Alfa-Bank explains how a team of skeptics shipped an internal HR service to production with GLM-5

Alfa-Bank shared a case in which a team of vibe-coding skeptics took a new HR service to production in a few weeks using GLM-5. The experiment showed that AI noticeably speeds up prototyping and routine development, but it does not relieve the team of responsibility for architecture, Git and release.

How the project was chosen

The pilot inside HR Tech was staffed by three people who initially did not believe in the practical value of vibe-coding: a team lead, a systems analyst and a product manager. They were free to pick the task, and instead of an abstract demo project the team chose a real feature from the Alfa People roadmap. That became "My Goals", a service where bank employees can set goals and link tasks to them inside the corporate HR platform.

The bet was not only on speed but on reasonable risk. If the project failed, it would not break critical business processes, but if it succeeded, it would meet a real need. Deadlines tightened quickly: the due date was moved closer several times, leaving only a few days for development. The team reused discovery work already done, collected user pain points and described minimal solutions, then handed the model the task of drafting the business requirements.

How development went

In practice, the hardest part was not GLM-5 itself but getting people to work together who had never written applied software and barely used an IDE or Git. The team lead first had to run a quick onboarding, set up the environment and explain the basics of branches, pulls and pushes. Without that, AI did speed up individual pieces of work, but the whole flow stalled at every conflict, misunderstanding and mismatch of changes.

The process then looked like this:

• the team lead, systems analyst and product manager worked on the service
• the interface was built from text descriptions and series of rapid iterations
• the team made 19 prototype versions in total before choosing a basis for the product
• because of heavy load on GLM-5 during the day, development often shifted to night hours
• among 15 pilot teams, this one managed to bring its service to production

AI mostly helped with layout, forms, validators and first versions of screens. The product manager described what the interface should look like, the model generated HTML and UI logic, and the team lead connected it to the backend and brought it to a working state.

After release, the service quickly saw real use: in the first week it recorded 10 thousand unique users, 9240 created goals and 981 linked tasks, which looked like a very strong start for an internal product.

Where the limits appeared

The case shows clearly that the main bottleneck in enterprise development with AI is not code generation but engineering discipline. When several participants change a project in parallel, not knowing Git turns into constant manual conflict resolution.

The second problem was infrastructure: 15 teams took part in the pilot at once, and during the day GLM-5 could take several minutes to respond, making the tool almost useless and breaking the work rhythm. Another telling moment came right before the demo: the model and plugins became unavailable, and the final bugs had to be fixed by hand. That sobers people up fast and dispels the illusion that AI can fully replace an engineer at a critical moment.

Architecture, integrations, security, log review and edge-case handling all remained with the experienced developer. In effect, it was the last-minute manual work that let the team get through the presentation and release without a failure.

"In large companies, vibe-coding is not a replacement for a developer

but a tool in the hands of an experienced specialist."

As a result, the team arrived at a more sober model of use: AI suits parallel work on routine tasks, interfaces and draft hypotheses, but it does not remove the need for technical leadership. If the release boundaries are not fixed, there is a temptation to expand the scope endlessly, because it seems the model will "do just a little more". In practice, a firm task frame, manual control and the ability to stop further tweaks in time got the service across the finish line.

What this means

The Alfa-Bank story matters because it shows vibe-coding without the advertising filter. Inside a large company it can already speed up the launch of internal products, especially at the prototype stage and for standard UI tasks, but it works only where there is an experienced engineer, a clear scope of work and a willingness to handle everything critical by hand.

In this scenario AI is not an autopilot but a team accelerator, and it pays off only with a mature process and firm accountability for the result.