How a lawyer from Kazakhstan without programming skills assembled an AI team for a company

Zhangali Baigaliev, city — Almaty, legal counsel at Bitrix24 Kazakhstan 

Why a lawyer needs their own AI

I am not a developer. I work with contracts, consultations, and legal risks. But today, “AI Lawyers KZ” is already working on our company’s portal — a fully functional multi-agent AI assistant for the legal department, which I built myself. And I didn’t write any code manually.

Requests to the legal department rarely come in neatly. More often, they are raw questions where facts are mixed with guesses, and the answer is needed “yesterday.” Public chatbots are not the best helpers here: a neural network might confidently cite an outdated version or an article that doesn’t exist. In law, such an error costs more than no answer at all.

I needed a tool that works within our workflows and thinks like a lawyer: first checks, then approves.

Where it all began

In our company, vibe coding is not just allowed — it is encouraged. We were given video lessons and access to tools, and then an external trainer was invited. In an hour, he showed how to vibe code from scratch and shared prompt examples for every area of work — from accountants to content managers.

The most interesting part started after that. We were given an hour of practice, and in that hour, every employee created a useful application or bot for their field. For example, our marketing director built a dashboard for marketing and PR metrics: all current projects, competitor analysis, budget spending, key performance indicators — everything on one screen. For me, this hour became the starting point: it became clear that you can build a tool for your own tasks yourself.

I built my system using Cowork/Code and Vibecode — tools from the Bitrix24 ecosystem for creating applications with the help of AI. Access to them was already provided in our company, so I didn’t have to separately configure servers and the technical environment for development. But this is not the only option. Any tool with code access will do: Claude Code, OpenAI Codex, or similar. The choice depends on what the team already uses and how much they are willing to pay for a subscription.

How “AI Lawyers KZ” is structured

In essence, it is a small legal department that works without days off and responds in minutes. You send a request — and the system itself decides which specialist will take it on.

Inside is a team of agents, each with its own role:

— dispatcher-orchestrator — analyzes the request and decides which agent to pass it to

— lawyer — answers general legal questions

— labor expert — handles labor law questions

— Research — searches for and verifies norms from official sources

— Contract — works with contracts

— QA — checks the final answer before it goes to a human.

This is not 10 separate chatbots. One request goes through a chain of roles: the dispatcher determines the route, the specialized agent conducts the analysis, Research confirms the legal basis, and QA checks the result before issuing it to the user.

The main thing in the system is answer discipline. It separates fact from norm, and interpretation from assumption. If a norm could not be confirmed from an official source, the answer says so. Unverified information cannot be presented as reliable.

The system is integrated into our work messenger: it has a bot. Colleagues don’t need to learn a new service — they write there, just as they always did.

I didn’t write code — I set the task

The most unexpected thing for me was the distribution of roles. I described what should be achieved, and the AI agent itself designed the structure, wrote the code, ran tests, and brought the program to launch on the work server. I didn’t need to choose a programming language or technical templates.

Setting a task for an agent turned out to be very close to legal work. It is like a technical specification for a contract: who is the user, what do they input, where to get data, by what rules to process it, what result is needed, and what to do in case of an error. The more precise the subject, the fewer disputes during acceptance.

Currently, the project has 86 automated tests, and the entire test suite passes successfully. They check key system operation scenarios: request distribution, agent interaction, document processing, work with large language models of neural networks, and behavior in case of errors.

You act as the architect and project manager, and the AI is your engineer.

Honestly about mistakes

Vibe coding is not a “make it beautiful” button. There were many mistakes, and it was precisely they that provided the most valuable experience. Three that stood out the most:

1. “Floating” tests. The same automated check would sometimes pass successfully and sometimes produce an error. The reason was a data race: checks were launched simultaneously and interfered with each other, like two lawyers editing the same contract file at the same time. We had to figure it out and separate them into individual “offices.”
2. A real model is not a training mockup. On test artificial data, everything worked, but the “live” neural network behaved differently. It returned data not in a strict scheme, added extra service symbols, and called tools in the wrong format. The system had to be taught to digest such answers. A lawyer would say: contract conditions are good on paper, but tested in execution.
3. Launching on the server gave a slap. Environment settings, platform gateway, authorization — at every step, something didn’t match up the first time. But every time the problem was solved, and each time I better understood how what I was doing was structured.

Where to start

My advice to colleagues who want to try:

1. Start with one task. Let it be something you regularly do manually. A multi-agent system can wait.
2. Write the task like a technical specification. What goes in, what comes out, what rules, what to do in case of an error, and what the agent is not allowed to touch.
3. Build verification into the system itself. Official sources, a separate controller agent, tests.
4. Don’t be afraid of mistakes. They will happen, and they are solvable.

Why vibe coding will be useful to every employee

After this project, I am convinced: vibe coding is not a skill for IT specialists, but basic literacy for a modern employee. And here is why.

The tool is made by the person who knows the task best. Previously, between “I need” and “ready” stood a technical specification, a queue to developers, and a retelling with losses. In the case of legal tasks, this is especially critical: it is important to take into account context, legal risks, document structure, and source requirements. Now, the tool can be assembled by the specialist who works with these processes daily and understands what the result should be. This is why “AI Lawyers KZ” was initially designed not as a universal chatbot, but as a tool for the real tasks of a lawyer.

This is a skill of setting tasks, not programming. Clearly formulating a goal, input data, rules, and result criteria is useful in any profession. Vibe coding trains this every day.

Critical view of AI. Those who assembled the agent themselves see where AI makes mistakes and know how to build verification. When AI is everywhere, this is a real competitive advantage.

The experiment costs almost nothing. If it doesn’t work out — an hour is lost, not a budget and months of development. So try boldly and confidently!

The entry barrier is low, and the result is a real working product. And if a lawyer was able to assemble a team of AI agents for their department, then an accountant, HR specialist, or sales manager will definitely manage.

About this, AI CRM CONF 2026 will be discussed by Alexander Vartanyan, Director of Marketing and AI/Martech at Bitrix24, in the session “Vibe Coding: How Business Will ‘Program’ Itself,” on October 15. I built a system for the legal department; someone else will build it for their routine — the principle is the same.

Vibe coding is not magic, but a new way of working. The specialist brings their subject knowledge and demandingness for the result, and the AI takes on the engineering part. For a lawyer, this is especially natural: we already formulate conditions, check execution, and look for weak points every day.