How to reach us

AI agents for business

The agent answers customers on WhatsApp, Telegram and your website, checks your CRM, ERP and documents, and takes the next step itself: opens a deal or hands off to a manager. What it may do alone is agreed before launch

Discuss the workflow What an agent can take on

WhatsApp

11:20

Hi! We need 12 Ergo Pro chairs for our new office by the end of the month. Can you make it? Any discount for the volume?

AI agent

Reads and acts

Reading the request

  • 12 × Ergo Pro
  • due: end of month
  • discount question

  • ERP

    Stock and delivery

  • CRM

    Customer record

  • Access and rules

Bitrix24 · Deal for a manager

Deal:
12 × Ergo Pro, by month end
Task:
approve the discount

Works with your tools

  • WhatsApp
  • Telegram
  • Bitrix24
  • 1С
  • Google Sheets
  • Notion

  • Acts, not only answers

    Writes to the CRM, opens deals and tasks

  • Rules set before launch

    Disputed cases go straight to a person

  • Quality tested up front

    On your real questions, before go-live

What an AI agent can take off your team’s plate

A good first workflow repeats every day, already has usable data, and ends with a result the team can check. We usually start an AI implementation project with a workflow like this

How the agent reasoned

AI agents already at work

Not concepts or demo prompts. These agents talk to real candidates, suppliers and customers, and write to live systems

  1. HR agent · WhatsApp · Magnum

    Magnum HR Agent

    An application in the HR system turns into a WhatsApp conversation. The agent runs the interview, picks the nearest store, explains the terms and books a trial day. The recruiter gets a candidate who is already screened.

    • 2–3 nearest stores with an open role offered to each candidate
    • 100 questions in the reference set the agent is tested on
    • Books the trial day right in the chat
    Read the case
  2. Onboarding agent · WhatsApp and Bitrix24 · Compass

    Compass

    A supplier scans a QR code at a Compass station and lands in WhatsApp. The agent collects company details, validates the tax ID, picks the category, accepts a price list in any format and creates a card in Bitrix24.

    • A QR code at every station, so the agent knows the city and site
    • Accepts price lists as PDF, Word, Excel or a photo
    • The supplier card goes straight to Bitrix24
    Read the case
  3. Support agent · WhatsApp, Instagram · Almaty Marathon

    Almaty 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
    Read the case

How an AI agent differs from a chatbot

The same customer writes to a store. The bot answers instantly with a menu. The agent answers like an attentive manager

Chatbot

Context Sees only the text and walks you through a menu
Action Doesn’t understand anything off script
Control Gives up and puts the customer in a queue

AI agent

Context Finds the order in the CRM and answers to the point
Action Moves the delivery itself and confirms it
Control Knows discounts are a person’s call and hands the question to a manager

Fits your systems, not the other way round

No need to replace your CRM or ERP. Before we build, we agree where the agent reads, where it writes, and what it does only with approval

  • WhatsApp
  • Telegram
  • Website chat

AI agent

Understands the request, decides where to look, and acts by your rules

OpenAI Anthropic
  • 1C, ERP stock and prices reads
  • Bitrix24, amoCRM customers and deals reads writes
  • Google Sheets price lists and rotas reads
  • Notion, Drive policies reads
  • Your API, database orders and statuses reads writes

You decide what the agent may do

Before launch we draw up the list: what the agent does on its own, where it asks for approval, and what always goes to a person

Does on its own

Repetitive and checkable

  • Answers about stock, timing and order status
  • Asks for missing details
  • Opens a deal and a task in the CRM
  • Books a meeting or an interview

Asks for approval

The action is ready, a person presses the button

  • Changes an order in the ERP
  • Sends an invoice or a contract
  • Reschedules a paid delivery

Always hands to a person

With the chat and the collected context

  • Discounts and off-list prices
  • Complaints, refunds, disputes
  • Legal and HR decisions
  • Any request to talk to a person

We build this list with you before launch. Every agent action is logged, so a disputed answer can be traced step by step

How we test the agent before launch

We agree the threshold before we start and test it on your real questions

A set built from your questions
100–200 real requests with reference answers from your staff. The set includes the awkward ones: typos, one-word messages, requests outside the script.
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.
Test set run 10/10 checked · graded by a separate model
  1. WA How much is delivery to another city? 0.92
  2. TG I want to book tomorrow after six booked and confirmed the time 0.88
  3. WEB Can we pay by company invoice? correctly handed to a manager 0.90
  4. WA Give me 20% off and I buy today judge: promised a discount outside the rules 0.64
  5. TG Where is my order? 0.93
  6. WA Get me a manager handed to a person right away 0.97
  7. WEB What warranty do you give? judge: the answer is not in the knowledge base 0.72
  8. WA Change the address in my request updated the deal in the CRM 0.85
  9. TG and if it does not fit can i return it 0.84
  10. WEB We need a contract and closing documents 0.86
Whole test set 0.85 8/10 passed the threshold average score · threshold 0.80

Illustration of the mechanics with sample data. The real set is built from your requests during the audit.

How we launch an AI agent

Five stages. After each one you decide whether to go on. Each stage is paid separately

Stage 1 of 5

Pick one process

2–3 working days
What we do

We find where the team loses hours and pick one process with an owner on your side and a result that can be counted.

What you get

A short memo: the process, the type of solution, what success means, and what stays out of the first stage.

Before the next 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 before an AI agent project

A chatbot usually follows a conversation script. An AI agent checks company knowledge and live data, then updates CRM, creates a ticket, or sends a notification. When the decision is risky or data is missing, it hands the conversation to a person with the context attached.

Start with a repeated workflow that has a clear outcome and usable data. Common examples include lead qualification, first-line support, HR screening, and internal document search. Important business decisions should still remain with people.

Yes. We build on top of the tools you already run: WhatsApp, Telegram, amoCRM, Bitrix24, 1C, Google Sheets, Notion, and internal APIs. We check the access model and limits of each integration before development. Early on we define what AI may write directly and what it should only suggest to a person.

We separate access by role, limit approved sources, and log what the AI does. Financial, legal, HR, and other sensitive decisions require a human handoff. Disputed answers are reviewed against real conversations and added to the test set.

A narrow working prototype can usually be tested in 1–2 weeks. An MVP with integrations, access roles, logs, and a team interface often takes 2–4 weeks. The exact timeline depends on data quality, API readiness, and the number of workflows.

An AI agent starts at $2,000. The total depends on the number of workflows and integrations, data quality, access requirements, and testing scope. Support after launch costs $200 to $1,000 a month. After a short workflow review, we fix the scope and price of the first stage.

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.

ChatGPT and similar tools help an individual employee write and search. Implementation connects a model to a company process: customer chats, the CRM, 1C and documents, with access rights respected. The system answers by your rules, writes results into your systems and hands difficult cases to a person. Quality is checked with a test set before launch and after every change.

The code, the knowledge base, the test set with reference answers, and documentation. Model keys and accounts are in your company’s name, so the system can keep evolving with us or without us.

Read before the call

Let’s talk

Send the workflow, the data sources, and the systems you need connected. We will estimate a practical first stage without hand-waving.

Azamat Galimzhanov
I read every request myself.
Azamat Galimzhanov Founder and tech lead Meet the rest of the team
Brief (optional)

By submitting this form, you agree to the privacy policy.