Most “AI for QA” pitches lead with test generation: point a model at your app, get a test suite back. That part works, sort of — it’s decent at scaffolding obvious cases and terrible at knowing which cases actually matter to your business.

The more useful shift is quieter:

• Failure triage. Flaky test detection and failure clustering across a CI run used to be a spreadsheet-and-vibes exercise. Models are good at reading a stack trace, a diff, and a flaky-history table together and telling you “this is the same root cause as last Tuesday.”
• Test maintenance. Locator rot and selector drift are exactly the kind of mechanical, pattern-matched fix an LLM handles well — given the old DOM, the new DOM, and the test file.
• Coverage gap analysis. Feeding a model your diff plus your existing test files to ask “what did we not cover” is a much better use of the technology than asking it to write the tests from scratch.

None of this replaces judgment about what’s worth testing. It replaces the tedious 60% of the job that made people burn out on writing tests in the first place.