AI Shouldn't Do Your Code Review. It Should Tell You Where to Look First.

November 13, 2026

Watch the video version:

Can't AI just do the review? Your team is producing more code than it ever has, so it is a fair question. The short answer is no. The useful answer is that AI belongs one step earlier, deciding which changes need the most human attention, and saying why.

At a roundtable in New York last month, one technology leader described the problem from inside it. "My biggest struggle right now is the review gate." Compliance requires specific employees to do the final review before anything reaches production, and the offshore team's output is speeding up with AI. They asked how to hire more people, then asked the harder question: who wants "the job where all they do is review?"

Another leader at the same table, whose job is redesigning how a large company releases software, named what goes wrong next. Teams say they are changing their process with AI, and what they change is one step of it: the governance committee, the architecture review board. "We do that committee to manage risk," they said. "We do architectural review boards to make sure it's alignment with tech strategy." Their answer was to start from that purpose, ask how you would meet it with AI, and redesign the process end to end. Those are their accounts of their own companies.

Most teams skip that question. They put AI inside the step: it pre-screens every pull request, summarizes it, and sometimes approves the easy ones. The step stays, and every change still gets the same pass, only faster. The safe version keeps a human in the loop to sign off on what the AI saw, which leaves the same gate with a person at the end of it.

Here is what both versions miss. A reviewer's seat has a Design Aim: what it was built to hit. For a reviewer, the aim is catching the risk a change carries, and changes don't carry equal risk. A pull request that rewrites a permission check and one that fixes a typo in a log message get the same look. The seat keeps working. The risk lands somewhere else.

Not all pull requests are created equal, and some need greater attention. That is where AI earns its place in review.

My team is building toward an initial assessment on every pull request. The AI reads each change against the signals that raise its risk:

  • security, authentication, or permissions
  • a path that runs under load
  • shared code that other parts of the application depend on
  • the data model or a migration
  • a new dependency
  • the size of the change
  • an author new to that part of the code

It flags the risky ones to be looked at first and given the most attention, and it has to say why. A flag with no reason is a guess the reviewer can't check. The reviewer still reviews every pull request and is still responsible for every one. The assessment is a guideline, not a rule.

So before you decide where AI goes in your review, write one line for each review seat you have: the risk it exists to catch. If nobody can write the line, AI will only make that step faster at missing. Where the line exists, put AI where it tells the reviewer where that risk is, and leave the judgment where it sits.

Here is the test. Pull your last ten merged pull requests. Did the riskiest one get more review attention than the smallest? If it was the same, your review is aimed at nothing, and putting AI inside it will only aim at nothing faster.

The Structural CTO™ is the book about reading what a seat was built to hit, and building for the load a company has not reached yet: https://TheStructuralCTOBook.com

How to tell whether a leadership seat was ever aimed at anything, from a calendar and an inbox, without having worked a day at the company, once you are handed both. (Chapter 5)

Anthony S. Jackson

Anthony S. Jackson

Anthony S. Jackson has spent 30 years inside technical organizations. He is the author of the Architecture Protocol Series: three books on the structural problems technical leaders were never told they would face. He writes the LeadershipOS™ Inner Circle, a monthly printed newsletter for CTOs and engineering managers who design teams that hold under pressure.

LinkedIn logo icon
Back to Blog