AI agents and automation

AI agents that work inside your systems

They answer customers in WhatsApp and on your website, create deals in Bitrix24, check payments in 1C and stock in your catalogue. We design around your processes, launch a pilot in 2–3 weeks and stay on for support.

Discuss the task See the cases

An AI agent connecting messengers, CRM, 1C and reports

Task analysis

Where an AI agent pays off fastest

We don't deploy AI for its own sake. We start with the process that has the most repetitive requests and a clear cost of error. Six areas where we already have ready scenarios, and the signs that it's time.

Sales and first enquiries

The agent replies within a minute at any hour, asks clarifying questions, qualifies the lead and creates a CRM deal with the full conversation attached. Hot leads go straight to a manager.

Time to act if: messenger enquiries wait hours for a reply, and half the questions are “how much” and “is it in stock”.

Customer support

Order status, payment, delivery, warranty, instructions — the agent answers from the knowledge base and your systems' data, and hands complex cases to an operator with a ready summary.

Time to act if: operators answer the same twenty questions, and customers get no reply in the evening or at weekends.

Online stores and catalogues

Product selection by parameters or by vehicle, stock checks, order contents and status from Kaspi or 1C, a link to the product and checkout right in the chat.

Time to act if: the catalogue has thousands of items and buyers can't find the right one without a consultant.

B2B and manufacturing

Quotes, specifications, invoices and documents: the agent collects the request parameters, checks availability, prepares a commercial offer and hands the manager a deal with the fields filled in.

Time to act if: managers spend days on routine quotes and re-sending invoices.

Internal processes

Staff routine: answers from internal regulations, reconciliations, document preparation, payment reminders, status checks in 1C and CRM without switching between windows.

Time to act if: employees spend longer searching chats and folders for answers than doing the actual work.

IT, hosting and service companies

The agent builds a dossier for each ticket — service, payments, server state, website errors — and drafts a reply for the engineer. That's how support works at our hosting brand ONEHOST.

Time to act if: engineers spend the first half hour of every ticket gathering context.

Architecture

How our AI agent is built

A chatbot with a prompt answers questions. An agent does the work: reads your documents, goes into CRM and 1C, performs actions by the rules and leaves a trace in the log. Here is what it consists of.

Diagram: communication channels, an AI agent with a language model and knowledge base, company systems and human control

The diagram scrolls sideways →

1Channels

Bitrix24 Open Channels, WhatsApp, Telegram, website widget, e-mail, internal staff chat. One agent answers in every channel and remembers the customer's history.

2Brain

Claude, GPT or a local model if data must not leave your perimeter. A vector-search knowledge base (RAG): regulations, price lists, documents, request history.

3Hands

Connections to Bitrix24, 1C, billing, Kaspi and the catalogue via APIs and MCP servers. We wrote the connectors to 1C, Bitrix24 and WHMCS billing ourselves and use them in our own products.

4Control

Every conversation and action is logged. An operator can step in at any moment. The agent never performs irreversible actions without an employee's confirmation.

Stack:Claude / OpenAIlocal modelsn8nQdrantMCPPythonPHPBitrix24 REST1CHosted inside your perimeter or in the ONEHOST data centre in Kazakhstan.

Cases

What already works for our clients and for us

Not demos or slide decks: the agents below handle requests every day. Each is connected to the client's CRM and accounting systems.

~200tickets a month

ONEHOST hostingBillingZabbixClaude

Assistant for support engineers

For every ticket it builds a dossier — the client's service and payments, server metrics, website availability, certificate expiry, DNS — and drafts a reply with a diagnosis for the engineer. The knowledge base holds 6,800 closed tickets. A human always answers the customer.

Bitrix24Laximo catalogueRAG

Car parts lookup by VIN from a photo of the registration document

An assistant inside the Bitrix24 deal card reads the VIN from a photo, searches parts by catalogue and part number, checks availability and answers customers from the company's documents.

View case →

WhatsAppKaspi APIOnline store

Virtual tyre and wheel salesperson

Selects by vehicle or size, reports order contents and status from Kaspi, advises on delivery and warranty, books tyre fitting and hands over to an operator on request.

View case →

Bitrix24BillingWhatsApp

Hosting customer assistant: services, invoices, tickets

After e-mail verification it checks services and payments, opens tickets, sends invoice reminders in WhatsApp, calculates a server configuration.

View case →

All ONELAB projects →

Examples

What a conversation with the agent looks like

Three typical scenarios from our projects. In each one the agent doesn't just reply — it writes to a system or pulls data from it.

Enquiry → CRM deal

Hello. I need a quote for steel structures, roughly 12 tonnes.

Good afternoon! Please tell me the delivery city, your preferred timeline and a contact phone — I'll prepare the request for a quote.

Almaty, two weeks, +7 7XX XXX XX XX.

Thank you! I've passed the request to a manager, who will contact you today before 18:00.

Result: a contact and a deal are created in Bitrix24, the conversation is attached, and a task is assigned to the manager.

Payment status from 1C

Good afternoon. Could you check the payment status for contract No. 4587?

Checking the accounting system… Payment for contract No. 4587 was received on 12 January, status “Paid”. The order is in production, planned shipment 20 January.

Could you send the invoice again, please?

I've sent invoice No. 10234 to this chat. If you need the completion certificate, I can send that too.

Result: the data and the document came from 1C and CRM; no manager was involved.

Product selection by parameters

I need M12 bolts, grade 10.9, in stock.

M12×40 and M12×50 grade 10.9 are in stock. Shall I send prices and stock for both?

