AI Voice Call Quality Assurance Checklist
Printable QA checklist for AI calls to traders: opening, comprehension, interruptions, objections, handoff, disclosures and disposition accuracy.
- Who it is for
- Broker ops leads, campaign owners and QA reviewers who listen to AI call recordings for dormant trader, deposit and KYC campaigns.
- Time to complete
- About 4 minutes per call reviewed
This checklist is for the person who listens to AI call recordings after a dormant trader, deposit or KYC campaign runs. It splits a call into the seven places an AI agent goes wrong: the opening, whether it understood the trader, how it handled being interrupted, what it did with objections, the handoff to a human, the AI and recording disclosures, and whether the outcome logged in your CRM matches what happened on the line. Score a weekly sample against it and you'll know which fix to make to the script, the voice or the routing before the next batch dials.
Pick the calls to review
A QA pass is only as good as the sample. Set the rules once so the same kind of calls get picked every week, whoever is reviewing.
Pull a fixed number of recordings per campaign each week, split across every disposition your CRM uses, wins and losses alike.
Reviewing only the funded or booked calls tells you the script works when it works. The dropped and 'not interested' calls tell you why it doesn't.
Include at least one call per language and per voice the campaign runs.
A script that sounds fine in English can come out stiff in Arabic or Vietnamese. Topcalls runs calls in 32 languages, so list the ones your campaign uses and cover each.
Add every call that was transferred to a human and every call tied to a trader complaint.
These are the calls where the most can go wrong. Review all of them, not a sample.
Add every call under 20 seconds and every call over your expected maximum length.
Very short calls are usually a bad opening or a wrong number. Very long ones are usually the agent looping.
Record who reviewed each call, the date, and the campaign and script version it ran on.
Without a script version on each review you can't tell whether a fix helped.
Open the recording and the transcript side by side for every call.
The transcript shows what the speech recognition heard. The recording shows what the trader said. The gap between them is a finding on its own.
Check the opening
The first ten seconds decide whether the trader stays on the line. Review this part of every sampled call.
Confirm the agent waited for the trader to speak before starting and didn't talk over the greeting.
An agent that starts its pitch over 'hello' sounds like a robocall. Note the timestamp when the agent first spoke.
Confirm the agent named the brokerage and the reason for the call in the first sentence.
'This is the assistant from [Broker] about your trading account' is a reason. 'I'm calling with an exciting opportunity' is not.
Confirm the agent used the trader's name as it appears in the CRM and pronounced it correctly.
Mispronounced names draw complaints on multilingual books. Flag any name the voice got wrong and add it to the pronunciation list.
Check the agent referenced something true about the account: last activity, platform, or the deposit or KYC step that's pending.
'You opened an MT5 account in March' tells the trader this is their broker. A generic opener could be from anyone.
Check the opening fits the campaign type: a KYC reminder shouldn't sound like a sales pitch, and a reactivation call shouldn't sound like a compliance notice.
Time the gap between the trader's first answer and the agent's first reply.
Topcalls targets sub-500ms response latency. Flag any opening where the pause was long enough for the trader to say 'hello?' twice.
Note whether the trader asked 'who is this?' or 'what is this about?' in the first 30 seconds.
If more than a few sampled calls have this, the opening isn't doing its job. That's a script fix, not a voice fix.
Check whether the agent understood the trader
Comprehension failures show up as the agent answering a question the trader didn't ask. Read the transcript with the recording playing.
Mark every place the transcript differs from what the trader actually said.
Account numbers, amounts, currency names and platform names (MT4, MT5, cTrader) are where recognition fails most. Count them.
Check the agent answered the question the trader asked, not a nearby one.
Trader asks 'do I still have money in the account?' and the agent explains how to deposit: that's a comprehension miss. Log it.
Check the agent followed a language or dialect switch if the trader made one.
Traders in Dubai or Kuala Lumpur often switch mid-call. Note whether the agent followed or kept going in the wrong language.
