Start a conversation

AI-written code is shipping faster than anyone can review it. That is the real risk

The productivity gains are real. The danger is a review process designed for human speed, now facing output it was never built for

Every floor of a building carries the ones above it. Review is the floor most teams forgot to strengthen. Illustration: Thirdfloor Labs

In brief

  • Teams using AI coding tools are producing several times more code, but review capacity has barely moved.
  • The risk isn’t bad AI output. It’s good-looking output that nobody has truly read.
  • Review has to change shape: smaller changes, automated checks first, and people focused on what machines can’t judge.

Ask any engineering lead what AI has done to their team and the answer is usually the same: more. More code, more pull requests, more features started in a week than used to be started in a month. The gains are real, and we see them in our own work every day.

But speed has a second-order effect that gets far less attention. When the cost of writing code falls close to zero, the scarce resource becomes the ability to understand it. Most teams have not changed how they review code since long before AI arrived.

The bottleneck has moved

For most of software’s history, writing code was the slow part. Review was a check on something that took days to produce. Now a single engineer with an agent can open more changes in an afternoon than a reviewer can read carefully in a day.

Output has outpaced review

Change in a typical team after adopting AI coding tools, year on year

Code written3.4×
Pull requests opened2.8×
Reviewer hours1.1×
Test coverage1.0×

Illustrative · Source: [Thirdfloor client and internal projects]

The result is a quiet shift in what “reviewed” means. Changes get approved because they look right, the tests pass and the author is trusted. That was a reasonable shortcut when the author had written every line. It is a weaker one when they have skimmed it.

The danger isn’t that AI writes bad code. It’s that it writes plausible code, and plausible code gets waved through.Thirdfloor engineering

What review was built for

Code review was designed to catch the mistakes people make: typos, missed edge cases, a forgotten permission check. AI makes different mistakes. It invents an API that almost exists, copies a pattern that is subtly wrong for this codebase, or quietly widens what a function can access.

Those errors are hard to spot precisely because the surrounding code is tidy. A reviewer reading for style will pass them. A reviewer reading for intent, asking what this change allows that wasn’t allowed before, will not.

Guardrails that scale

None of this is an argument for slowing down. It is an argument for spending the time AI saves on the parts that matter. On every project we run, that means four things:

  1. Small changes, always. An agent will happily write a thousand lines. We ask for fifty, so a person can genuinely read them.
  2. Machines check first. Types, tests, linting and security scans run before a person looks, so review time goes on judgement, not syntax.
  3. Review for intent. Every change answers one question in writing: what can the system do now that it couldn’t before?
  4. A named owner. Whoever approves a change owns it in production. AI can draft; it never signs off.

What to ask your team

If you buy software, or run a team that builds it, three questions will tell you most of what you need to know. How big is a typical change? What runs before a person reviews it? And who, by name, is accountable for what goes live?

Faster delivery is one of the best things to happen to software in a generation. It should mean better work, not thinner checks. The teams that get this right will ship more, with fewer surprises, and that is the standard we hold ourselves to.

Thirdfloor Labs is a small team of engineers who build, run and support AI-first software, for ourselves and for the businesses we work with.

Talk to us about AI-assisted development