AI assistant
What it is
The AI assistant is an employee you describe in words rather than program. It reads the customer's incoming messages, replies on its own, finds out what you need to know, and hands the conversation to a real person at the right moment.
The assistant works in WhatsApp, WhatsApp Business API, Instagram and the website widget. By default it replies in all of these channels at once, but you can keep only the ones you need — in the Channels line at the top of the section, click "Change".
The assistant subscription costs 39,000 ₸ a month for the whole company — all channels are included, there is no need to pay for each channel separately. You can pay for one month or for 6 months at once with a 10% discount — in the Subscriptions section. Building the assistant, filling in every section and testing it in the sandbox is free: the subscription is exactly what lets it start replying to real customers.
Getting started
In the left menu open Automation → AI assistant and click Add an assistant — it's free. The settings are split into four groups: Setup, Scenarios, Knowledge and Connections. Sections marked with an asterisk are the ones the assistant will not work without — the rest can be filled in later, they add knowledge but are not required to launch.
The minimum to get the assistant working:
- Persona — who it speaks as. For example: "You are a sales manager at company X".
- Goal — what exactly it needs to find out from the customer.
- Qualification checklist — at least one item: what the assistant should find out during the conversation.
While these three fields are empty, you cannot save the assistant switched on — a hint next to the save button will tell you which sections are missing.
Don't want to fill in the fields one by one? Click Set up by text and describe your business in plain words — for example: "we are a dental clinic in Astana, open from 9 to 18, collect the name and phone number". The assistant will fill in the settings itself. You can ask for changes the same way: "let it write 'Good afternoon' instead of 'hi'". Any change can be undone.
Who it replies to and when
- Who gets a reply — "Everyone who writes" or "Only people we messaged first" (with a WABA broadcast or a message triggered by a new row in a Google Sheet). In the second case, a person who wrote on their own goes to an operator, and the assistant stays out of that conversation.
- Working hours — by default the assistant works around the clock. On the Schedule tab you can turn on Limit working hours and choose what it does outside them: stays silent, or replies but warns that a manager will join during working hours.
Setting its character
- Tone of voice — brief or detailed, formal or casual, emoji allowed or not.
- Guardrails — what the assistant must never promise. This rule outranks the scenario: it holds even when the customer pushes hard. For example: "never promise a discount or an exact delivery date".
- Additional instructions — anything that did not fit the other fields. For example: "the Abay branch is closed on Saturdays — offer Satpayev instead".
- Max questions in one message — not a suggestion but a guarantee: if the model tries to ask more, the extra questions are trimmed before the message reaches the customer.
0means no limit.
Guardrails hold more reliably than requests placed in other fields. If something is critical — "never quote a price over the phone", "never agree to a refund" — put it in Guardrails.
When a human is called in
The Hand off to a manager when field describes the condition under which the assistant stops running the conversation. For example: "the customer asks for a discount or a call", "asks for a company invoice", "complains".
There is also a quick way — Trigger words for handing off to a manager. If the customer's message contains one of these words (for example, "refund"), the conversation goes straight to a manager, without a reply from the assistant.
After a handoff the assistant leaves a note for the manager in the chat and stops replying in that conversation — from then on it is the human only. This is deliberate: a customer who has already been passed to a human should not start receiving robot answers again.
What happens when qualification finishes
When the assistant considers the task done, it leaves the manager a note with the collected data. You can also turn on:
- Write the form to a Google Sheet — a row is added to the sheet after every completed qualification. The columns come from the "Data to collect" section, and Google Sheets themselves are connected under "Connections".
- Add a CRM note — the summary lands in the lead or deal timeline.
- Also move the deal to a stage — you pick the pipeline and the stage.
Moving the deal overwrites its current stage in the CRM. If a manager is working that deal in parallel, the assistant can roll it backwards — enable this action deliberately.
CRM actions are available only after a CRM is connected in the CRM integration section.
Product catalog
If you sell products, the assistant can work as a salesperson using your catalog. The catalog is filled in under Knowledge → Catalog:
- Upload from a file — a JSON file with a list of products (name, price, SKU, category, stock, specs). Products with the same
offer_idare updated, new ones are added; you can also hide products that are not in the file. - Quotes and discounts — the address your website uses to build a quote from a list of products, a bundle discount (for example, 5% from 3 items), installment months and the markup on them, and the default city for checking stock.
- Kaspi payment (ApiPay) — your company's ApiPay key, so the assistant can issue a Kaspi invoice for the catalog amount by itself.
- Products — a searchable table: what's in the catalog now, with prices and stock.
To let the assistant use the catalog, turn it on in Connections → Assignments: the "Product catalog: search, prices, quotes" item and, if needed, "Send a Kaspi invoice". The assistant then finds products from the customer's description ("I need a 60 cm range hood, white"), names the price and stock, calculates a bundle with the discount and sends a link to the quote. It takes prices only from the catalog and never makes them up.
The catalog is saved with its own buttons, separately from the assistant's main save button. After uploading, check in the Test how the assistant picks products.
Testing before launch
The Test button opens the sandbox — a conversation with the assistant that no customer ever sees. In it, the assistant behaves just as it does with real customers: it uses its knowledge, asks the checklist questions, and hands over on your condition. The sandbox answers using the last saved settings — after making changes, save first.
Run three scenarios in the sandbox before switching it on:
- A customer who answers willingly and reaches the end of the checklist.
- A customer who immediately asks for a human — to confirm the handoff fires.
- A question the assistant has no answer to — to see whether it invents one.
Knowing that it works
Go to Analytics → AI assistant. The key numbers:
| Metric | What it means |
|---|---|
| Conversations | How many distinct conversations the assistant handled |
| Forms completed | Where it reached a complete form and left a note |
| Completion rate | What share of conversations reached a form |
| Handed to a human | Where the assistant went quiet and called a manager |
| Replies per conversation | On average. Too many means the customer got tired; too few means conversations were breaking off |
| Spend | What the model calls cost, in dollars |
A separate Build the summary button analyses every conversation in the period as a whole: the AI looks for patterns invisible in a single dialogue — where people drop off, what they ask about repeatedly, how things skew by city and time of day. The analysis takes up to a minute.
Assistant statistics start from 16 September 2026. For earlier conversations there is no technical way to tell an assistant reply from a broadcast, so they are not included.
Common questions
Does the assistant always reply instead of a manager? No. It stays silent in conversations already handed to a human or assigned to a team member, and if a manager wrote to the customer themselves in the last 5 minutes. It also gives way to auto-replies: if this message should trigger the "First message" or "Off hours" auto-reply, the assistant does not reply.
What if it does not know something? The more thoroughly the "Knowledge" sections are filled — reference tables, documents, frequent questions, exceptions — the less often it gets stuck. Check with the sandbox: that is where you see whether it invents an answer or honestly calls a human.
What happens if the subscription ends? The assistant stops replying to customers. The settings and the sandbox stay in place — nothing is deleted.
Can it be switched off temporarily? Yes, with the toggle at the top of the section — don't forget to save. The settings are kept — switch it back on and it continues with the same knowledge.