AI for customer support
The agent takes the first line on WhatsApp, Telegram and your website. It answers from your knowledge base, checks the order in the CRM and asks for what is missing. Refunds, complaints and disputed cases go to an operator along with the chat and a short summary
WhatsApp · customer
14:07Hi! I bought a robot vacuum from you last week and it stopped charging. Is that covered by the warranty? Can I swap it for a new one?
Support agent
Reads and acts
Sorting the ticket
- not charging
- warranty question
- exchange request
CRM · Customer order
Customer order
Knowledge base · Warranty and exchange
Warranty and exchange
Agent permissions
Reply to the customer
14:08Hi Dana! The vacuum is under warranty, and in the first 14 days it can be exchanged after a check. An operator handles exchanges: Asel will message you here within the hour. Meanwhile, please send a photo of the serial number sticker.
Bitrix24 · Exchange for an operator
- Ticket:
- warranty exchange, order 7302
- Collected:
- model, purchase date, fault
Works with your channels and CRM
-
Answers from your knowledge base
No answer there means it calls an operator
-
Operators get the context
The chat, the order and the reason for handoff
-
Tested before launch
On your past support tickets
What the agent takes off your support team
Start with what repeats every day and already has a written answer in the knowledge base, the CRM or your policies. This part of support is a common first step in implementing AI across the company
Where the agent looked
Support agents already at work
One agent answers race participants, the other answers staff from internal documents and links the source. Support operators can work the same way
Scenario transcript
Question
- Customer: Hello! I’d like to sign up for the marathon
- Under the hood: Race calendar. Knowledge base, races 2026, Almaty Marathon 2026: registration, Astana Half Marathon 2026: registration, Tau Jarys 2026: registration. Three races on the calendar
- Agent: Good afternoon! I’m the Almaty Marathon assistant. Which race would you like to know about? The calendar now has Almaty Marathon 2026, Astana Half Marathon 2026 and Tau Jarys 2026.
Race kit
- Customer: Almaty Marathon. Where do I pick up my race kit?
- Under the hood: Race kit pickup. Knowledge base, Almaty Marathon · race kit, Where: pickup point, How: QR code from the email. Pickup point, QR code
- Agent: Race kits are handed out at the pickup point before the race. Show the QR code from your registration email.
Refund to a manager
- Customer: And if I can’t run, will I get my money back?
- Under the hood: Agent permissions. Refunds: manager. A person decides on refunds
- Under the hood: Handoff to a manager. WhatsApp, To: manager, Summary: race, kit, refund. The manager has the summary
- System: Refund → manager · chat summary attached
- Agent: I’ve passed your question to a manager. They’ll reply right here.
First-line support · WhatsApp, Instagram · Almaty MarathonAlmaty Marathon
The agent answers participants on rules, schedule, registration and race packs. Refunds, complaints and VIP requests go to a manager with a summary of the chat.
- 9 in 10 conversations resolved without an operator
- ~2,130 conversations with race participants
- WhatsApp and Instagram around the clock
- Screen recording from the live botKnowledge base assistant · Telegram · Magnum
Olzhas: answers for Magnum staff
An employee asks in their own words: how to request leave, who takes the sick note, where to get an employment certificate. The assistant finds the answer in approved documents and shows which document it came from.
- 100 real employee questions in Russian and Kazakh, run against every update
- The team uploads and updates documents in an admin panel, no developers needed
- Understands Russian, Kazakh and a mix of both
Connects to where your support already works
No need to move to a new support system. The agent takes answers from your knowledge base and statuses from the CRM or ERP, and files tickets where operators already work
- WhatsApp
Telegram- Website chat
Support agent
Finds the answer, asks for missing details and passes to an operator what a person should decide
OpenAI
Anthropic- Bitrix24, amoCRM tickets and customer history reads writes
1C, ERP orders, payments, warranty reads 
Notion, Google Docs knowledge base articles reads 
Google Sheets rates and on-call rotas reads 
Your API, database order and request statuses reads writes
Where the agent answers and where it calls a person
We write the limits down before launch. Otherwise the team ends up arguing about them in front of live customers
Answers on its own
The answer is written down and checkable
- Opening hours, addresses, delivery terms
- Order or request status by number
- Warranty terms and required documents
- A link to the right guide
Prepares, an operator decides
The draft and the data are ready, a person sends
- A warranty exchange
- Compensation or a bonus for a mistake
- Moving a paid service to another date
Hands to a person right away
With the chat and a short summary
- Refunds and payment disputes
- Complaints about staff
- Threats, health, safety, legal claims
- A request for an operator or the same question a third time
Once an operator takes over a conversation, the agent stays out of it. Every agent reply is saved with its source, so a disputed case can be traced step by step
Testing answers on your own tickets
First the agent silently answers past tickets, and we compare its replies with your operators’. The threshold is agreed before we start
- A set built from past tickets
- 100–200 requests from your support with your operators’ best answers. The set must include angry customers, one-word messages, outdated rules, and questions where the right answer is “passing you to an operator”.
