An honest look at what large language models are good at in a QA workflow, and where they still cost more than they save.
Given a user story and acceptance criteria, a model produces a solid list of positive, negative and boundary scenarios in seconds. Treat the output as a checklist to review, not a test plan to ship — the value is in coverage prompts you might have missed.
Pasting a stack trace or an unfamiliar framework class and asking for an explanation is a genuine time saver, especially when onboarding onto an old Selenium codebase.
A model cannot see your DOM at runtime. Locators it invents look plausible and fail in CI. Always regenerate them against the real page with codegen or the inspector.
Self-healing tools patch over unstable locators. That is useful during a migration and dangerous as a permanent strategy — it hides the drift instead of fixing it.
Use AI to accelerate the writing and the reading, keep human review on everything that becomes an assertion. Quality decisions stay with the engineer.
Book a free demo class and work through these patterns with a mentor.
