Let me share a case that clearly shows what the path from "I want AI" to a working solution looks like in practice.
The client is an accounting outsourcing company: 13 employees, 150 clients on service, 10+ years in the market.
By the time we started working on a specific solution, the CEO — let's call her Anna — and I had already completed a full audit of all business processes, and we had several AI-ready tasks on hand. But developing for each area separately would take weeks and a substantial budget.
I proposed a different path.
A good first step for AI implementation is to digitize accumulated knowledge. A digitized knowledge base provides a solid foundation for future AI initiatives.
What kind of knowledge?
Over the years, companies accumulate knowledge and experience in documents on network drives and paper. Finding anything when needed is nearly impossible.
And worst of all — when knowledge and experience live in employees' heads. If they leave, the knowledge leaves with them. Digitization solves these problems.
Preparing for digitization
I told Anna about the capabilities a "living" knowledge base unlocks, and her first words were:
"I want my employee to be able to just ask the AI a question in Russian and get an answer from our knowledge base — instead of trying to find it in folders, go crazy, and give up."
But first we had to decide what exactly to digitize. We got on a call, Anna shared her screen, and we went through Yandex.Disk together. Folder by folder, she commented, I asked questions — a full picture emerged as we went.
After settling on the data from Yandex.Disk, we moved to PlanFix. Hundreds of tasks, the company knowledge base: instructions, regulations, organizational documents. And client cards — 150 companies with their entire interaction history.
This is where the gap between how you imagine your processes and how they actually work surfaced.
Anna knew there was a lot of data. But for the first time she saw the full picture:
- Instructions exist in two systems simultaneously: newer versions in PlanFix, old ones on Yandex.Disk. Employees habitually go to Disk.
- 5% of files are broken PDFs that can't be opened.
- Some instructions haven't been updated in years, but employees keep working from them, just guessing what changed.
- Contract templates thought to be current are stored in a folder marked "old, but let them be."
Together we set priorities and ended up with 58 knowledge domains — from taxes and accounting to client onboarding and sales scripts. What to digitize now, what later. Client data stayed outside the perimeter.
Call recordings we decided to process through a local model to keep everything inside. We didn't digitize video courses but semantically analyzed their descriptions — so the agent knows where things are.
Building the infrastructure
To avoid overloading my own machine and keep everything within the client's environment, I asked their admins to set up a remote workspace for me. That's where I did the digitization. For employee access, I set up a virtual server with Docker, installed Qdrant for knowledge base vectorization, and a chat interface to work with it.
The result: 28,000 search chunks, 58 file collections — one per domain.
Here's how it all works:
An employee asks a question, the assistant searches the entire base and returns an answer with the source: domain, file, fragment.
A personal "ChatGPT" over your own knowledge base.
One of the most curious moments came when Anna opened the chat and asked a question. The assistant answered with a citation and document reference. Anna said:
"I didn't even know we had that written down anywhere."
The first solution built on the digitized KB
Next, we tackled commercial proposals.
On the client's server, I deployed Tsekh. Loaded in service descriptions, pricing, and the company's brand style. Copied the proposal creation logic and the service selection matrix based on the client's tax regime. Described the task execution rules and packaged it all into a proposal agent.
How proposals are created now:
- Anna's manager uploads client data to the chat.
- Tsekh pulls in the proposal agent and builds the structure.
- Anna approves.
- Tsekh pulls in the design agent, which rebuilds the proposal in brand style: fonts, colors, logo.
Total: 20 minutes for what used to take 2–3 days.
A few days ago, Anna's client received the first maintenance proposal — 256,000 RUB/month. Built by agents, approved and sent by a human.
Of course, there were failures
DeepSeek incorrectly digitized 7% of data on the first pass. To fix it, I downloaded a local AI model: 300 MB, runs directly on the client's machine, no data leaves the premises. We re-digitized.
Another snag: the chat interface didn't show the source document. Employees could see the answer but couldn't open the source file. Fixed that too — essentially, I wrote a script to reassemble the document from search chunks.
Honestly, I thought digitizing this volume would take a month. In reality, it took a week.
Tsekh deployed agents on the fly: one digitizes, another verifies, a third vectorizes. I can deploy Tsekh onto a client's machine, do the work, and remove it. Or leave it, if needed. A year ago, this simply wasn't possible.
Anna and I have already planned two more agents: a contract agent and an agent that listens to client chats and automatically enriches the CRM client card. Employee training is planned for each completed solution. The project is alive and growing layer by layer.
I started with what I always recommend: a knowledge base as the first step. This project is a textbook example. When the foundation is digitized, agents stack on top in hours. That's a production line.