Quality control for 100% of calls: why to drop sampling
· Updated
Quality control for 100% of calls means every recorded conversation is transcribed and checked against a checklist, not just the handful of calls a manager had time to listen to. The AI does this: it scores every call against your criteria, flags the problem ones and shows the quote each score is based on. Managers spend their time reviewing problems that have already been found instead of hunting for them.
Let's look at why sampling does not work, how scoring every call is set up and what to do with the results. The feature itself is described on the speech analytics page.
Why spot-checking does not work
A team lead physically cannot listen to every call. Usually they listen to a small share, often random calls or the ones someone complained about. This approach has three weak spots:
- Rare failures get lost. A mistake that happens on one call in a hundred almost never lands in the sample.
- Scores depend on the reviewer. Two supervisors score the same call differently.
- Feedback comes late. The agent learns about the problem a week later, after repeating it many times.
How the AI scores every call
- The recording comes from the PBX. Calls arrive with their recordings from your PBX or a cloud PBX such as A1 VATS. Recordings can also be uploaded by hand.
- Transcription. Speech becomes text with speakers separated. On our test environment recognition ran at an RTF of 0.27–0.37: a minute of audio is processed in roughly 16–22 seconds. Names and numbers can be misrecognized, so every score links to the moment in the recording that can be replayed.
- Checklist scoring. The AI goes through your criteria and scores each one, with a quote. A criterion the call did not touch is marked "not applicable".
- Call goals. Separately, it marks whether the goal was achieved: a booking, a sale, an issue resolved.
- Outcome. The score, a short summary and the caller's sentiment appear on the call record and in the CRM.
What changes for the manager
| Before | After |
|---|---|
| Listens to random calls | Opens the calls with low scores |
| Learns about problems from complaints | Gets an alert in Telegram |
| Argues about a score | Sees the quote and the moment in the recording |
| Builds reports by hand | Sees trends per agent and per team |
How to write the criteria
A score is only as good as the checklist. Criteria should be specific and observable: "named the company and their own name", not "was polite". A detailed guide with examples is in our post on the call quality scorecard.
Keywords and reasons for refusal
Beyond the checklist, it helps to watch the words. In the dictionary you set products, competitors and stop words, and the system finds them in transcripts across word forms. The AI also flags prices and competitor names on its own. For managers there is a breakdown of refusal reasons: which objections come up most often, with quotes, and whether they were handled.
Recordings from any source
Calls do not have to come from a connected PBX. You can upload a recording by hand in common audio formats. A stereo recording is split into agent and caller, and uploading the same file twice is rejected. This is handy for scoring an archive of calls before telephony is even connected.
What to do with the results
Scores on their own change nothing. You need actions:
- Rule-based alerts. Set a metric and a threshold — for example, a failed key criterion — and the manager gets a Telegram message.
- Coaching sessions. The moments where a criterion scored low are collected with a quote and a link to the second in the recording. A review takes minutes.
- Promises to customers. The AI finds promises agents made, such as "I'll call you back tomorrow morning", and flags the overdue ones.
- Questions across the whole base. "Why did customers decline last week?" — the answer is built from calls over a period of up to 93 days, with links to specific conversations.
Common objections to scoring every call
- "Agents will push back." Criteria are known in advance and every score is backed by a quote. That is more transparent than random spot-checks.
- "The AI makes mistakes." It can, especially with names and numbers. That is why a score leads back to the recording and a person replays disputed calls.
- "We don't have time to review that many scores." You do not review all of them, only the flagged ones: low scores and alerts.
Sentiment: the caller's only
The AI tracks the caller's sentiment over the course of the call: where frustration appeared, where an objection came up. Agents' mood and emotions are not assessed. This is a deliberate choice: scoring is for reviewing conversations, not for watching people.
It works for human agents too
Speech analytics scores any recorded call, not only conversations with the voice bot. That is why teams often connect it first: it does not change a single conversation but immediately shows what is happening on the line. A bot for routine calls can be added later — see AI calls.
How often to look at the results
Daily: alerts and low-scoring calls — a few minutes of work. Weekly: a team summary that arrives in Telegram on a schedule, with call count, minutes, average score and missed calls. Monthly: questions across the whole call base and a review of the checklist if business rules have changed. This rhythm needs no dedicated analyst and keeps problems from piling up.
How to start
- Connect your PBX or upload recordings.
- Build a checklist and goals for your industry.
- Compare AI scores with manual ones on a few dozen calls.
- Turn on alerts and review the first results.
We will show a report on your recordings as part of a demo and pilot.
FAQ
Why is scoring every call better than spot-checking?
A sample misses rare failures and depends on the reviewer. When AI scores every call against one checklist, problem conversations are found right away.
How fast are calls transcribed?
On our test environment the RTF was 0.27–0.37: a minute of audio is processed in roughly 16–22 seconds. This is an early measurement, not a guaranteed turnaround.
Can I dispute an AI score?
Yes. Every score is backed by a quote and a link to the moment in the recording, and a supervisor can add a manual score next to the AI's.
Can old recordings be analyzed?
Yes. Recordings can be uploaded by hand or imported from a connected PBX's history, then transcribed and scored.
Is agent mood assessed?
No. Sentiment is tracked for the caller only.
Does analytics work for calls without the bot?
Yes. It scores any recorded calls by your employees, not only conversations with the voice bot.