AI call analysis means turning a conversation into text, producing a short structured summary, checking the sales script and writing the result into the customer record. A manager reads the review instead of listening to recordings one after another. The starting point is not the model but the legal side: a call recording is personal data, and it can only be processed on a lawful basis.
Key takeaways
The review runs as a chain, and each link can break on its own, so all of them are described before anything is connected.
The result appears in the record after the call, not at the moment the other person hangs up: the recording has to be pulled from the phone system, transcribed and reviewed. An hour-long conversation takes noticeably longer than a three-minute one, which matters if the rep needs the review before the next call.
Chats are reviewed the same way, except that the system takes the ready text of the dialogue instead of a transcript. One metric is added here — response time: in chat a customer more often leaves because of silence than because of price.
A call recording and its transcript are processing of personal data. The Law on Personal Data No. ZRU-547 treats collection, storage, use and transfer of such information as processing, and Article 18 permits it on one of the listed grounds: most often the person's consent or the need to perform a contract with them.
Important Under Article 21 consent is given in any form that lets the fact of receiving it be confirmed, so the recording notice at the start of the call is captured in the recording itself, while the legal basis and retention periods are described in a data processing policy. Under Article 22 a person is entitled to know the purposes and methods of processing, retention periods, who the data is disclosed to and whether it was transferred abroad. Employees whose calls are reviewed are data subjects too and are notified separately.
Voice is a separate question. If a person's identity is established from their voice, that is biometric data: Article 26 allows processing it only with the subject's consent, and Article 27-1 as amended by Law No. ZRU-1125 of 26 March 2026 requires storing it inside Uzbekistan. An ordinary call review does not identify anyone by voice, but where that line falls in your own process is worth agreeing with a lawyer before launch.
The cost of a mistake has gone up. Law No. ZRU-1115 of 21 January 2026 added a provision to Article 46-2 of the Code of Administrative Liability: unlawful processing of personal data using artificial intelligence technologies carries a fine of fifty to one hundred base calculation units with confiscation of the instruments of the offence. The same law forbids relying solely on the conclusions of artificial intelligence systems when taking legally significant decisions that affect a person's rights.
One more limit concerns outbound calling. Under Article 30 of the Law on Advertising No. ZRU-776, advertising distributed through telecommunications services is allowed only with the subscriber's prior consent. A list of "customers worth calling back with an offer" assembled from call reviews does not by itself amount to that consent.
Transcription accuracy depends on your audio rather than on a vendor's marketing: a phone line compresses speech, and room noise, interruptions and speakerphone add to it. A clean headset recording gives a noticeably better result than a conversation from a shop floor.
Cloud recognition services do support Uzbek: the Google Cloud Speech-to-Text language list includes uz-UZ with the Chirp and Chirp 2 models, and Russian is supported there as well. It gets harder where Uzbek and Russian are mixed inside one sentence and product names sound in a third language.
Tip Do not decide on the strength of a demo run on someone else's recordings. Take twenty of your own conversations, noisy and bilingual ones included, transcribe them and check by hand: that shows the real error rate on your audio and where a person will have to correct the transcript.
An error in the transcript travels down the chain: a misheard figure lands in the summary and from there in a task for the sales rep. That is why amounts, dates and contract numbers in a summary are flagged for checking rather than pasted into documents automatically.
A useful summary is not a retelling but a set of filled fields: those can be compared across calls, counted and filtered in a CRM. Free-form text gives none of that.
Example Reason for the call: a quote for a chain of three shops. Request: moving accounting off spreadsheets, integration with the till. Objection: "expensive compared with an off-the-shelf program". What the rep answered: compared the cost of customisation over a year. Agreement: send a quote for three locations by Friday. Next step: call on Monday. Tone: calm.
The set of fields is designed around your process: in retail it is the product and delivery time, in services the scope and budget, in a clinic the treatment and a convenient time. After the call the rep checks the summary instead of writing a report from scratch.
