A voice AI agent is software that handles a phone conversation instead of an operator: it hears the customer’s natural speech, understands the meaning, answers in a synthesised voice and writes the outcome of the call to the CRM itself. It takes inbound calls and can also call your customer base. For a business in Uzbekistan one question decides everything: does the agent understand Uzbek, Russian and a mix of both inside one sentence. The agent has four layers: telephony over SIP on top of your PBX, speech recognition (STT), a language model that picks the next step, and speech synthesis (TTS). Unlike IVR, the customer presses no buttons and simply talks, and complex calls go to an operator with context. Zukko.AI, an AI agent platform for business, runs its own STT/TTS stack with an Uzbek voice. AI calls, inbound and outbound, are available only on the Enterprise plan: from 15,000,000 UZS a month with 2,200 minutes included and 1,900 UZS per extra minute, so 3,000 minutes a month cost 16,520,000 UZS.
What a voice AI agent is
A voice AI agent is an AI employee whose working channel is the phone rather than chat. It does what an operator on the line does: greets, finds out why the customer is calling, answers from the knowledge base, books an appointment or records a promise to pay. The difference is that the agent does not get tired by evening, answers the same way at 9 am and at 11 pm, and never forgets to log the result in the customer card.
People often search for it as a "voice robot". The word is familiar but misleading: a classic robot plays a recorded message and waits for a key press, while the agent holds a dialogue and reacts to the meaning of the answer. For the wider picture of how an AI employee differs from scripted automation, see what an AI agent and AI employee are.
How a voice AI agent works: from the customer phrase to the CRM record
Every line in the conversation travels the same five-step path. Knowing this chain helps when you choose a vendor: a weak link in any layer is audible to the customer, even if the others work well.
- Telephony. The call runs over your line: the agent connects over SIP to your existing PBX or SIP trunk. Numbers and hardware stay the same; details are on the telephony (VoIP) page.
- Speech recognition (STT). The customer’s voice becomes text. This is where the key question for Uzbekistan is settled: does the layer recognise Uzbek, Russian and a sentence where both are mixed.
- Understanding and decision (LLM). The language model works out what the customer wants, checks the script and the knowledge base, and picks the next step: answer, clarify, book or hand over to a person.
- Speech synthesis (TTS). The answer is spoken aloud. The customer can interrupt the agent, and it stops and listens, as a live person would.
- Logging the outcome. The result of the conversation, the transcript and the next step land in the customer card in the AI CRM, and the recording goes to call analytics.
Each layer adds a pause before the answer. If the total delay is noticeable, the customer starts talking over the agent and the conversation falls apart. That is why quality is tested on a live call, not on a recording. On the Enterprise plan all three layers (STT, LLM and TTS) run on Zukko.AI’s own stack and can be deployed inside the customer infrastructure by quote.
How a voice AI agent differs from IVR and robocall messages
All three tools answer a call without an operator, but they work differently. IVR walks the customer through a tree of buttons, a robocall plays a pre-recorded message, and a voice agent talks.
| Parameter | IVR (touch-tone menu) | Robocall message | Voice AI agent |
|---|---|---|---|
| How the customer answers | presses buttons | does not; only listens | speaks in their own words |
| Speech understanding | no | no | Uzbek, Russian and a mix of both |
| Call direction | inbound | outbound | inbound and outbound |
| If the answer is off-script | back to the menu | the message plays to the end | clarifies or hands over to an operator with context |
| What remains after the call | the key pressed | the fact of connection | the meaning of the answer, a transcript and the next step in the CRM |
| How the script changes | reprogram the tree | re-record the message | edit a text instruction |
The practical takeaway: IVR is good for routing ("press 2 for delivery"), a robocall for one-way notices. When the next step depends on what the customer says, you need an agent.
Which calls to give the agent and which to keep for people
Give the agent calls where the script is predictable and the volume is large. On the inbound line that means:
- routine questions: opening hours, branch address, order status, stock availability;
- booking, rescheduling and cancelling visits;
- night and weekend calls when operators are off;
- first-pass qualification: find out the task and pass a hot lead to a manager.
Outbound calls (confirmations, reminders, surveys, winning back dormant customers) are separate work with their own rules. How to build such a campaign, what script to give the agent and what a minute costs is covered in automated customer calls: scenarios, script and price per minute.
People keep disputes, emotional complaints, exceptions to the rules and large deals. The agent recognises such conversations by meaning and transfers them to an operator together with the transcript, so the customer does not have to repeat themselves. Where the line falls is covered in voice robot or operator.
