resolution (25)
    Points
    15
    Yes, partly or no means
    Partly: right team, but no reference given
  • Line: Next step stated, with a date

    Area (area points)
    Resolution (25)
    Points
    10
    Yes, partly or no means
    No: the call ends with no next step
  • Line: Empathy and courtesy

    Area (area points)
    Tone and empathy (15)
    Points
    10
    Yes, partly or no means
    Partly: polite but flat with an upset caller
  • Line: Clear speech, no jargon

    Area (area points)
    Tone and empathy (15)
    Points
    5
    Yes, partly or no means
    Partly: fast or full of internal terms
  • Line: Recording notice and required disclosures

    Area (area points)
    Compliance and records (15)
    Points
    8
    Yes, partly or no means
    No: missing where your rules need it
  • Line: [Outcome code](/blog/call-disposition-codes/) and notes correct

    Area (area points)
    Compliance and records (15)
    Points
    7
    Yes, partly or no means
    Partly: code right, notes thin
  • Line: Total

    Area (area points)
    Five areas
    Points
    100
    Yes, partly or no means
    Fatal errors are scored separately

The areas add up as 15 + 30 + 25 + 15 + 15 = 100.

Here is how a score is worked out. Say Meena audits one of Rohit's calls. He did everything well, except he talked over the caller once (discovery, partly: 5 of 10) and was flat with an upset caller (empathy, no: 0 of 10). The score is 100 − 5 − 10 = 85.

The sample scorecard: five weighted groups that sum to 100, with a separate box of fatal errors that fails the call whatever the score.

This card is a deliberately fuller audit version. The simpler card in our call center training manual is the one to use when you teach a new agent the standard. Use that for training and this one for audits.

What are fatal errors in a call audit?

A fatal error is a mistake so serious that the call fails whatever its score. "Fatal error" is real BPO language, and we recommend keeping the list short enough that every analyst can recite it. A starting list, which is our own advice:

  1. Sharing account details before the identity check.
  2. A promise of guaranteed returns or a false claim about a product.
  3. Discussing a loan or policy with someone other than the customer.
  4. Abusive, threatening or mocking language.
  5. Missing the recording notice where your rules require it.
  6. Data shared outside the approved process.

Score a fatal call as zero, and keep a separate count. A centre can have a good average and still have five fatal errors a week, which is a risk the average hides.

How many calls should you audit per agent?

Most centres audit a small share of their calls. In a Call Centre Helper poll of 115 centres, run in 2013, over two thirds monitored 0 to 6 calls per agent per month (Call Centre Helper poll). The split was 7% for 0 to 1, 34% for 2 to 4, 26% for 5 to 6, 19% for 7 to 10 and 14% for 11 or more. It is old and small, so treat it as a picture, not a rule.

A reader roundup on the same site says about 4 to 8 calls per agent per month is standard, with some teams aiming for 10 (Call Centre Helper readers). Those are opinions, not a study. SQM Group says manual QA typically covers less than 2% of calls.

Why does a small sample give a wide margin of error?

Because with few calls, one odd call moves the score a lot. The standard way to measure it is the normal approximation for a proportion, which the NIST/SEMATECH handbook sets out (NIST sample sizes for a proportion). At 95% confidence:

margin of error = 1.96 × √(p × (1 − p) ÷ n)

Here p is the pass rate (we use 80%, or 0.8) and n is the number of calls audited. The part under the root is 0.8 × 0.2 = 0.16. So the margin is 1.96 × √(0.16 ÷ n). The table shows the results. These are our calculations, and each number is rounded.

