By: TitApeX BoHan /October 20, 2026

AI Writes 30% More Code, Yet Delivery Isn't Faster: Ask About Code Review Before Commissioning Software

AI is writing code faster than ever, so many business owners reasonably assume that custom software should now be cheaper and quicker to deliver. A 2026 working paper from Harvard researchers offers a sobering answer: after companies adopted AI coding agents, developers wrote about 30% more code, yet the amount of finished, shipped work did not measurably increase. The bottleneck had moved to the AI code review stage. Here is what the research found, and how to raise it when you commission custom software.

30% More Code, No More Finished Work

Developers in an open-plan office writing code quickly on laptops

The paper, by Harvard PhD candidates Fiona Chen and James Stratton, draws on data from the engineering management platform Jellyfish, covering 718 companies and roughly 720,000 workers from 2021 through March 2026. The pattern is clear:

Output went up: lines of code added per engineer per month rose about 30%, commits about 20%, and pull requests (PRs) about 23%.
Completion did not: measured by work items deployed to production, finished work showed no statistically detectable increase.

In other words, AI made writing faster, but software is not done when it is written. Code has to be reviewed, tested, merged and deployed before it becomes a feature you can use. The paper is still an unreviewed working paper, but it points in the same direction as other industry data.

The AI Code Review Bottleneck: Merge Times Up 49%

A senior engineer facing a stack of printed code and screens full of review comments

The researchers found that the time from submitting a PR to merging it rose by about 49%, the share of PRs sent back for changes climbed sharply, and review comments per PR increased 35%. By March 2026, nearly 80% of the firms had used an AI review tool, yet AI wrote only about 23% of review comments. Most of the gatekeeping still fell on senior engineers.

A benchmark report released in early 2026 by DevOps platform Opsera, covering more than 250,000 developers, saw something similar: AI cut time-to-PR by up to 58%, but AI-generated PRs waited 4.6 times longer in review and introduced more security vulnerabilities. Earlier, METR's 2025 randomized trial even found that experienced developers took 19% longer to finish tasks with AI, while believing they had been faster.

None of this means AI is useless. It means that writing code is no longer the expensive part; judging whether the code is right is.

Four Code Review Questions to Ask Before You Commission Custom Software

A business owner asking a development team lead about the review process

You do not need to read code to ask the right questions. These four help you tell whether a development team uses AI to accelerate, or simply to pile up volume:

1. Who reviews the code AI writes?

Ask the team to walk you through their process. Does a senior engineer look at every change before it is merged? Is an AI review tool an assistant, or the only checkpoint?

2. Where are the review standards written down?

Good teams keep an explicit review checklist covering security, performance, naming conventions and whether each change has matching tests. Our earlier piece, AI writes the code, but who tests it?, pairs well with this question.

3. How big is each delivery?

A single update with thousands of lines of code is the hardest to review and the easiest place for problems to hide. Small, frequent deliveries usually signal more rigorous review.

4. Can problems be traced back to a specific change?

Ask for version control history and change notes, so anyone can later understand why a piece of code exists.

Reading Quotes and Timelines: Where Does the Saved Time Go?

If the bottleneck is review, a fair quote will not drop dramatically just because a team uses AI. The research suggests AI mainly saves writing time, not total delivery time, which hardly supports the hope that "AI makes it half the price." If a quote is astonishingly low or a timeline unreasonably short, ask what was cut: the writing time, or the review and testing time?

We suggest thinking of the budget in three parts: write (AI can speed this up a lot), review (requires experienced people) and verify (testing and pre-launch checks). In a healthy project, the last two are never squeezed to almost nothing.

Conclusion: AI Brings Speed, People Bring Judgment

A developer and a client shaking hands in front of the finished system

Software development has entered a stage where producing code is easy and producing the right code is rare. That is TitApeX's view in a sentence: AI handles flawless efficiency, people handle the judgment that matters. In our custom development projects, AI speeds up writing and first-pass checks, but every line that goes live is reviewed, tested and explained by a person. If you are evaluating a custom system, take a look at our services or talk to us about what you need.