AI code review tools, run automatically on every pull request, reliably catch a specific and genuinely useful category of issue: obvious bugs, inconsistent patterns against the rest of the codebase, missing error handling, and style inconsistencies — largely the same territory static analysis and linting already cover, but explained in plain language rather than a terse rule violation, which makes the feedback more actionable for whoever's reading it.
What it structurally cannot catch is anything requiring context outside the diff itself. Whether a change actually solves the right problem, whether it interacts badly with a piece of business logic three files away that isn't part of the current change, whether the approach chosen is appropriate given constraints the AI has no visibility into — a deadline, a planned future change, an organisational decision made in a meeting last month. These require the kind of judgement covered in the piece on code review that actually works, and an AI reviewer has no access to that context regardless of how well it reads the code itself.
The risk worth watching for is human reviewers treating an AI-approved pull request as sufficiently reviewed, and skimming rather than genuinely reading it themselves. AI review should reduce the mechanical burden on a human reviewer — catching the small things automatically so a human's attention goes to the judgement calls — not replace human judgement about whether the change is actually a good idea. Teams that let AI review substitute for human review, rather than filter what human review needs to focus on, lose exactly the oversight that catches the most consequential mistakes.
Where it earns a clear place is on high-volume, lower-stakes changes — dependency bumps, small refactors, routine bug fixes — where the mechanical checks it performs well are genuinely most of what review needs to accomplish. Reserve dedicated human attention for changes to core business logic, security-sensitive code, and anything architecturally significant, where the judgement AI review cannot provide is exactly what the change most needs.