- A different model grades
- A separate judge model, not the one that answered, scores the answers. We check its scores against your staff’s own grades.
- Every miss gets reviewed
- An overall score can hide individual errors, so every answer below the threshold is reviewed on its own and fixed before launch.
- Acceptance on new questions
- We sign off on questions the system never saw during tuning, so the result can’t be fitted to familiar ones.
- WA Why am I not getting the login code? 0.90
- TG internet is down router blinking red found the fault at the address 0.87
- WEB How do I change my plan? 0.93
- WA Refund me, the item arrived broken handed to an operator right away 0.95
- TG Can I extend the warranty? judge: made up a condition not in the rules 0.62
- WEB Can’t export the report, error E-502 collected the details and opened a ticket 0.88
- WA How long until the courier comes? 0.84
- TG Third time you’re not answering my actual question judge: replied with a template, did not call a person 0.74
- WEB Where do I get a reconciliation statement? 0.86
- WA vacuum not charging is that warranty 0.89
Illustration of the mechanics with sample data. The real set is built from your requests during the audit.
How we launch a support agent
Five stages. After each one you decide whether to go on. Each stage is paid separately
Pick one process
2–3 working daysWe find where the team loses hours and pick one process with an owner on your side and a result that can be counted.
A short memo: the process, the type of solution, what success means, and what stays out of the first stage.
Process, metric and owner are agreed.
Timelines start once access is granted: without chat exports and a test API the audit cannot start. WhatsApp number verification with Meta takes from a few days to three weeks, so we start it on day one. Projects with several processes or heavy integrations with 1C and custom systems take 3–4 months, but even there a working version ships in the first month.
Questions heads of support ask
Yes, and it is often the safer start. The agent drafts a reply with a link to the source, and the operator sends or edits it. The edits show which topics the agent already handles well, and those open to customers first.
It stops and moves the conversation to your CRM or a Bitrix24 open channel with a short summary: what the customer wanted, what they have already been told, which details are collected and why a person is needed. The operator does not have to reread the chat from the start.
Every section needs an owner on your side. Articles are edited where they already live, in Notion, Google Docs or an admin panel, and the agent picks up the changes. If the rules change every week and the knowledge base does not, the agent will keep repeating the old answer with confidence. That is why questions it could not answer go into the monthly report.
Yes, when answers are built on a knowledge base or RAG. Sources are usually useful for internal operators. For customers, they can be shown selectively or kept for quality control.
It does not decide them. It collects the facts: order number, what happened, a photo if one is needed. Then it passes the case to an operator. The agent can draft the reply, but a person sends it.
From $2,000. The first working version ships 2–4 weeks after access is granted. The price depends on the number of channels and topics, the CRM and ERP integrations, and how much work the knowledge base needs. Support after launch costs $200 to $1,000 a month: test set runs, reviews of disputed answers and knowledge base updates.
Yes. We choose a reliable integration path and discuss message templates, limits, consent, conversation storage, and when the dialog should move to a person. Employees’ private chats stay separate from the work system.
A process owner who makes decisions, 30–100 real requests, conversations, or documents, a list of systems with access to their test APIs, handoff rules, and examples of good team answers. The knowledge base does not have to be perfect: we build the test set together.
Worth reading before the call
- AI for support teams Where to start: the ticket queue, drafts, limits of automation and what to measure
- How RAG reduces support load Why uploading PDFs into a chat is not enough and who should own the knowledge base
- How AI answers customers in WhatsApp What to answer automatically and when to pass the chat to a person
- What are AI evals and why projects need them How to tell the agent is ready to answer customers on its own
Let’s talk
Tell us what customers ask most often, in which channels, and where the answers are written down. We will suggest which topics to give the agent first and estimate the pilot.
I read every request myself.