AI-assisted test generation — real productivity boost or just hype?

Over the past few months I've been experimenting with Claude and GitHub Copilot for test case generation at work. Mixed results. **Where it genuinely helped:** - Generating happy path tests from API specs (OpenAPI → pytest skeletons) - Writing assertion boilerplate for repetitive CRUD tests - Suggesting edge cases I'd missed (null values, boundary conditions) **Where it fell short:** - Complex async flows with multiple service dependencies - Understanding business logic that isn't written down anywhere - Generating tests for legacy code without docs — hallucinations were a real problem Net result: ~20–30% faster on greenfield work, no meaningful speedup on legacy code. Anyone else tracking this systematically? I'm wondering if there's a better workflow I'm missing.