Automated call quality assurance in Uzbekistan: scoring agents without manual QA

Automated call quality assurance means every agent call is scored by AI instead of a supervisor sampling a few recordings. The platform transcribes the conversation, splits the agent’s and the customer’s lines and checks the call against a five-criteria rubric, with separate rubrics for sales and support. The result is an AQS (Agent Quality Score) from 0 to 100, with a quote from the call under each criterion, so the score can be checked against the recording. By McKinsey’s estimate, manual review covers fewer than 5% of conversations at 70–80% accuracy, while a mostly automated process can exceed 90%. Tone of voice, FCR and predicted CSAT and CES are calculated alongside the score. Calls are analysed in Uzbek, Russian and English, including mixed-language phrases; without IP telephony, the Zukko Recorder app for Android collects the recordings. Supervisors keep the disputed calls, the coaching and the criteria themselves.

How call QA without manual review works

In manual QA a supervisor picks a handful of recordings, listens and fills in a checklist. The weak point is arithmetic, not people: there is never enough time for the whole shift, and the result depends on which call happened to be picked. Automated scoring does not pick. It runs every conversation through the same steps.

1. The recording goes in for analysis. The platform connects to your PBX or telephony provider; if you have no IP telephony, the Zukko Recorder app for Android collects the recordings.
2. The conversation is transcribed. The text is split by speaker, agent and customer, and the language is detected automatically: Uzbek, Russian or English, including Uzbek-language calls where Russian is mixed into the same sentence.
3. The call is scored against the rubric. The AI works through five criteria, leaves a quote from the call under each and produces an AQS from 0 to 100.
4. Related metrics are calculated alongside. Agent and customer voice tone, whether the issue was resolved on first contact, predicted satisfaction and customer effort.
5. Results roll up per agent. The dashboard, the leaderboard and the morning coaching report show the whole shift rather than a single call.

Why sampling stops working, and what speech analytics checks in each call, is covered in the article on call quality control. This article answers a different question: how the agent score itself is built, and who is responsible for what once it is in place.

What the AQS score is made of

AQS (Agent Quality Score) rates how well an agent handled a specific call, as a number from 0 to 100. It is assigned against a five-criteria rubric, and sales calls and support calls use different rubrics, because a salesperson and a support agent cannot be measured with the same ruler.

Each criterion shows the quote behind it. That is the main difference from a paper checklist: the score can be checked against the recording, the agent sees what it was given for, and any argument is with their own words rather than with a supervisor’s opinion.

Beside the score the platform calculates separate indicators. They are not part of AQS; they add to it. Agent and customer voice tone is labelled across 20 tones in every segment of the call. For support calls the platform detects FCR (was the issue resolved on first contact) and predicts CSAT on a 1–5 scale and Customer Effort Score on a 1–7 scale. For sales calls it detects whether the deal closed. What each of these metrics measures is explained in the speech analytics guide.

What stays with people and what automated scoring takes over

Automated scoring takes the largest and least creative part of the job off the supervisor: listening and filling in checklists. Decisions about people, and about the rules themselves, stay with a human. The split looks like this.

Once scoring is in place, the supervisor’s role shifts from hunting for mistakes by ear to working with a finished review. Three tasks remain entirely theirs. First, disputed calls, where an agent disagrees with the score or the situation is unusual. Second, coaching: the report shows the growth areas, but people are taught by people. Third, tuning the criteria to the company’s product and script.

How to tailor the scoring criteria to your team

The standard sales and support rubrics work as soon as telephony is connected. Large contact centers usually need a rubric of their own, built around their product, script and compliance requirements. It is not bought off the shelf; it is assembled in this order.

1. Describe the reference call. Collect 10–15 conversations the team agrees are exemplary and write down what exactly makes them good.
2. Break it into checkable criteria. Each item must be read one way only: “stated the return terms”, not “was generally polite”.
3. Define criteria by meaning, not by exact wording. An agent may say “this call is recorded” or “we are recording the conversation”; the criterion should accept both.
4. Allow for language switching. In Tashkent, agents and customers move between Uzbek and Russian within a single call, so a required statement can come in either language.
5. Run the rubric on your archive and compare it with manual scores. Where the AI and the supervisor disagree, fix the wording of the criterion instead of adjusting the score.
6. Review the criteria whenever the script changes. Products and promotions change, and yesterday’s mandatory condition may no longer apply.

The more a criterion matters to the business, the more precisely it should be described. A sales team will spell out objection handling and closing; a bank contact center will spell out process, meaning customer identification and mandatory disclosures. There is no single correct rubric, only one that honestly reflects what you pay agents for.

How to calibrate automated scores against manual ones

Calibration is what separates a working rubric from a tidy spreadsheet. Take an archive of calls, get automated scores for them and have supervisors score the same recordings by hand. Then look at the gaps: where the AI and the human disagreed most, and why. The aim is not to nudge scores until they match but to find ambiguous criteria and rewrite them so everyone reads them the same way.

Most disagreements come from criteria defined by words rather than meaning. A typical case is the rule “the agent does not promise terms that do not exist”. The agent says, “No, there is no discount on this plan.” The word “discount” was spoken, yet nothing was promised. If the criterion is written as a banned word rather than a banned promise, an honest call loses points. It is worth reserving one full pass over the archive for context errors like this before agents ever see their scores.

A practical threshold: if after two or three iterations the automated score agrees with the averaged view of several supervisors within a reasonable spread, the rubric is ready for use. A perfect match is neither achievable nor needed, since supervisors disagree with each other too. McKinsey puts manual scoring accuracy at 70 to 80 percent, while a largely automated process can reach more than 90 percent.

From score to agent coaching

An AQS is where the conversation with an agent starts, not where it ends. Its value is that it shows more than “72 out of 100”: it shows where the points went. Need discovery was skipped, an objection went unanswered, the call ended without a next step. Coaching is built on those specific gaps, not on a general “do better”.

Overnight the platform prepares a coaching report for everyone who took calls: the day’s KPIs, the best and worst call of the shift, three strengths with quotes, three growth areas with a better alternative, and three recommendations for the next shift. The supervisor takes two or three calls with the same problem, goes through them with the agent and checks a week later whether the score on that criterion has improved.

All of this runs on the telephony you already have, with no infrastructure replacement. Setup takes from 2 hours for small and medium businesses and up to 30 days for large contact centers that need CRM integration and custom rubrics. Call analytics goes up to 300 reviews a month on the Start plan, and the Business plan lifts that limit; current terms are on the pricing page and the Call Analytics page.

Conclusion

Call QA without manual review does not remove the supervisor; it redistributes the work. The AI scores every call against a five-criteria rubric, assigns an AQS from 0 to 100 and backs it with quotes, while voice tone, FCR and predicted CSAT sit beside it as separate indicators. People keep the disputed calls, the coaching and the criteria. The most underrated step is calibration: without it, a loosely worded criterion penalizes honest calls. And the principle that matters most is that a score is useful as a breakdown of causes, because that is what an agent training plan grows from.

Want to see the scoring on your own calls? Get a demo call or explore the Call Analytics module.