A customer writes in WhatsApp at 23:40: "Hello, is there a slot tomorrow?" In the morning the administrator replies, but the customer has already booked at another salon. Sound familiar? That is exactly the gap an AI agent closes: it replies at once, finds a free slot itself and books the person while they are still online.
Everyone is talking about AI in business now. In Kazakhstan, 2026 has been officially declared the Year of Digitalization and Artificial Intelligence (Forbes Kazakhstan), and on Kaspi Listings an "AI bot for WhatsApp" is sold from 10,000 ₸. Meanwhile, both WhatsApp's rules and message rates for Kazakhstan changed over the past year. Below we cover what an agent can really do, what it costs and where to start if you decide to implement one.
AI agent vs. a regular chatbot: what's the difference
A regular bot lives by buttons. "1 - prices, 2 - address, 3 - operator." As long as the customer presses numbers, all is well. But the moment they write "can I come on Saturday afternoon, to Aigerim?", the bot sends the same menu again.
An AI agent understands such messages. In Russian, in Kazakh and in the mix people here actually write in. But what matters more is that the agent is connected to your systems. It sees the schedule, creates a deal in the CRM, checks an order's status and sends a payment link. It takes prices and terms from your price list and regulations, so it makes nothing up.
To be honest: if you get the same five questions all the time, you don't need AI. A button bot is enough and will save you money. An agent makes sense when managers can't keep up and you lose customers while they wait.
Where an agent pays off
We advise starting with one task. A universal "assistant for everything" usually does everything a little and nothing well. Kazakhstani teams that deploy agents say the same: a targeted agent for a specific process can free up to 40% of employees' time (Digital Business).
Salons, clinics, car services
Booking, rescheduling, a reminder the day before the visit. The administrator opens the system in the morning and sees ready bookings. Clinics have an important limitation: the agent does not give health advice. If a person describes symptoms, the dialogue goes straight to a live employee.
Online stores and Kaspi sellers
Half the messages here are "is it in stock?", "when is delivery?", "which size should I take?". An agent that sees the catalog and orders answers these itself, suggests a replacement if the item is out and sends a payment link.
Courses and schools
Parents and students write in the evenings and at weekends: how much, is there an installment plan, when does it start. The agent answers, books a trial lesson and hands those ready to pay to a manager. During enrollment season for UNT prep courses this greatly relieves the sales team.
Real estate and B2B
The agent asks qualifying questions: budget, timing, what exactly is needed. The answers land in the lead card, and the manager calls already prepared.
Inside the company
An agent can receive job applications in Telegram and answer employees' questions about leave, certificates and sick leave. HR will thank you.
WhatsApp or Telegram
If the agent talks to customers, you almost always need WhatsApp: in Kazakhstan, people write to businesses there. Telegram is simpler and cheaper to build for. It suits internal tasks, online schools and selling digital products. Often one agent is built and connected to both messengers.
What changed in WhatsApp
From 15 January 2026, Meta banned general-purpose AI assistants, where the AI itself is the product, from the WhatsApp Business API. That is why ChatGPT and Copilot left WhatsApp (TechCrunch via Yahoo Tech). This does not affect an ordinary business. Meta states directly that the API is for customer support and notifications, and your agent is exactly that.
The conclusion is simple: the agent should deal only with your business. It answers about products, bookings and orders, and it will not write a schoolchild's essay. For sales, that is even better.
Official API or a "grey" gateway
Cheap bots often work through unofficial services connected to an ordinary number. They launch quickly. The downside is that WhatsApp can block the number, and you lose all your customer conversations. The official platform requires company verification and message templates, but the number stays yours. If WhatsApp is your main sales channel, don't risk it.
When to choose Telegram
Telegram's Bot API is free: no payment for messages and no template approval. There are Mini Apps: you can build a catalog with a cart inside the chat. If your audience already follows your Telegram channel or the agent is for employees, start here.
How much it costs
Money goes on four things. One is one-off, the rest are monthly. Monthly costs are often forgotten when comparing contractors' offers, and then the bills come as a surprise.
- Development and launch (one-off): process analysis, knowledge base, integrations, testing.
- WhatsApp messages (monthly): Meta charges for every delivered message.
- Language model (monthly): payment for AI by conversation volume.
- Support (monthly): knowledge base updates, conversation reviews, improvements.
