To automate routine tasks in a company, there is no need to overload the technical department's backlog. The experience of Siberian developers has proven that if you provide non-technical specialists with basic knowledge of working with large language models, a bottom-up initiative can reduce the routine workload of entire departments by almost half.
Format: Workshops Instead of Boring Lectures
Top-down attempts to enforce the use of neural networks rarely yield results. Siberian.pro took a different approach by launching weekly, one-hour experience-sharing sessions ("AI clubs") with strictly voluntary attendance.
The meetings were divided into two independent tracks:
- For developers — focusing on in-depth technical nuances of code generation and integrations;
- For non-programmers (marketing, sales, HR, accounting) — emphasizing solutions for practical office tasks.
Participants analyze real-world use cases, review colleagues' homework projects, and test new tools under the mentorship of senior specialists.
From Prompts to Custom AI Agents
Training is built on the principle of gradual complexity. At the start, participants learn the basics: how to use editor interfaces like VSCode for everyday tasks, comply with data security standards (including regulatory requirements like 242-FZ), and understand why confidential information should never be uploaded to public cloud chats.
Next, the program moves on to practical application:
- Exploring extensions and auxiliary tools: RooCode, Kilo Code, MCP protocols;
- Building custom skills for autonomous AI agents (Hermes, zeroclaw);
- Automating specific processes — from automated documentation proofreading to time tracking.
A prime example: marketers and accountants began independently building scripts and even small macOS applications that fully cover their repetitive routine tasks without involving in-house software engineers.
All successful solutions, once verified by the technical team, are submitted to a shared corporate Git repository. Today, the knowledge base includes dozens of ready-made AI skills available to every employee in the company.
The Cost and Measurable Results
The infrastructure costs for maintaining the ecosystem turned out to be remarkably modest. The core stack relies on local solutions via OpenWebUI and LiteLLM:
- Around $80–100 per month is spent on LLM subscriptions for the non-technical team;
- Up to 10,000 rubles per month goes toward infrastructure expenses;
- 8–10 hours of a Senior-level expert's working time per month is dedicated to mentoring.
With these minimal costs, this systematic approach delivered tangible economic impact over the course of a year:
- 80% of non-technical staff began regularly using AI tools in their daily workflow;
- Data collection and template preparation accelerated by 30–40%;
- Pre-sales data preparation time dropped from 12 hours to 3–4 hours;
- Routine workloads in accounting and HR decreased by 40%;
- The marketing department saves a total of up to 200 work hours each month.
Human Always in the Loop
The most important principle of internal training is the human-in-the-loop concept. Neural networks are positioned strictly as working tools, not human replacements. Direct automated publishing and document approvals without manual verification are prohibited. Mandatory fact-checking by an employee eliminates risks related to language model hallucinations and fosters a responsible, conscious approach to generative technologies.