Confirm the agent asked for clarification when the trader was unclear instead of guessing.
'Sorry, did you say you'd like a callback or that you'd like to close the account?' is the right behaviour. Guessing wrong on an account closure is a complaint.
Check the agent handled background noise, a bad line or a speakerphone without repeating itself more than twice.
After two repeats the agent should offer a callback or a text, not loop.
Confirm the agent didn't invent account details, balances, bonus terms or spreads it wasn't given.
Any number the agent said that isn't in the campaign data or the knowledge base is a serious finding. Escalate it the same day.
Note any question the agent couldn't answer and what it did next.
'I don't have that information, I can have someone from the account team call you' is fine. Silence or a subject change isn't.
Check interruption handling
Traders interrupt. A good agent stops talking, listens, and picks up the thread. Score every interruption in the sampled call.
Count how many times the trader spoke while the agent was talking, and how many of those the agent stopped for.
Write both numbers on the review. The ratio is your interruption score for the call.
Check the agent stopped within a word or two when interrupted, not at the end of its sentence.
Finishing the sentence anyway is the tell that gives away an automated call.
Confirm the agent dealt with what the trader interrupted with before returning to its script.
If the trader cut in with 'I already funded last week', the agent should respond to that, not resume the deposit pitch.
Check the agent didn't treat a cough, an 'mm-hm' or background talk as an interruption.
Stopping for every noise makes the agent sound hesitant. Flag calls where it paused for something that wasn't speech.
Check the agent didn't restart its sentence from the beginning after a false interruption.
Traders notice when the same opening clause plays three times. It's a frequent complaint alongside long pauses.
Confirm the agent resumed at the right point after a real interruption, without repeating what it had already said.
Note any moment where both sides talked over each other for more than a couple of seconds, and who backed off first.
Check objection handling
Dormant traders have a stock set of objections: they lost money, they moved brokers, the spreads were bad, they don't have time. Check each one the agent met.
List every objection the trader raised and match each to a branch in your objection flow.
An objection with no matching branch is a script gap. Send it to whoever owns the flow.
Check the agent acknowledged the objection in one sentence before responding, without parroting it back word for word.
Confirm the agent didn't make a promise the brokerage can't keep: guaranteed returns, bonus terms, spread claims or a callback time nobody owns.
Anything that sounds like investment advice or a performance claim goes to compliance the same day. Don't wait for the weekly review.
Check the agent stopped after the trader said no twice.
A third attempt is the difference between persistent and pushy. Set the number in the script and check the agent respected it.
Confirm 'remove me from your list' was treated as an opt-out, with the agent confirming it and ending the call politely.
Check the suppression list the next day to see the number actually landed there.
Check that an objection about a specific past problem (a stuck withdrawal, a margin call, an open support ticket) was routed to a human rather than argued with.
The agent can't fix a withdrawal from March. It can get someone who can.
Note the tone the agent took on each objection: matter of fact, apologetic, or salesy.
Write one word per objection. Patterns across calls tell you whether the voice or the script needs adjusting.
Check the handoff to a human
For calls that reached a human, or should have, check that the transfer happened at the right moment and that nothing was lost on the way.
Confirm the agent offered a human the first time the trader asked for one.
Check the agent triggered a handoff on the conditions your workflow sets: deposit intent above a threshold, a complaint, a compliance keyword, a request to close the account.
Pull the handoff rules from your workflow template and tick each one that applied to this call.
Check the trader was told what was happening before the transfer: who they'd speak to and roughly how long the wait would be.
Confirm the human agent had the call context on screen: trader name, account, campaign, and what the AI had already covered.
Ask the human agent who took the call, or compare the CRM note timestamp with the transfer timestamp.
Check the trader didn't have to repeat anything they'd already told the AI.
Every repeated question is a point against the handoff.
Check what happened when no human was free: a callback was booked with a specific time, and it was actually placed.
