What drives the price
Mostly, how many scenarios the agent runs and how many systems surround it. The Scenario tier covers one job: inbound calls or reminders, in Uzbek and Russian. The Call centre tier adds outbound calling, outcomes in the CRM, booking into the schedule and transfer to an operator. The Platform tier is for when scenarios change often: a builder that lets you edit them without a developer, a brand voice, topic analytics and a quality score for every conversation. Connecting your number and PBX, and an SMS after the call, are separate lines in the configurator. Dialogue complexity matters too: confirming a visit is simpler than helping someone choose a service.
How the work runs, week by week
We start by listening: we take recordings of your real calls, pick out the frequent scenarios and decide which one to automate first. Then we write the script: the greeting, the clarifying questions, what the agent does itself and when it calls a person. We build the agent on OpenAI models and run it against your recordings before it ever hears a live customer. We connect your number and PBX, the CRM and the schedule. Launch starts on a share of calls: every conversation is transcribed, and we fix the script wherever the agent stumbled. You see a working version every two weeks, and the agent takes the full flow only after the pilot.
Uzbek and Russian speech: how we test it
In Uzbekistan people often mix Uzbek and Russian in one sentence, use old street names, dictate a phone number in chunks and call from a noisy car. So we test recognition on recordings of your own calls rather than on a demo, separately for each language, and show where the agent makes mistakes. Everything the result depends on — name, phone number, date and time — the agent repeats back and asks to confirm. After the call the customer receives an SMS via Eskiz with the address and time, so the customer spots a recognition slip right away, not on the day of the visit. The agent replies in the language it was spoken to in and switches when the customer does.
Risks and common mistakes
The first mistake is trying to automate everything at once. An agent that runs one scenario very well earns more than one that answers any question poorly. The second is leaving no way out to a person: if the caller asks for an operator or the question falls outside the script, the agent transfers immediately. The third is long pauses: people hang up when a reply lags, so we measure latency during the pilot. The fourth is outbound calling without rules: call at reasonable hours and only people who expect the call. Finally, a call outcome that never reaches the CRM is as good as lost, so every call leaves a trace on the customer record.