ELMA365: three iterations to a useful AI for testing

⚡ The gist in 5 seconds - The gist: the KORUS Consulting team automated ELMA365 testing with an AI tool, going through three iterations. - Where it's available: Tatyana Veselova, head of the ELMA direction in the CRM&BPM department, described the case study on Habr. - Limitation: it was not possible to fully remove the human from the process — the final solution is semi-automatic. ### 🔍 What was found The team implements and customizes the ELMA365 low-code platform. There are no dedicated testers, so analysts manually check forms, roles, approvals, and regression scenarios before every release. The first version of the AI tool was "vibe-coded" with Claude in just a couple of weeks. It was supposed to study a config export, come up with test cases on its own, run them, and produce a report. In practice, the AI only checked the functionality of individual UI elements, such as form requiredness, and reported "everything works." Such "green tests" did not confirm that the user scenario passed end to end and that conditions triggered correctly. In the second iteration, the AI was connected to the system via API: the tool created records in the backend and filled in form fields. But this approach completely ignored UX: it did not check interface logic, field requiredness, dependencies on specific data, or complex sequences. Only the third iteration, with Playwright, produced a meaningful result: the tool logs into the browser under a specific user's account and repeats a human's actions in a real scenario. Even here, though, the AI often stopped midway: it did not understand what final status a process should reach, and it could not switch between accounts with different roles on its own. The result was a semi-automatic solution: the AI executes scenarios in the browser, while an analyst validates the expected outcomes. ### 💡 Why it matters Practical benefit — after three iterations, the tool reduced analysts' workload by roughly 30%. For teams without dedicated testers, that is a noticeable win: routine no longer completely eats up the time of people who handle requirements, documentation, architecture, and clients. The case also shows that AI does not replace a specialist in testing but shifts their role toward validating scenarios and expected results. This is a realistic automation scenario: first the tool, then error analysis, then embedding the human into the process. ### 🧩 Context ELMA365 implementation teams have project managers, developers, and analysts, but no dedicated testers. Analysts run regression testing before every release, even if only minor tweaks were made after a client demo. It was this routine that pushed Tatyana Veselova to experiment: clients asked why, with AI available, everything was still done by hand. The first two iterations — analyzing the config export and connecting via API — yielded no useful result. Only the switch to Playwright made it possible to get a working semi-automatic solution, where the AI emulates user actions and the human [text cut off]

🔗 Read on habr.com