What do you actually review in AI-written code, and what can slide?

November 18, 2026 5 answers
Asked by r.684~quindaro

Our team uses AI to write much of our web app code, including some CRUD flows that write back to an ERP. Should review focus on the code itself, the tests, or checking the UI and database? Does readability still matter if AI will help maintain it?

Answers

m.438~gasketspool

Start with anything that can change business data. Trace an ERP write from the UI through validation, authorization, database changes and the external writeback. Check what happens on retries or partial failures, including whether the same action can be applied twice.

Permissions and payments deserve the same close attention. Exercise the flow through the UI and check the resulting data, but also read the implementation. A successful click through one path won't tell you much about failure handling.

r.684~quindaro:

That makes sense. The writeback is a small part of the app by code volume, but a mistake there would have a much bigger impact.

Reply to this answer
_d75!orbitladle.6

I would read the whole change before asking anyone else to review it, just as I would for code written by a teammate. That doesn't mean every variable name needs a debate. A slightly awkward name is easier to live with than an abstraction nobody understands or a flow that hides where data gets changed.

Readability still matters. Someone has to diagnose the problem when the generated code does the wrong thing, and an AI tool can't take responsibility for the fix. If a change is too large to understand, break it into smaller pieces before it goes up for review.

Reply to this answer
Nexviro-4h

Tests are useful evidence, not a replacement for review. If the same model wrote the implementation and the tests, check whether the assertions reflect the actual business rules rather than just the behavior it happened to generate.

I'd look for independently specified cases around permissions, invalid inputs and failed writes, then run the app and inspect the stored results. You can also ask an AI tool for a second pass, but treat its comments as things to investigate, not approval.

_e91+drimble.7:

Checking the stored result is a good complement to the tests. A passing UI flow alone doesn't show whether the right data landed in the right place.

Reply to this answer
Veldixa+5q

Don't write off the frontend as just styling. Check that the user flows make sense, but also look at render times and how many API calls a screen makes. A page can look fine and still be doing a lot of unnecessary work.

Basic browser automation can help catch broken flows. Visual polish can be a lower priority than correctness, though the interface still deserves a deliberate look rather than whatever layout the generator happened to produce.

Reply to this answer
c.927+quorblet

There is a practical limit to line-by-line review when the output gets huge. For less critical work, running the app as a black box and checking the results can be more useful than trying to decipher every generated line.

That does leave implementation risks unseen, so I wouldn't use it as the only check for ERP writes or permissions. For a non-critical change, an imperfect name or some inelegant code may be worth accepting if the behavior is right.

Reply to this answer

Add your answer

Answers are reviewed before publication.

Related Discussions