What AI-powered QA actually looks like

"AI-powered QA" usually means a model that writes unit tests from a function signature. That is a code generator, not QA.

QA is a person opening the app, doing something slightly wrong, and noticing what breaks. The interesting question is whether an agent can do that. It can, with the right setup.

The agent needs hands

A QA agent needs to operate the product the way a user does. In practice that means:

• a real browser it can drive, with screenshots and the accessibility tree;
• a real mobile simulator or device farm for iOS and Android;
• access to logs, network traffic and the database state;
• a place to write findings that a developer will actually read.

Without hands, the agent can only reason about code. With hands, it can reason about behavior.

Exploratory, not scripted

Scripted tests check what you already thought of. Exploratory testing finds what you did not.

The agent gets a charter, not a script:

charter: checkout flow, logged-out user, slow network
time box: 20 minutes
report: anything that blocks purchase or looks wrong

It navigates, tries edge cases, takes notes and files findings with reproduction steps, screenshots and log excerpts. Each finding is verified a second time before it is reported, so developers do not get noise.

Regression that writes itself

Every confirmed finding becomes a candidate regression test. The agent writes it, runs it against the fix, and adds it to the suite. Over time the scripted layer grows out of real bugs rather than guesses.

What it does not replace

It does not replace product judgment. An agent will tell you that a button does nothing on Android 14. It will not tell you that the feature is a bad idea.

It does replace the hours of clicking through the same flows before every release. That is the part nobody wanted to do anyway.