  • Calls audited: 5 for one agent

    Working
    1.96 × √(0.16 ÷ 5) = 1.96 × 0.179
    Margin on the pass rate
    ± 35 points
  • Calls audited: 10 for one agent

    Working
    1.96 × √(0.16 ÷ 10) = 1.96 × 0.126
    Margin on the pass rate
    ± 25 points
  • Calls audited: 20 for one agent

    Working
    1.96 × √(0.16 ÷ 20) = 1.96 × 0.089
    Margin on the pass rate
    ± 17.5 points
  • Calls audited: 50 for one agent

    Working
    1.96 × √(0.16 ÷ 50) = 1.96 × 0.057
    Margin on the pass rate
    ± 11 points
  • Calls audited: 500 across the floor

    Working
    1.96 × √(0.16 ÷ 500) = 1.96 × 0.018
    Margin on the pass rate
    ± 3.5 points

To find how many calls you need for a given margin, turn the formula around: n = 1.96² × p × (1 − p) ÷ margin². For a margin of ± 5 points (0.05) at an 80% pass rate, that is 3.84 × 0.16 ÷ 0.0025 = about 246 calls. If the pass rate is near 50%, the worst case, it is 3.84 × 0.25 ÷ 0.0025 = about 384. COPC, a quality standards body, offers a sample size calculator if you want to try your own numbers.

The meaning is simple. Five to ten calls a month is enough to coach an agent on what you heard, and a larger pool across the floor is enough to judge the team. It is too few to rank agents against each other or to act on a single score. The math also assumes the calls are drawn at random. A targeted sample answers a different question: it finds problems, but it does not estimate the average.

What does this look like in a real centre?

Take the Jaipur collections BPO again: 120 agents taking 30 calls each a day, over 22 working days, with two analysts.

  1. Calls a month: 120 × 30 × 22 = 79,200.
  2. Our assumption is that a 6-minute call takes about 15 minutes to audit (listen, score, write the evidence). That is 4 audits an hour.
  3. With 7 productive hours, one analyst does 28 a day, and the pair do 56.
  4. Over 22 days: 56 × 22 = 1,232 audits, or about 1,230.
  5. Share of calls: 1,232 ÷ 79,200 = 1.6%.
  6. Per agent: 1,232 ÷ 120 = about 10 calls a month.

Per agent, the margin is ± 25 points (the 10-call row above). For the floor as a whole, 1.96 × √(0.16 ÷ 1,232) is about ± 2.2 points, if the draw is random. So the Jaipur team can trust the floor number and cannot trust a ranking of agents.

Back to Meena and Rohit. Her 72 and the other analyst's 84 may both fall inside Rohit's wide band, and the gap between them points to a card that two people read differently. The fix is to widen the evidence, not to hold the score against him: add targeted calls, and calibrate.

What is the call audit process, step by step?

The process is a loop, and the loop only works if it closes. The steps below are our recommended order, based on the standard manual audit.

  1. Sample. Draw a random set of calls, then add targeted calls: complaints, long calls, short calls, transfers and low customer scores. Our guide to call recording software in India covers how to find the right call quickly.
  2. Score. Listen to the whole call. Score each line yes, partly or no. Mark fatal errors separately.
  3. Write the evidence. One line for each low mark, with the time in the recording. A score with no evidence cannot be coached.
  4. Calibrate. Compare scores across analysts, as below, and settle any disputed lines before the score reaches the agent. You coach on agreed scores.
  5. Coach. Give feedback within 48 hours of the audit. This is our recommendation, not a standard: a call from last month is a poor basis for a talk. Pick two or three behaviours, not twelve, and play the call back to the agent.
  6. Re-audit. Audit the same agent on the same lines to see whether the coaching worked, and track the fix. The re-audit feeds the next sample.
The audit loop: sample, score, calibrate, coach and re-audit, with the re-audit feeding the next sample.

What is call calibration, and how do you run it?

Calibration is scoring the same calls separately, then agreeing on why the scores differ, so every analyst reads the card the same way. Without it, you get what Call Centre Helper describes as "favoritism, actual or perceived" and, over time, people leaving (Call Centre Helper on calibration). The same page suggests holding sessions at the start and on an ongoing basis, and inviting agents as well as analysts and team leads.