Script checking works when the stages are written as observable events rather than wishes such as "the rep was polite". A system cannot verify "established rapport"; it can verify "named the company and introduced himself".
| Stage of the call | What the system checks | What to do with the result |
|---|---|---|
| Opening | The rep named the company, introduced himself and gave the recording notice | A miss is a reason to rewrite the opening lines |
| Understanding the task | Questions about the task, the deadline and the budget were asked | No questions means the quote is made blind |
| The offer | The solution, the scope and a price or a range were named | No price named and the customer goes comparing elsewhere |
| Objections | An objection was voiced and answered | Recurring objections are material for the script and the website |
| Agreement | A specific next step with a date was named | No next step and the deal stalls without a task in the CRM |
| Closing | The outcome was stated and confirmed by the customer | A mismatch in understanding shows up straight away in the transcript |
Scoring by stage gives a picture across the team: where everyone dips, the script is the problem; where one person dips, it is a coaching question. How to line these stages up with deal stages is covered in our article on the sales pipeline in a CRM.
What gets connected is what belongs to the company and shares data officially. An employee's private chats in a personal account are not among them.
Hence an organisational conclusion: if sales run from personal numbers, the first move is to bring the conversations into work channels. Otherwise the review sees only part of the traffic and conclusions about the team rest on incomplete data.
The order of the steps matters more than the choice of model: the first two happen on paper and with a lawyer, not in code.
Access to call recordings is separated in the same way as access to the customer base: a rep sees their own calls, a manager sees the team's. How that is arranged in a CRM is covered in our article on access rights in a CRM.
Call review exists for the sake of decisions, so the set of metrics is fixed in advance and watched regularly rather than once during a demo.
The last metric is watched separately: it speaks to recognition quality on your own audio. If the share of corrections grows, the review starts leading you astray, and either the scoring rules or the recording source needs changing.
Call review usually stops working not because of the model but because of the way it was introduced.
On the service page, review of calls and chats is part of automating internal routine — together with document recognition and assembling recurring reports.
Above the starting figure the budget is driven by call volume, the number of channels, the state of the phone system and any requirement to keep processing inside your own perimeter: your own model on your own server removes the transfer of data outside but adds hardware and maintenance. What the work includes is described on our bots and AI assistants page, and the scope of support on our technical support page.
Syntra Systems starts with the legal side and the criteria: we fix the basis for processing and the recording notice, describe script stages in observable terms and design summary fields around your sales process. Then we test transcription on your own recordings, noisy and bilingual ones included, and show the real error rate.
Next we connect the phone system and the work chat channels, file the result into the customer record and turn the next step into a dated task. We separate access to recordings and set a retention period, and after launch we keep checking a sample of scores by hand. An assistant that answers customers from your documents is a neighbouring task, covered in a separate article.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
Yes, where consent is the basis for processing. Under Article 21 of the personal data law consent is given in any form that lets the fact of receiving it be confirmed, so the notice is spoken at the start of the call and stays in the recording. Purposes, retention periods and recipients are described in the processing policy.
They can, if these are the company’s work channels: a WhatsApp business account through the official interface and a company Telegram bot or business account. A sales rep’s private chats in a personal account cannot be connected officially, and working around the rules ends with a blocked number.
Cloud recognition services support Uzbek, and the Google Cloud Speech-to-Text language list includes uz-UZ. Accuracy in a specific business depends on recording quality and on how often languages are mixed inside a conversation, so it is measured in a pilot on your own recordings.
A score is a reason to look into something, not a ready decision. The law on artificial intelligence forbids relying solely on such a conclusion for legally significant decisions affecting a person’s rights, so a manager rechecks disputed cases against the recording.
The law sets no fixed number of days: you define the period yourself to match the purpose of processing and record it in your policy. Recordings are deleted once the purpose is achieved, the consent period expires, or the person withdraws consent.
Someone still has to review calls, but the role changes: instead of spot-checking calls, they recheck disputed scores and audit a sample of transcripts to catch any drop in accuracy. Decisions on coaching, the script and a specific manager still rest with the head of sales.
The review needs the recording files and data on each call — number, sales rep, duration and direction. These are pulled from the cloud PBX or the telephony server, so what matters isn't the age of the system but whether it can hand over its recordings.
Sources
