Claude frees QA engineers to focus on judgement, not repetition

An automation QA engineer on L'Oréal's web program uses Claude and Claude Code to generate test scenarios, calculate coverage, analyze failures and refactor test code — all while keeping a human review gate at every handoff, from business logic through to the final commit.

Test development and maintenance can carry a heavy manual load: writing scenarios from requirements, calculating coverage, chasing gaps and triaging failures.

Within L'Oréal's web program, that load could be especially heavy because no issue tracker was connected to the workflow. Context that was needed to explain a failure, a known bug, a content change, had to be reconstructed by hand every time.

That overhead scaled with every release, regardless of whether the underlying work was straightforward or genuinely complex.

We introduced Claude and Claude Code. Now, the QA engineer runs test scenario generation, coverage calculation, coverage gap analysis and failure report analysis through those platforms, then uses Claude Code to refactor the test suite itself to keep token use down as it grows.

Rather than automate the workflow end to end, she’s built in a deliberate human gate at every transition:

• Business logic review of every generated scenario

• A flakiness check on automated tests

• Manual adaptation for cross-device cases

• Manual addition of context whenever a known bug or a content change explains a failure

• Manual sign-off for commit and push, always

She has also improved the system over time, adding generic, reusable component steps and sync date stamps to persistent memory so the workflow does not re-derive the same context on every run.

The result is closer to a well-governed assistant embedded in her existing process than to an autonomous pipeline, which is precisely why the human judgement in QA has not been designed out of it.

The effect is uneven by design, and that unevenness is the useful finding.

Gains are consistent on straightforward, well-structured coverage work, where generating and validating scenarios is now materially faster. Gains shrink on messy maintenance and failure analysis, where judgement and missing context still dominate the time spent.

That pattern is now shaping where the team leans on Claude first, and where the team still expects a person to lead.