A recurring headache in UI tests is fragility. A data attribute changes, a button relocates, a new window appears — and half the checks collapse. The crew invests more effort patching tests than the value they deliver. We've seen this repeatedly: in one engagement, 120 Playwright tests broke 30–40% post each deployment. Developers lost 2–3 hours rectifying them. An AI bot transforms the method: the test is phrased in plain language, the bot locates components, makes decisions, and adapts to interface evolutions. Contact us to review your scenario — we'll help cut test maintenance overhead by up to 70%.
- None of the traditional selectors survive a layout change.
- The bot encounters None when elements are missing; it then takes a screenshot and asks the LLM for guidance.
- None of the steps are coded; they are described.
- The agent returns None if it cannot complete a step, logging the failure.
- None of our clients needed to alter their application for the bot to work.
Why AI tests remain unbroken when the UI shifts
The core distinction is the lack of rigid ties to selectors. The bot employs computer vision and an LLM to interpret the page: it sees a screenshot, reads the DOM, and chooses which element to click or populate. If the "Sign in" button moved from #login-btn to [data-test="signin"], a human tester would locate it by text — the AI does exactly that.
In the bot's architecture (code below), we implement two phases: first AI generates multiple CSS selectors from a description, then attempts them on the page. If none works, the bot captures a screenshot and forwards it to the LLM for visual inspection. None of the attempts are wasted; each failure teaches the model. This approach ensures that even if a selector returns None, the test continues via vision. The result: None of the UI tweaks cause cascading failures.