How much does call analysis cost for a call center in Uzbekistan

There is no single market price for call analysis: five drivers make up the bill — audio minutes, recognition languages (Uzbek STT), cloud or on-prem, depth of analysis, and integrations. That is why call analytics pricing is compared by the vendor’s billing model, not by one number: per minute, per seat, per token, or a fixed fee with overage. The baseline is manual review: McKinsey estimates that it covers less than 5% of calls, so the fair comparison for automated review is the payroll of a supervisor team, not zero. For CIS banks, on-prem costs more than cloud: servers, security and a rollout of up to 30 days come on top of the subscription. In Zukko.AI, an AI agent platform for business, call analysis is counted in tokens: an average call review uses about 65,000 tokens. The Start plan (699,000 UZS a month on monthly billing, 90 million tokens) allows up to 300 reviews a month; the Business plan (3,900,000 UZS, 600 million tokens) has no such limit, and its tokens cover 7,600–11,300 reviews.

A reference point for the maths: McKinsey estimates that manual QA covers less than 5% of calls with 70–80% scoring accuracy, while automated review covers 100% and can reach accuracy above 90% (McKinsey). So the cost of analytics is never measured against zero — it is measured against what sample listening already costs you. Do not confuse it with the price of call tracking either: call tracking is usually billed monthly for rented substitute numbers, forwarding minutes and a subscription fee, and it answers “where did the call come from”, not “what was said in it”. Below: how to turn your own call volume into a budget, what a subscription usually leaves out, and how to test payback on your own data.

Five drivers of call analysis cost

Before you compare price lists, it helps to know what you are paying for. The price of speech analytics almost never hangs on one number — several parameters pull it up or down at the same time. Here are the five main ones.

• Call and minute volume. The main driver. 500 calls a month and 50,000 sit in different price leagues. Many models count audio minutes, so long conversations cost more than short ones.
• Recognition languages. Uzbek STT (speech-to-text) is harder and more expensive to maintain than Russian or English, and in Uzbekistan agents and customers switch languages readily inside one call. Quality on Uzbek-language calls is what buyers actually pay extra for.
• Cloud or on-prem. Cloud is cheaper and quicker to start; deployment inside a bank’s perimeter (on-prem) costs more because of infrastructure and rollout, but is sometimes required by regulation.
• Depth of analysis. A plain transcript is one price; tagging up to 20 voice tones, scoring against the AQS rubric, finding script violations and predicting churn is another. The deeper the review, the higher the price.
• Integrations. Connecting the existing PBX, call recording and CRM is usually part of the rollout cost. Custom integrations and API work make the project bigger.

In short — what the price of call analysis depends on, which way each driver moves the bill, and what to clarify with the vendor before you compare offers.

Billing models for speech analytics

Comparing vendors by a single number is pointless, because they sell in different units. Some charge per audio minute, some per agent seat, some per token, and some a fixed fee plus a surcharge for overage. Below are the four models, their logic and who each is built for.

No model is “better” in a vacuum: per minute pays off on short calls, per seat on a stable team with heavy traffic, per token when the AI handles chats and the CRM as well as calls. The Zukko.AI model is built on tokens — more on that below and on the pricing page. Call tracking solves a different task altogether: it attributes a call to an ad channel, while speech analytics reviews the conversation itself.

How do you budget for your own call volume?

The vendor works out the exact price for your volume, but you can estimate the order of magnitude yourself. Take a hypothetical call center (the numbers are illustrative, not the Zukko price list).

1. Count the volume. For example, 5,000 calls a month at an average of 4 minutes is about 20,000 minutes of audio.
2. Convert it into the model’s unit. Per minute: 20,000 minutes × the vendor’s rate. Per token: an average review of a 1–1.5 minute call ≈ 65,000 tokens, so 5,000 such calls ≈ 325 million tokens a month. A 4-minute conversation uses more, about 100,000–120,000 tokens, so the flow in this example comes to ≈ 500–600 million tokens a month.
3. Add depth. Do you need voice tone, the AQS rubric, violation search? A basic transcript is cheaper than a full review.
4. Account for deployment. Cloud is the minimal start; on-prem adds infrastructure and a rollout of up to 30 days.
5. Compare with the plan. Match the resulting number of tokens against the plans on the pricing page — that shows which plan you fit into.

In our calculations with customers we almost always see the business underestimate its volume in minutes: the calls “don’t seem like many”, yet conversations of several minutes each turn a modest flow into tens of thousands of minutes. So it is safer to budget in minutes or tokens than in a rough count of calls.

Hidden line items beyond the subscription

The subscription is only one line on the invoice. The others surface after signing if you did not ask about them in advance. Below is what most often turns out to be “not included” and is worth agreeing before launch.

What does it cost a business of your size?

The order of the budget depends not on the industry but on conversation volume and data requirements. There are deliberately no specific sums here: every vendor has its own model, and our plans live on the pricing page. But the set of things you will pay for is predictable — check it against your profile.

Is automated review cheaper than manual QA payroll?

The main mistake in estimating cost is comparing the price of speech analytics with zero, as if quality control used to cost nothing. It did cost something: supervisors did it on salary. And the maths was poor: McKinsey estimates that manual QA covers less than 5% of calls with 70–80% scoring accuracy, while a largely automated process can reach accuracy above 90% and save more than 50% of QA costs (McKinsey).

Less than 5% of the same 5,000 calls a month is fewer than 250 conversations. To listen to noticeably more by hand you need not one QA specialist but several — and you still will not reach 100%. The right cost comparison looks like this: a monthly subscription for automated review versus the payroll of a supervisor team that still hears only a small share of conversations. In that comparison automation is almost always cheaper per unit of coverage, because one lost major customer among the “unheard” 95% outweighs hundreds of routine calls.

