Is AI Making Developers More Productive Or Just Busier?
By Muzammil Rawjani, Founder of Software Orca, delivering mobile, custom, and AI-powered software solutions.
Ask a room of engineering leaders whether their teams got faster this year, and most will say yes without hesitation. But then ask them to point to a release date pulled forward, a defect count driven down or a customer complaint resolved, and the room will likely go quiet.
The gap between confidence and evidence is the real question, and I think it deserves closer scrutiny than most adoption headlines give it.
Speed And Output Are Not The Same Thing
The instinct to equate AI use with productivity is understandable. Watching code appear on screen in seconds feels like progress, and at the level of a single keystroke, it is. But software delivery was never bottlenecked by typing speed.
It was bottlenecked by understanding the problem correctly, integrating new logic without breaking what already worked, and catching mistakes before they reached a customer.
Faster code generation does not remove any of those three constraints. Instead, it adds pressure to them. A function written in seconds still needs to be understood, tested and fitted into an existing system, and a human still has to do all three, at the same pace as before.
What a Controlled Study Actually Found
Most productivity claims about AI coding tools come from vendor surveys or self-reported estimates, both prone to optimism bias. I believe METR ran a rare exception: a randomized controlled trial in which experienced open-source developers completed real tasks either with or without AI assistance.
The AI-assisted group finished 19% slower, despite predicting beforehand a 24% improvement.
The gap between prediction and outcome matters more than the slowdown itself. It suggests developers cannot reliably judge their own AI-assisted speed from inside the work, which means self-reported productivity metrics, the kind most companies still rely on, measure a feeling rather than a fact.
The Debt That Accumulates Quietly
Task-level slowdown is one problem. A second, quieter one is what AI-assisted code does to a system’s long-term health. Git Clear’s analysis found 211 million lines of code refactoring, the deliberate work of simplifying and consolidating code, fell from 25% of all changes to under 10% between 2021 and 2024, the same stretch AI coding tools went mainstream.
Fewer developers are pausing to simplify what they build; more are generating new code to solve today’s problem and moving straight to the next one.
This shift does not appear in a sprint retrospective. It surfaces eighteen months later, when a routine feature request takes three times longer than it should, because nobody can safely touch the surrounding code anymore.
Software built that way accumulates weight rather than value, and its usable lifespan shrinks instead of extending.
Where The Real Gains Concentrate
None of this argues against AI adoption. Gartner’s research found 90% of engineering leaders report genuine productivity improvement, with a measured average gain of 19.3%. This figure and the METR result are not really in conflict.
They describe two different populations: organizations where review capacity and testing discipline scaled alongside code generation, and organizations where generation raced ahead of both.
I find that this distinction is entirely about governance, not tooling. Recent AI in software development statistics show that 60% of enterprises continue to use untested code. Put plainly, many organizations have already tipped into the second group without deciding to.
Making The Distinction In Your Own Organization
If you want to know whether AI is actually improving engineering performance, try these things:
• Compare defect and incident rates from before and after AI adoption. Rather than relying on developer sentiment, use delivery metrics. As already evidenced, perceived speed and actual speed can diverge sharply.
• Track the ratio of new code to refactored code over time. A shrinking refactor share is an early warning sign of maintainability decline.
• Require a defined review and testing gate for AI-generated code before it merges. Untested code reaching production is now a documented pattern rather than an edge case.
• Give engineering leadership explicit ownership of the gap between generation speed and review capacity. One should not silently outpace the other over successive release cycles.
• Revisit the productivity question quarterly using delivery data instead of adoption data. Tool usage alone says nothing about whether the business is shipping durable software faster.
AI does not make developers more productive or busier on its own. It magnifies the discipline already present in an organization’s engineering process, for better or worse. The only way to know which direction it is pulling your team, and your codebase’s long-term health, is to measure outcomes instead of activity.
Forbes Business Council is the foremost growth and networking organization for business owners and leaders. Do I qualify?