Our routine for a monthly session:

  1. Choose 5 calls: one clean, one with a fatal error, one with an upset caller, one with a transfer and one that nobody expects to argue about.
  2. Each analyst scores all 5 alone, without talking.
  3. Put the scores side by side, line by line, not just the total.
  4. Find the lines where scores differ by more than one step (yes against no).
  5. Talk through those lines until you agree what the evidence is, then rewrite the line's definition so it is clear next time.
  6. Re-score once to check the new wording.

Our target is that analysts' totals land within 5 points of each other on the same call. That number is our recommendation, not a published standard. Pick your own, and tighten it as the card matures. The totals matter less than the lines: two analysts can reach 85 by different routes.

Add one more check. SQM Group suggests comparing QA scores against post-call customer satisfaction. If calls that score high have unhappy callers, the weights are wrong, not the callers.

Which Indian rules should a call audit check?

A call audit should turn each rule that applies to your calls into a line on the card, so the check happens on every audit. This is what to check, not legal advice. Each row links to our own guide, which carries the primary source and its date.

  • Rule: RBI: recovery calls must be recorded, from 1 January 2027, with the record kept six months or until a case closes

    What the audit checks
    The call is in the archive, the borrower was told it is recorded, and the loan was discussed only with the borrower
    Our guide
    [RBI rules for recovery agent calls](/blog/rbi-rules-for-recovery-agent-calls/#must-recovery-calls-be-recorded)
  • Rule: RBI: recovery calls between 08:00 and 19:00 unless the borrower asked otherwise

    What the audit checks
    The call time sits inside the window, or a recorded request is on file
    Our guide
    [Recovery call hours](/blog/rbi-rules-for-recovery-agent-calls/#what-hours-can-a-recovery-agent-call-a-borrower)
  • Rule: IRDAI: strict rules on mis-selling and disclosure in insurance sales calls

    What the audit checks
    No guaranteed-return promise, and the required disclosures were said
    Our guide
    [IRDAI rules for insurance sales calls](/blog/irdai-rules-for-insurance-sales-calls/)
  • Rule: Recording notice and consent

    What the audit checks
    A notice line at the start, in the caller's language
    Our guide
    [Is call recording legal in India](/blog/is-call-recording-legal-in-india/)
  • Rule: DPDP Act and Rules: keep personal data only as long as needed

    What the audit checks
    Recordings are kept and deleted by your retention rule
    Our guide
    [DPDP rules for call centers](/blog/dpdp-rules-for-call-centers/)
  • Rule: TRAI: promotional calls from the 140 series, service calls from 1600, and the DND list respected

    What the audit checks
    The right number series for the call type, and the list was scrubbed
    Our guide
    [140, 160 and 1600 series numbers](/blog/140-160-1600-series-numbers-india/) and [TRAI's new rules](/blog/trai-new-rules-for-business-calls/)

Two notes make this table easier to use. The RBI rows come from the RBI's notification, which our recovery-calls guide links to and explains (RBI notification). And the TRAI row is partly about the call list, not the call. An audit of one recording cannot show that the DND list was scrubbed, so audit the list as well. Our guide to cold calling in India also covers the number series and the DND basics.

Should you audit calls with AI or by hand?

Use both. Software can read every call, and people should read the calls that matter most. Here is how they compare.

  • Point of comparison: Coverage

    Manual audit
    A small share of calls, often under 2% according to SQM Group
    AI scoring from transcripts
    Every call, if you let it run on all of them
  • Point of comparison: Consistency

    Manual audit
    Varies between analysts until you calibrate
    AI scoring from transcripts
    Applies the same rules each time, though it still needs checking against your people
  • Point of comparison: What it can miss

    Manual audit
    Most of the calls
    AI scoring from transcripts
    Tone of voice and line noise, and it can mis-transcribe Hinglish
  • Point of comparison: Best use

    Manual audit
    Judgement calls, coaching and fatal errors
    AI scoring from transcripts
    A first pass, so people read the lowest scores

A sound setup has three parts. Software scores every call it can. People read the lowest scores and every fatal flag. And the analysts calibrate the software like another analyst, by scoring the same calls and comparing. We have not seen a published accuracy figure we would trust for AI scoring, so check it yourself on your own calls before you rely on it. The software layer behind that first pass is speech analytics.

