1C Developers Switch to AI Agents: McKinsey Sees Just a 3% Productivity Gain
A 20-person team writing for 1C, the Russian business accounting and automation platform, moved to AI agents over six months. Manual coding dropped sharply, while code review grew. McKinsey estimates the average productivity gain from AI at just 3%, and at 30% of teams productivity actually fell. So the team first mapped out who reviews AI-generated code and when, and only then turned to hardware.
AI-processed from Habr AI; edited by Hamidun News
When AI agents join a team, everyone asks the same question: what should developers do now? The head of a team of roughly twenty people published a report on Habr describing how they made this transition over several months. The team includes developers working on 1C (a Russian platform for business accounting and automation) and adjacent stacks, plus testers and analysts.
The author compared their experience with industry data. The most sobering figure comes from the consulting firm McKinsey: the average productivity gain from AI is about 3%, and at 30% of teams it actually declined. McKinsey's conclusion: the winners are not those who bought the tool, but those who rebuilt roles and processes around it.
That is why this team's story is useful as an example of where to start.
How it developed in the team
There was no plan. In spring 2026, some colleagues simply tried AI agents on personal projects. By summer it had become a working practice: manual coding dropped sharply, but the load on code review and architectural decisions grew. By late summer the team began experimenting with local generation on its own hardware and assembled a shared knowledge base, from which a separate 1C knowledge base emerged. Knowledge-sharing meetings on AI tools are held once a week. Without them, according to the author, it would have been hard to keep up the pace.
In the autumn, they rolled out a cloud agent as the main tool for the whole team. Before that, passwords, keys and other sensitive data were scrubbed from the code. Around the same time, they launched a pilot of spec-driven development: the specification becomes the main artifact, and the code is derived from it. To evaluate models, they used the PRISM benchmarking project, so as not to rely on marketing promises and instead run models on their own tasks.
There is also an unpleasant finding: some team members reported worse sleep and well-being amid the rapid immersion in the technology. The author believes this has to be factored into the pace of adoption.
What the industry says
Gartner predicts that most code will be written by AI agents and checked by experienced developers. By its estimate, 80% of engineers will need retraining by 2027. IDC describes the same shift as a move toward supervision: setting the task, checking the result, and refining the agent's output. A few more figures from the review:
- The share of local and open-weight models (with publicly available weights) in token traffic for corporate AI, according to a survey of 200 companies, grew from 23% to 34% over the year. Vercel, on its gateway, sees growth from 11% to 36% in three months.
- Self-managed hosting in the cloud and on-premises (on the company's own hardware) split the market almost evenly: 46% versus 42%.
- According to BlueOptima, AI tools succeed in less than 23% of cases on real tasks at large companies. Sourcegraph explains the cause: the SWE-bench benchmark tests small, self-contained repositories, not large production codebases.
- Gartner warns that without governance, token costs will exceed the average developer salary by 2028.
- 64% of companies are deploying AI agents, but only 33% of them measure return on investment (ROI).
The author notes that estimates of AI prevalence vary across sources from 59% to 90%+, so individual percentages should be treated with caution.
Plans and review rules
In the middle of next year, team leads will each receive one NVIDIA DGX Spark, a compact AI computer, as a sandbox for experimenting with local generation. By the end of the year, they plan a shared server that can serve the entire team simultaneously on a 400B+ class model (400 billion parameters or more). The exact server specification is being deferred to a checkpoint after several months of working with the sandboxes. If their own server is not ready in time, the fallback is managed open-weight hosting.
The first version of the plan focused more on tools than on process. After reviewing McKinsey's findings, the team wrote three rules:
- Architectural review moves to the specification level, before code is generated.
- A separate checklist for AI-generated code is introduced: internal conventions, business-logic edge cases, and checks for hallucinations.
- Before AI-generated code is merged, a specific person signs off on the change, rather than it being approved "by default."
What to take away
The author draws conclusions for other teams. AI speeds up work but does not replace people, and it is better to say so openly from the start. Your own benchmark on your own codebase matters more than general leaderboards, especially for niche platforms like 1C. Process matters more than the tool. The team's well-being is part of the adoption plan. And even a simple quarterly self-assessment of results puts you in the minority of companies that measure ROI at all.
Caveat: this is one team's experience, and the industry data come from different sources using different methodologies. The material is useful as a starting point for an adoption plan, but not as proof of a universal effect.
- Nov 21, 2026Empatra Studio 0.29 Cuts Frontend Agent Costs by 59% and Runs Twice as Fast
- Nov 21, 2026Ego lite Gives Each AI Agent Its Own Workspace Space in the Browser
- Nov 21, 2026Instinct, valued at $10 billion, leads the market for AI agents you text like a messaging app
- Nov 21, 2026Docker rebranded for AI agents, but Sandboxes shipped back in September
Key Questions Answered
How it developed in the team?
There was no plan. In spring 2026, some colleagues simply tried AI agents on personal projects. By summer it had become a working practice: manual coding dropped sharply, but the load on code review and architectural decisions grew. By late summer the team began experimenting with local generation on its own hardware and assembled a shared knowledge base, from which a separate 1C knowledge base emerged. Knowledge-sharing meetings on AI tools are held once a week.
What the industry says?
Gartner predicts that most code will be written by AI agents and checked by experienced developers. By its estimate, 80% of engineers will need retraining by 2027. IDC describes the same shift as a move toward supervision: setting the task, checking the result, and refining the agent's output. A few more figures from the review: The share of local and open-weight models (with publicly available weights) in token traffic for corporate AI, according to a survey of 200 companies, grew from 23% to 34% over the year.
What are the key details on Plans and review rules?
In the middle of next year, team leads will each receive one NVIDIA DGX Spark, a compact AI computer, as a sandbox for experimenting with local generation. By the end of the year, they plan a shared server that can serve the entire team simultaneously on a 400B+ class model (400 billion parameters or more). The exact server specification is being deferred to a checkpoint after several months of working with the sandboxes. If their own server is not ready in time, the fallback is managed open-weight hosting.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.