Why development prices differ so much
On Kaspi Listings an AI assistant for WhatsApp is offered for 10-50 thousand tenge (example). Agencies name hundreds of thousands, sometimes millions, for what sounds like a similar task. These figures cannot be compared directly.
For 10-50 thousand you usually get a configured builder with a ChatGPT prompt, often via an unofficial gateway. It doesn't know your stock and can't see the schedule. Turnkey development includes process analysis, a knowledge base, integrations, handoff to an operator, testing on real conversations and analytics. When choosing a contractor, ask two things: which systems the agent will be connected to and who will support it after launch.
WhatsApp rates for Kazakhstan
Since July 2025, Meta charges for every delivered message rather than per conversation. From 1 October 2026 Kazakhstan has its own rates: previously we were in a common regional group (Alibaba Cloud). From the same day, service messages, that is replies to a customer who wrote first, became paid. The first thousand per month are free (RichAutomate).
A service message to a Kazakhstani number costs about $0.018 according to Meta partners' price lists. Promotional broadcasts cost more. Meta may change rates once a quarter (Meta documentation), so check the current table before calculating.
Let's calculate with an example
A salon in Almaty gets 1,500 people writing per month. The agent sends each about five messages, 7,500 in total. The first thousand are free, the remaining 6,500 × $0.018 ≈ $117. Add the model fee and support. A full-time administrator costs more, and the agent also answers at night.
How to implement it
Start with the conversations you already have. Choosing a neural network can wait.
- One task. Take what consumes the most time and where losses are visible in money. Usually that is booking, the first consultation or order questions.
- Knowledge base. Export 200-500 real conversations and write out the recurring questions. Add the price list, delivery and return terms, addresses and hours. The agent will answer from these materials, so everything must be up to date and in two languages.
- Boundaries. Decide in advance what the agent does not do: no discounts beyond what is allowed, no promising deadlines that are not in the system, no arguing. If the customer complains or asks for a human, the dialogue passes to a manager with a short summary. The customer won't have to explain everything again.
- Integrations. Without them the agent just chats. Connect the CRM (amoCRM, Bitrix24 or your own), booking, 1C, the catalog and payment links.
- Testing. Run a couple of hundred old questions through the agent, especially odd and conflict-prone ones. Then switch it on for part of the flow, for example only at night. If all is well, move it to the whole flow.
- Numbers. Track three metrics: how many conversations the agent closed on its own, how many of them ended in a booking or order, and how many went to an operator. Once a week read a sample of conversations and extend the knowledge base. Without this routine the agent quickly goes stale.
Ask a question or leave a request
Send us an example of your customer conversations. We will tell you which process to hand to the agent and what it will cost per month.
By submitting the form, you agree to the processing of personal data.
Personal data
The agent collects names, phone numbers and order details. That means the Kazakhstan law "On Personal Data and Their Protection" applies. The minimum: at the start of the dialogue the customer receives a link to the privacy policy and agrees to it, data is stored as the law requires, and conversations are visible only to those who need them for work. More in our article "Personal data and your website: Kazakhstan law and business requirements".
Another danger is that a neural network can confidently name a non-existent price or promotion. So we build the agent to take figures only from your systems. If there is no answer, the agent says so and calls a manager. All dialogues are saved, so a disputed case can always be reviewed.
Where people usually stumble
Most often people try to replace the whole sales team with an agent at once. The result is an agent that does ten things mediocrely and the conclusion "AI doesn't work". Start with one process.
Another common story is when the customer is "locked" in with the bot. The person wants to talk to a live manager and can't. It is better to state it honestly from the start: "I'm an AI assistant, I'll call a manager if you need."
Sometimes the agent is launched and forgotten. Three months later prices are different, promotions are different, and the agent talks about the old ones. Support must be budgeted from the start.
And the Kazakh language. An agent in Russian only loses part of its customers, especially in the regions. Prepare the knowledge base and test dialogues in both languages.
Sources
- Forbes Kazakhstan - AI is changing business
- Digital Business - how AI agents save up to 40% of time
- TechCrunch via Yahoo Tech - WhatsApp bans general-purpose AI bots
- Alibaba Cloud - WhatsApp rates from 1 October 2026
- RichAutomate - WhatsApp API rates by country 2026
- Meta - Pricing on the WhatsApp Business Platform
- Kaspi Listings - example offer