If your goal is not to size a budget but to close the quality-control blind spot without growing the QA headcount, the path is laid out in how to implement speech analytics.

How do you know whether call analysis will pay off?

Payback is not calculated as “savings on people” but as two flows: what your current quality control already costs you, and what the conversations you never hear are worth. Work it out in UZS on your own data — someone else’s payback percentages are useless here.

1. Cost of today’s QA. Supervisor payroll × the share of their time spent listening. That is your current bill for quality control, even if nobody calls it “analytics”.
2. Price of one lost customer. Average deal size × margin. That is what one conversation costs when nobody reviewed it and nobody called back.
3. Size of the blind spot. If sample listening covers less than 5% of conversations (McKinsey estimate), the remaining 95% is your unmeasured risk, not “business as usual”.
4. Break-even threshold. Divide the monthly cost of the platform by the profit on one deal — that gives you how many deals a month you need to save to break even.
5. A check on a real month. Take last month, run its recordings through the review and see how many script violations and unhandled requests turn up. That is a fact, not a forecast.

Ways to cut the cost without losing the point

Not every call center needs maximum depth from day one. There are honest ways to shrink the bill without turning the project into an imitation of quality control.

• Start with one team. One sales team or one support line instead of the whole contact center: audio minutes drop several-fold, while the findings on scripts and objections already work.
• Split depth by call type. Full review for sales and complaints, a plain transcript for enquiry and internal calls.
• Do not keep the archive longer than needed. Retention beyond the regulatory minimum is almost always a surplus line item; check the requirements for your own segment.
• Do not pay separately for routine. If reminders, tags and status changes are handled by automation inside the same platform, you do not need a separate contractor for them.
• Plan for peaks, not the average. Sale season, Ramadan and ad launches bring a surge in minutes — discuss overage in advance, not when the invoice arrives.

Why is on-prem more expensive than cloud for banks?

For SMBs, cloud is the cheapest and fastest option: a subscription, onboarding from 2 hours, no infrastructure of your own. For CIS banks the picture is different: call review touches personal data, and the requirements of a bank’s own security function and its regulator are stricter than for an SMB. The law itself, after the ZRU-1125 amendment (26 March 2026), makes in-country storage mandatory only for certain categories — biometric and genetic data and data on telecom subscribers; the rest may be processed abroad under the conditions of Article 27-1 (gazeta.uz). More in our breakdown of ZRU-547 and data localisation.

Hence the price gap: deployment inside the bank’s perimeter (on-prem) adds the cost of servers, security and a rollout of up to 30 days, but the recordings never leave the bank’s infrastructure. In Zukko.AI, data is stored in Uzbekistan, and on-prem inside the bank’s perimeter is an Enterprise option. More on what bank deployments involve — in the guide to speech analytics for businesses in Uzbekistan.

Questions to ask a vendor before signing

Commercial offers are compared not by the bottom-line figure but by the answers to awkward questions. This list closes off most second-month surprises.

1. Which unit the bill is counted in — minute, seat, token or pack — and what happens on overage.
2. How recognition accuracy was measured on Uzbek speech and on mixed Russian–Uzbek conversations, on what data and by whom.
3. Whether rollout, scoring-rubric setup and connecting your PBX are inside the subscription or billed separately.
4. Where recordings and transcripts are physically stored and what happens to them when the contract ends.
5. What it costs to extend the retention period and to add minutes for a seasonal peak.
6. Who is responsible for building the integration with your CRM and how long it takes.
7. What support includes: language, working hours in Tashkent time (UTC+5), incident response time.

How does Zukko.AI calculate the cost?

Zukko.AI does not charge separately “for call analysis” as an isolated service — call review is part of one platform together with the CRM, channels and automation. The billing unit is the token: an average call review ≈ 65,000 tokens. One million tokens costs about 7,767 UZS on the Start plan (699,000 UZS for 90 million) and 6,500 UZS on the Business plan (3,900,000 UZS for 600 million), so an average review costs about 505 UZS on Start and about 423 UZS on Business for a call of 1–1.5 minutes. All AI functions draw on the same tokens; call reviews go up to 300 a month on Start, and that limit is lifted from the Business plan up.

• Tokens — you pay for the work the AI actually does, not for merely having a CRM. Plan tokens → purchased packs → overage: 8,000 UZS per 1 million tokens on Start and Business.
• Seats — the number of CRM seats depends on the plan and can be added as the team grows.
• Plans — from Start for small businesses to Enterprise for banks, with on-prem and a volume set by contract.

Take the same example. 5,000 average reviews are 325 million tokens, about 2.1 million UZS of the 3,900,000 UZS Business plan, and that flow does not fit into Start, which allows up to 300 reviews a month. If conversations run to 4 minutes, the same flow takes up almost the whole Business volume. The price of a review here is its share of the plan’s tokens when they are fully used, not a separate charge. Exact sums, tokens and packs are on the pricing page, which is also the easiest place to match your estimated tokens against the plans. What the module can do is on the Call Analytics platform page.

Conclusion

“How much does call analysis cost” is a question without a single number, because the price is set by audio minutes, Uzbek STT, cloud or on-prem, depth of review and integrations. Compare vendors by model (per minute / per seat / per token / fixed + overage), and budget in minutes or tokens rather than by eye. Above all, compare the cost of automated review not with zero but with the payroll of a manual QA team that still hears less than 5% of calls (McKinsey estimate). For SMBs the start is cheaper and faster (cloud, onboarding from 2 hours); for banks it costs more — for on-prem inside their own perimeter (rollout of up to 30 days). The actual plans are on the pricing page.

Want to size the cost for your own call volume? Get a demo call or explore the Call Analytics module.