How to choose a voice AI agent: 7 questions for the vendor
Every demo reel sounds good: it is recorded on a lucky example. The difference shows on your script and your customers. These seven questions filter out most unsuitable options before any test.
| Question | The answer you need | Why it matters |
|---|---|---|
| How does the agent sound in Uzbek on my script? | a call recording on your text | a ready-made reel does not show how the agent copes with your questions |
| Does it understand mixed speech? | yes, within one call | customers in Uzbekistan often switch from Uzbek to Russian mid-sentence |
| Does it connect to my PBX? | over SIP, without changing numbers | replacing telephony stretches the launch into weeks |
| Where does the call outcome go? | to the customer card in the CRM, with a transcript | otherwise the result lives in the vendor’s report, not in your funnel |
| How does the agent hand over to a person? | with context, without repeat questions | a customer asked the same thing twice leaves annoyed |
| Where is conversation data processed? | a specific environment and a contract | the customer’s number and voice are personal data |
| What makes up the price? | the billing unit and what happens above the plan | minutes, lines and tokens are counted differently |
If a vendor will not let you hear a call on your own text before you pay, that is an answer in itself.
How to test the agent before launch
Testing takes a day and saves a month of firefighting after launch. Make a few dozen test calls with your own team and make them hard on purpose:
- Noise. Call from the street, from a car, from a bazaar: customers do not call from a quiet studio.
- Mixed speech. Start in Uzbek, continue in Russian, say a product name in English.
- Interruption. Speak in the middle of the agent’s answer: it should stop and hear you out.
- Off-script question. Ask something that is not in the knowledge base. A good agent admits it does not know and offers to connect you to a person.
- Refusal. Say "do not call me again": the number should go to the stop list.
- Outcome in the CRM. After each call open the card: does the recorded result match what was said.
How rollout goes: up to 30 days
AI calls run on the Enterprise plan: rollout takes up to 30 days and voice setup is by quote. The timeline depends not on the agent but on how many systems around it need connecting.
- Line and script. You connect the line over SIP, load the knowledge base (prices, addresses, frequent questions) and describe the script. Then test calls and launch.
- Systems around the agent. Integrations with the accounting system, handover rules and reports for the manager are added.
- Environment and compliance. Security review, work inside a protected environment, the own STT/LLM/TTS stack and agreeing scripts with compliance.
What the personal data law requires
A customer’s phone number, name and voice recording are personal data. Their processing is governed by the Law on Personal Data (ZRU-547). In March 2026 it was amended to soften the localisation requirements: only certain types of data now have to be stored inside the country (gazeta.uz, 27.03.2026). The practical minimum for a voice agent: a lawful basis for the call and for processing data, disclosing that an AI is speaking, and respecting refusal. More in is it legal to call customers with a robot. This is an overview, not legal advice: your lawyer gives the final assessment.
Bottom line: what to check before choosing
A voice AI agent is neither a "voiced menu" nor a replacement for the whole call centre. It is an employee on the phone line that closes predictable conversations and leaves the complex ones to people. Choose it by a call on your own script, test it on noise, mixed speech and refusals, and only then connect it to the working line.
Want to hear the agent on your own script? Get a demo call or explore the voice AI module.
Frequently asked questions
What is a voice AI agent in simple terms
It is software that talks to a customer on the phone instead of an operator. It recognises speech, understands what the person wants, answers by voice and writes the outcome of the conversation to the CRM. The customer does not press buttons and speaks as they would to an administrator.
Does the agent understand mixed Uzbek and Russian speech
Yes. The agent recognises Uzbek and Russian and can switch between them within one call if the customer switches. Check this on test calls before launch: ask your staff to speak the way your customers do.
Can a voice agent take calls at night and at weekends
Yes, the agent has no shifts. At night it answers routine questions, books appointments and logs the request in the CRM, and anything that needs a person becomes a task for the morning. Out-of-hours calls stop getting lost.
Do we need to change our PBX or phone numbers
No. The agent connects over SIP to your existing PBX or SIP trunk and works on your numbers. Customers call the same place they called before, and no separate hardware purchase is needed.
Can the voice agent be deployed in our own infrastructure
Yes, on the Enterprise plan, by quote. Speech recognition, the language model and speech synthesis run on Zukko.AI’s own stack and can be deployed inside the customer infrastructure. Companies with strict security requirements choose this option; rollout takes up to 30 days.
How do we know the agent is ready for launch
Run test calls with noise, mixed speech, interruptions, an off-script question and a refusal. If the agent behaves correctly in every case and the outcome in the CRM card matches what was said, it can go on the working line.
How long does rollout take
AI calls run on the Enterprise plan: a rollout with a SIP line, a knowledge base, a script, integrations, a security review and work inside a protected environment takes up to 30 days. Chats and the CRM launch on the Start plan from 2 hours.