Writing / Thinking Educational

The AI Code Review Bottleneck: Why Writing More Code Doesn’t Mean Shipping Faster

AI can increase the volume of code changes, but delivery still depends on review, testing, integration and operational confidence. Engineering leaders need to manage the whole flow, not just generation speed.

AI coding tools make it easier to produce a working-looking change. That is useful, but it does not automatically make software delivery faster.

Every change still has to be understood, reviewed, tested, integrated, deployed and supported. If the rate of code production rises faster than a team’s ability to validate changes, the bottleneck simply moves downstream.

The difference between generating and delivering

It is tempting to measure AI adoption through accepted suggestions, lines of code, pull requests created or time spent in an editor. These are activity measures. They tell us something about tool usage, but very little about whether customers receive value sooner.

GitHub’s recent discussion of agentic engineering highlights this tension. It points to research in which pull requests spent substantially longer in review as teams moved from lower to higher AI adoption. The exact result should not be treated as a universal law: team size, task mix, review policy and tooling all matter. But it is a useful warning against assuming that faster implementation means faster delivery.

Why review becomes harder

AI-generated changes can arrive quickly, but reviewers still need to understand intent and consequences. A patch may be syntactically correct while misunderstanding a domain rule, weakening an invariant, duplicating existing logic or missing an important edge case.

More changes can also mean more context switching. Reviewers must reconstruct why a change exists, whether the tests are meaningful and how it interacts with other work. If the pull request is large or poorly explained, the cost of understanding can exceed the time saved during implementation.

Five changes to the engineering workflow

1. Make work smaller before making it faster

Clear acceptance criteria and smaller changes reduce the cognitive load on both agents and reviewers. Ask for a plan before implementation when the task crosses multiple components. Split work when independent changes can be reviewed safely in isolation.

2. Make the agent show its evidence

A useful hand-off should include the problem understood, files changed, design choices, tests run, test results and known limitations. A summary that says “all tests passed” is not enough if it does not identify which tests ran and what they cover.

3. Separate self-review from independent review

An agent can inspect its own diff, look for obvious mistakes and run static checks. That is valuable. But self-review is not an independent quality gate. The workflow still needs automated checks and a human reviewer for changes whose risk warrants it.

4. Improve the test system

AI-generated tests can increase coverage, but coverage alone does not prove that the tests assert the right behaviour. Teams need representative fixtures, contract tests, integration tests and regression cases that reflect real failure modes.

5. Measure the whole delivery system

Track review waiting time, time actively spent reviewing, rework after review, escaped defects, change failure rate and end-to-end lead time. Segment the data by type of work. A result for routine tests or documentation should not be assumed to apply to complex architectural changes.

A practical 30-day experiment

Pick one team and one class of work. For the first week, record baseline measures: time from ready-for-development to merge, review waiting time, review rework and defects found after merge. During the next two weeks, introduce AI assistance with explicit requirements for smaller pull requests and evidence-rich hand-offs. In the final week, compare similar tasks and interview developers and reviewers about where time moved.

Do not declare success because more code was produced. Look for a reduction in end-to-end time without deterioration in reliability or an unacceptable increase in reviewer effort.

The leadership question

The objective of AI in software engineering is not to maximise code output. It is to improve the flow of valuable, reliable changes from idea to production.

If AI makes implementation twice as fast but doubles the work required to validate the result, the organisation has not become twice as productive. It has created a new queue.

The teams that benefit will treat review, testing and integration as first-class parts of the AI-enabled workflow, not as downstream clean-up.

Further reading

KEEP EXPLORING

More writing & thinking.

Explore the archive