Yes, M12×50.

M12×50, 10.9 — 1,840 pcs in stock; the product link and price are below. Place the order or hand over to a manager?

Result: the agent works with the catalogue and stock; the customer gets a reply at 22:47 with no waiting.

Rules that are not negotiable

An agent with access to your data is a responsibility

We run hosting and billing for hundreds of clients ourselves, so our security rules existed before our first agent. They are built into every deployment.

Your data stays with you

We host the agent inside your perimeter or in the ONEHOST data centre in Kazakhstan. For cloud models, personal data and secrets are stripped before sending; where the requirement is strict, only a local model is used.

Conversations are data, not commands

A customer's text never becomes an instruction for the agent. The address to check, the service and the amount come from your systems, not from the message. That is how we close prompt injection.

No place for secrets in the model

Passwords, keys, card numbers and tokens are cut out of requests before the model sees them. The rule came from real support conversations, where customers send credentials as plain text.

Irreversible actions — humans only

The agent performs whitelisted actions: create a deal, send an invoice, lift a suspension after payment. Deletion, refunds and permission changes are confirmed by an employee. Everything is logged.

Where to start

How we can work together

We start small and measurable. The pilot shows numbers on your data; after it, the decision on full deployment is made on facts, not on a presentation.

Pilot · 2–3 weeks

One process, one channel, your data

We pick the scenario with the largest flow of repetitive requests, build a knowledge base from your documents and launch the agent in one channel.

a working agent in WhatsApp, on your website or in Bitrix24;
a knowledge base built on your regulations and price lists;
a report: response time, share of conversations without an operator, cost per conversation.

Start a pilot

Deployment

An agent built into your systems

Integrations with Bitrix24, 1C, billing, the catalogue and messengers. Least-privilege access, an action log, operator hand-off scenarios, team training.

timeline and price are fixed after the process analysis;
staged delivery: we show results every week;
documentation and credentials remain with you.

Support and development

The agent gets better every month

We review the conversation log, extend the knowledge base, add scenarios and channels, and watch model costs and answer quality.

a monthly metrics report;
incident response and integration updates;
new scenarios without stopping the ones that work.

What we measure in every project

Time to first replybefore and after the agent goes live

Conversations without an operatorshare of requests closed by the agent

Conversion to dealfrom enquiry to CRM deal

Cost per conversationmodel spend per conversation

How we work

Six steps to a working agent

Timings are for a pilot. A deployment with several integrations follows the same steps, but each takes longer.

Process analysis 1–2 days

We look at real conversations and requests, count where the routine is heaviest, choose the pilot scenario and the metrics we'll judge the result by.

Design 1–2 days

Conversation scenarios, operator hand-off rules, the list of actions the agent performs itself, data and security requirements.

Integrations 3–4 days

We connect channels and systems: Bitrix24, 1C, billing, the catalogue, messengers. Access is granted with least privilege and documented.

Knowledge base and training 3–4 days

We gather documents, regulations, price lists and request history, test answers on your customers' real questions and tune the wording to your company's style.

Test and launch 1–2 days

A run on live conversations in shadow mode, where the agent drafts replies and an employee sends them. Then launch in one channel, then in all.

Ongoing support

We review the log, extend the knowledge base, add scenarios and channels, and report on metrics monthly.

FAQ

Frequently asked questions

Which models do you use and where does the data go?

We choose the model for the task and requirements: Claude, GPT or local models deployed inside your perimeter or in the ONEHOST data centre in Kazakhstan. For cloud models, personal data and secrets are stripped before sending. If data must not leave the company, we work only with a local model.

What about errors and “hallucinations”?

The agent answers from the knowledge base and your systems' data, not “from memory”. When it isn't sure, it says so and hands the conversation to an operator with a summary. Every reply is logged: we review disputed cases and add them to the knowledge base, so quality grows month by month.

Can the agent break something in CRM or 1C?

No. It only performs actions from an agreed whitelist with minimal permissions: create a deal, send an invoice, book a service. Irreversible actions are confirmed by an employee. A customer's text never turns into a command for the agent.

How much does it cost and how long does it take?

A pilot takes 2–3 weeks and has a fixed scope, so we quote its price right after a short review of the task. A full deployment is estimated after the design stage, once the list of integrations and scenarios is clear. Model costs are counted separately: usually tens of tenge per conversation.

Do I need Bitrix24 to start?

No. The agent works with any system that has an API, and with messengers alone. With Bitrix24 and 1C it's faster: we have ready connectors and hundreds of deployments behind us, and Open Channels give the agent every channel at once.

How is this different from CoPilot in Bitrix24?

CoPilot is an employee's helper inside the portal: it suggests text and summarises. Our agent works with customers and with your systems: it replies in messengers, goes into 1C and the catalogue, performs actions by your rules and runs on the model you choose, including a local one.

What happens after launch?

In the first month we watch especially closely: reviewing the log, adjusting scenarios, extending the knowledge base. After that — ongoing support with a monthly metrics report, or a hand-over to your team with documentation if you want to develop the agent yourselves.

Discuss the task

Tell us which process you want to offload

No specification needed. Describe where requests come from and what happens to them now — at the first meeting we'll propose a pilot scenario and metrics.

We reply within a business day and offer a meeting online or at our office in Almaty or Astana.
We'll show the working agents from the cases above on live data.
We name the pilot price after the first meeting.

Have a specification or an export of conversations? Send it to info@oneit.kz (PDF, DOCX, XLSX up to 20 MB) — and mention it in your message.