A Company's Immune System and 7,522 Hours Saved: What Actually Worked When Rolling Out AI at Alpina

• Task breakdown across the company: 69% through the AlpinaGPT platform, 26% through autonomous agents, 5% hybrid
• The gap between proficiency levels: 158 hours saved per year for orchestrators versus 13 hours for beginners (12x)
• The leader in time savings converted to money: the sales department (about 3.5 million rubles a year), with development in second place
• Confirmed volume: 910 documented cases from 99 employees, 7,522 hours after a cutoff at 66%

In the publishing business, a book takes roughly nine months to come to life, and most of the money is spent long before the print run ships. Any speedup in the production chain directly frees up working capital. When we started weaving artificial intelligence into the processes of the Alpina publishing group back in 2023, there were no ready-made corporate playbooks — every hypothesis had to be tested on real people.

Today, out of more than three hundred employees, about 250 regularly solve tasks in AlpinaGPT, and 130 people are building their own working tools in Claude Code, including the executive board. Let me walk through where we hit the biggest snags and how our evaluation methodology works.

Why do employees slide back into old habits even with free access to

the best models?

If a manager just sends the team a cheerful email with passwords to a neural network, 90% of people will be back to their usual routine within two weeks. The organization's immune system kicks in. Employees aren't resisting the technology itself: an experienced specialist doesn't want to feel like a trainee again, someone else worries about their own job security, and someone else got a made-up link from the model once and closed the tab for good.

We've shut down failed experiments ourselves. For example, we tried to catch narrator slip-ups in audiobooks using automatic speech recognition, but the language model would guess at missing words based on meaning precisely where we needed it to catch the error.

Distrust can only be cured by leading from the top by example and by setting aside dedicated practice time. I started by working through personal work tasks with members of the board of directors and the CEO. Once the CEO personally built a working librarian-agent prototype in one evening and handed it to the engineers as a live spec, the question of "why should we give people hours for training" disappeared on its own.

What lets orchestrators save 12 times more time than average chat users?

Inside the company, we split employees into four maturity levels: beginner, user, operator, and orchestrator. The data cutoff showed that one universal course for everyone doesn't work: a beginner needs to lock in at least one weekly operation, while an advanced user needs permission to run scripts and integrations.

The numbers reveal enormous stratification: an average chat user frees up 13 hours a year on average, while an employee at the orchestrator level saves 158 hours — a twelvefold difference. To pull the team up, we designated 21 internal champions, each of whom documented at least 10 of their own cases. When a marketer or marketplace manager sees a colleague down the hall build a "Marketplace Pulse" dashboard with order, margin, and inventory summaries without any programmers, the fear barrier disappears.

How do you structure the first 90 days of AI adoption without bloating

headcount or wasting money?

According to our audit, the top spot for the cost of freed-up time went not to developers but to the sales department — about 3.5 million rubles a year from automatically parsing calls, checking scripts, and generating draft commercial proposals from meeting transcripts.

If you're launching such a setup from scratch, break the first 90 days into three clear phases: 1. Month one — inventory: pick 3–5 narrow processes together with managers where the result can be measured in hours and money, and record a baseline. 2. Month two — applied hackathons on employees' real tasks, with mandatory testing on tough edge cases. 3. Month three — a control measurement of cost, the survival rate of the tools built, and the quality of the results.

I post honest breakdowns of corporate case studies and ROI calculation methods in the channel [@jemal_hamidun](https://t.me/jemal_hamidun). For companies that need a single managed framework for working with agents, we built [Nomad AI](https://nomadnet.ai). You can go from user to orchestrator at [Hamidun Academy](https://academy.hamidun.com) (there's a free 72-hour trial), and dig into implementation architecture at masterminds and pick up 370+ skills at [Vibe Club](https://hamidun.com/club).