Look up the callback in the CRM. A booked callback that never happened is worse than no offer at all.
Confirm the recording and transcript cover the whole call including the human part, or that the two parts are linked.
A regulator or a complaints team asking for a call wants the entire call.
Check the AI and recording disclosures
Your compliance officer signed off specific wording. Check the agent said it, word for word, at the right moments.
Confirm the agent identified itself as an AI or automated assistant in the opening, using the approved wording.
Compare the transcript to the approved script line. Paraphrases are a finding even if the meaning is the same.
Confirm the recording notice was given where your compliance review requires one, before anything substantive was discussed.
Which jurisdictions need it and how it's worded is a compliance decision. Your job here is to check it was said.
Check the agent answered 'is this a real person?' honestly and offered a human.
Play any call where the trader asked this to your compliance officer.
Check that opt-out instructions were given when the trader asked how to stop the calls, and that the agent didn't try to talk them out of it.
Confirm the disclosures were said in the trader's language, rather than only in the campaign's default language.
Check the voicemail, if one was left, carried the same identification and a callback route.
Voicemails are easy to forget in QA. Pull at least two per week.
Log any call where a disclosure was missing, late or altered, and send it to your compliance officer or counsel the same day.
This isn't a weekly-report item. If the cause is the script and not a one-off, pause the campaign until it's fixed.
Check disposition accuracy
The disposition is what your CRM, your dashboards and your next campaign act on. If it's wrong, everything downstream is wrong too.
Compare the disposition the agent logged to what actually happened on the call, and mark it right or wrong.
Track the share of wrong dispositions per campaign per week. That's the one number from this checklist that belongs on your dashboard.
Check that 'interested' or 'callback requested' has a specific next step attached: a time, a channel, an owner.
'Interested' with no next step is a lead that goes cold in your CRM.
Confirm 'not interested' calls where the trader raised a fixable problem were tagged with that problem rather than closed.
'Not interested, withdrawal issue' and 'not interested' are different leads.
Check that opt-outs, wrong numbers and third-party answers were logged with the correct code and reached the suppression list.
Look the number up the next day.
Check the call summary the agent wrote matches the transcript and doesn't include anything the trader didn't say.
Summaries drift toward what the script expected. Read them against the transcript, not on their own.
Confirm the fields that feed your KPI worksheet were filled: outcome, deposit intent, handoff, and reason code.
Empty fields become zeros in reports and make the campaign look better or worse than it was.
Check the disposition reached the CRM and the MT4/MT5 or back-office record within the window your integration promises.
Compare the call end time to the CRM update timestamp on three calls per week.
How to use this
- 1
Print one copy per call reviewed, or copy the sections into your QA sheet with a pass, fail and not-applicable column per item.
- 2
Pick the weekly sample using the first section, then work through the other seven sections for each call with the recording and transcript open.
- 3
Write one line of evidence next to every fail, with the timestamp in the recording.
- 4
Escalate the same day anything in the disclosure section, anything that sounds like a performance claim, and any invented account detail.
- 5
Total the fails per section across the week's sample. The section with the most fails is what you fix next: script, voice, routing or CRM mapping.
- 6
Re-review the same section on the next batch after a fix so you can see whether it worked.
Next step
We listen to a few of your recordings with you, map the findings to script, voice and CRM fixes, and scope a pilot on your dormant book.
Book a 30-minute QA walkthroughRead next
Related resources
AI Call Quality Scorecard
10-metric scorecard for AI calls at forex and CFD brokerages: definitions, a 1 to 5 scale, weights, auto-fail rules and a weekly review sheet.
AI-to-Human Handoff Workflow Template
A fill-in handoff workflow for AI voice calls to traders: trigger conditions, warm transfer vs callback, context passed, desk hours and fallbacks.
Trader Objection Handling Flow Template
Objection handling flow for forex and CFD broker calls: ten trader objections with acknowledge, answer and next-step branches, plus end and handoff rules.