AI Coding Tools Boost Output but Not Releases, MIT Research Finds

The Research in Plain Terms

Researchers from MIT Sloan and the University of Pennsylvania tracked how developers used three categories of AI coding tools across six stages of the software development lifecycle:

• Autocomplete tools — suggest code as you type
• Sync agents — write and edit code in real time alongside the developer
• Async agents — work autonomously on assigned tasks without developer oversight

The productivity gains were real and measurable at the code-writing stage. Autocomplete tools alone boosted coding activity by 40%. Add sync agents and that number climbs to 140%. Stack async agents on top and you hit 180%.

But here’s where it gets interesting.

Productivity Gains Fade as Code Moves Downstream

The further code moves through the development lifecycle — from writing to reviewing, merging, and releasing — the more those AI-driven gains erode.

The cumulative effect of AI tools translated to just 50% more projects and 30% more actual software releases. That’s still meaningful, but it’s a steep drop from 180%.

The culprit isn’t the AI. It’s the human processes that follow code generation. Code review, pull request approvals, testing, and release management are still largely human-led workflows. AI tools, at least in their current form, aren’t built to handle those stages — and those stages are where the bottlenecks live.

As MIT Sloan associate professor Mert Demirer put it: “Some bottlenecks are truly technical and mechanical tasks that humans have to step in and complete because AI isn’t good at doing them.”

More Apps, Fewer Downloads

The researchers also looked at data from four major app stores and found a broad increase in new app releases starting in early 2025 — roughly when agentic AI coding tools became widely available.

But there was no corresponding increase in app downloads or user reviews.

That’s a telling signal. AI is lowering the cost of writing code, which means more software is getting built. But more software doesn’t mean better software, or software that finds an audience. Testing, iteration, and market fit still require human judgment — and those processes haven’t changed just because code generation got faster.

Where the Real Bottlenecks Are

The research points to a structural problem: most software development workflows were designed before AI existed. They weren’t built to absorb a 180% increase in code output at the front end while keeping everything else the same.

Two practical responses emerge from the findings.

Make teams smaller and more focused. AI tools boost individual output, which means you don’t necessarily need the same headcount to produce the same volume of code. Smaller teams move faster and have fewer coordination costs. The freed-up resources can then be redirected toward the downstream work that AI can’t yet handle — merging, releasing, and maintaining products.

Map where time is actually being lost. Not every bottleneck is a coding problem. Some are process problems: slow approvals, redundant meetings, unclear handoffs. AI tools can help here too — summarizing discussions, drafting documentation, accelerating prototyping — but only if leaders actively look for those opportunities rather than assuming productivity gains at the code level will ripple outward automatically.

Demirer’s framing is worth keeping: “Look at where people are spending too much of their time and see how AI can resolve that.”

What This Means for AI Tool Adoption

If you’re evaluating AI coding tools — or already using them — the research offers a useful reality check.

Measuring success by lines of code written or commits made will give you an inflated picture of actual impact. The more meaningful question is whether AI tools are helping your team ship faster, not just code faster.

That means auditing your full development lifecycle, not just the writing stage. Where do pull requests stall? Where does testing slow things down? Where do releases get delayed? Those are the stages where AI investment — or workflow redesign — will generate the most leverage.

AI coding tools are genuinely useful. The productivity gains at the code-writing stage are well-documented and significant. But treating code output as a proxy for engineering productivity is where teams get misled.

The bottleneck isn’t the code. It’s everything that happens after.