Theory of constraints, AI, and code review
Swizec Teller
AI writes code faster now, but companies aren't shipping faster. The real bottleneck moved to code review, not code writing.
Based on reporting by Swizec Teller — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI coding tools have made writing code absurdly fast. Studies show individual gains everywhere: coders finishing tasks 26% faster, support tickets closed 15% quicker, whole categories of busywork shrinking. Yet zoom out to the company level and the numbers barely move. Research across thousands of firms and executives shows productivity bumps in the low single digits, sometimes under 1%. Something is eating all those individual gains before they reach the business.
The explanation comes from an old manufacturing idea: theory of constraints, coined by Eliyahu Goldratt back in the 1980s. A factory only moves as fast as its slowest station. Speed up everything else and you just pile up inventory in front of the bottleneck. Software teams are discovering the same thing, except the inventory is pull requests. Engineers now produce more code than ever, but that code has to pass through humans reading it, testing it, and approving it — and that step hasn't gotten any faster just because an AI wrote the diff.
For years, teams chased faster releases by shrinking deploy cycles and adding linters and automated tests to catch obvious mistakes before a human ever looks at the code. That worked because it thinned out what reached review, letting reviewers focus on real judgment calls instead of style nits. AI breaks that balance by flooding the front of the pipeline with volume no linter or test suite was built to filter. The result: PRs stack up, conversations sprawl across days in comment threads, and a team that feels frantically busy still ships maybe one or two real changes a day.
The fix isn't skipping review, even though plenty of AI boosters suggest just trusting the model. Comparing generated code to compiled code doesn't quite hold — code generation is still an artisanal, prompt-shaped process, not a settled abstraction layer like a compiler. Skip review now and the scrutiny just relocates to prompts and design docs, which senior engineers already spend much of their time on anyway. The actual leverage move, straight out of Goldratt's playbook, is to subordinate everything else to the bottleneck: slow down code generation, put more senior time into review, and treat mentoring through review as one of the highest-impact things an experienced engineer can do — not a rubber stamp to clear on the way to the next ticket.
My take — AI-written commentary, not fact-checked reporting
This tracks with what every AI-productivity headline conveniently ignores: individual speedups don't automatically compound into organizational speedups, and pretending otherwise is how you end up drowning in generated code nobody has time to actually check. The honest fix is boring — spend more human attention on review, not less — which is exactly the kind of unsexy advice that never trends on AI Twitter.
Read more about this at: Swizec Teller