How the QA team's work shifts when AI takes calls is in AI in call centers.

How Bananaflow fits

Bananaflow is an AI voice agent platform. Its Call QA reviews the calls its own AI voice agent takes, after each call ends.

What it does is score each finished call from 1 to 10 for every step of your workflow, tag the calls with lines quoted as evidence, and return caller sentiment and a short summary. You can write your own scorecard in plain words and name the fields you want back. A second QA step can run a separate review, such as a compliance check. By default it reviews every call except those under 15 seconds and voicemail, and you can lower the share. Results appear on the call page, go out by webhook and come back through the API.

Call QA reads the transcript with timestamps and response times, and people should still read the low scores. For more on how a BPO might use the platform, see Bananaflow for BPOs.

What should you do next?

Start with the card, then fix the sampling, then calibrate. This week, take the sample scorecard above and change the weights to match your calls. Write your fatal-error list in six lines or fewer. Time one real audit, from play to written evidence, and use that to work out how many calls you can audit in a month. Then hold your first calibration session with 5 calls, and rewrite the first line that splits you. A call audit is only as fair as the card and the sample behind it.

FAQ

Frequently asked questions

  • What is call audit in BPO?

    A call audit in a BPO is a structured review of a recorded call against set parameters, which ends in a score and written evidence. The score feeds coaching, and the evidence feeds compliance reports. It is more formal than call monitoring, which is the wider habit of listening in on calls, often live.

  • What are fatal errors in a call audit?

    A fatal error is a mistake serious enough to fail the whole call, whatever else was scored. Typical ones are skipping the identity check before sharing account details, a promise of guaranteed returns, discussing a loan with a third party, abusive language, and missing the recording notice where your rules require it. Your team writes its own list, so keep it short and unambiguous.

  • How many calls should a QA audit per day?

    It depends on how long each audit takes. In our Jaipur example, an audit of a 6-minute call takes about 15 minutes, so one analyst finishes about 28 a day in seven productive hours. That is our own assumption, not a standard. Time one real audit on your own floor and plan from that.

  • What is a good QA score?

    We found no verified public standard, so we will not give a number. Set a pass mark from your own data: score a month of calls, see where your best agents land, and check that a high score goes with a satisfied customer. If a call scores 90 and the caller is still unhappy, the card is rewarding the wrong things, a gap SQM Group also warns about.

  • What is the difference between QA and call audit?

    Quality assurance is the whole programme: the scorecard, the sampling, the coaching and the reports. A call audit is one step inside it, the scored review of a single call. People often use the two terms for the same job, and the person doing it may be called a quality analyst, QA or call auditor.

  • Is auditing 5 calls per agent a month enough?

    It is enough to coach an agent on what you heard, but too few to rank agents by score. At 5 calls and an 80% pass rate, the margin of error is about 35 points either way, which we calculated with the standard normal approximation. Use the sample to coach, and use targeted checks on complaints, transfers and long calls to catch the risky ones.

  • Can AI replace a quality analyst?

    Not fully. AI that reads transcripts can score every call, which a person cannot, but it cannot hear tone of voice or line noise, and it can mis-transcribe Hinglish. A sound setup lets software do the first pass, then has people read the lowest scores and every fatal flag, and calibrate the software against their own scores.

  • Which calls does Bananaflow Call QA review?

    Bananaflow Call QA reviews the calls its own AI voice agent takes, from the transcript, after the call ends. If you want to audit people, use the scorecard in this guide with your own recordings and analysts.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

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

Plan a pilot