A recent Fortune article stopped me cold: experienced software developers using AI coding assistants took 19–20% longer to complete tasks than when working without AI. Even more striking, those same developers predicted they would save 24% of their time. That’s a 44-point swing from expectation to reality.

The AI productivity revolution everyone’s been hyping? It’s running late. But before we dismiss AI-assisted coding as overhyped, we should understand what’s actually happening. Because today’s slowdown may be setting up tomorrow’s breakthrough.

The Context Gap

The METR study followed 16 experienced developers across 246 tasks. The slowdown wasn’t because AI couldn’t write code. It can. The issue was fit. Developers spent significant time cleaning up AI-generated output so it aligned with their architecture, standards, business rules and constraints.

AI knows everything generally. Your project needs everything specifically. Software development has always been about understanding what to build and why. The how is implementation. AI accelerates the how, but it doesn’t inherently solve the why or the constraints. That context gap is the bottleneck.

The study captured developers on day one with these tools. Not month six. Not year two. Day one.

Anyone who has lived through shifts like waterfall to agile or manual deployments to CI/CD knows the pattern: initial slowdown, eventual fluency. Working with AI requires a new muscle. Developers trained in precise syntax are now being asked to communicate intent in natural language and trust probabilistic output.

That fluency takes time. Teams gradually learn where AI excels, where it fails and how to prompt strategically. The real question isn’t whether the 20% slowdown exists. It does. The question is whether it’s permanent.

The Human Equation

There’s also a psychological dimension. Senior developers aren’t simply afraid. They’re evaluating tradeoffs. If AI makes them slower and lowers quality without clear upside, adoption feels like a handicap. Junior developers face a different concern: if AI handles foundational tasks, are they building durable skills or dependency? These concerns deserve engagement, not dismissal. The study shows the costs are real. The equation only changes if capability improves and workflows adapt.

Today’s AI coding tools lack persistent memory and deep architectural awareness. They don’t truly understand your codebase, your issue tracker or your design intent. They generate plausible output, but without durable context.

That will change.

As tools begin to maintain project memory, integrate across systems and surface uncertainty instead of hallucinating confidence, the context gap narrows. We’ve seen waves of development automation before that overpromised. The difference now is semantic understanding and tighter integration directly inside the developer workflow.

Here’s the uncomfortable idea: some of that 20% friction may be useful.

Reviewing AI output forces architectural clarity. Reconciling suggestions with constraints makes assumptions explicit. That friction may prevent long-term technical debt that traditional velocity metrics fail to capture.

The study measures task completion time. It does not measure system quality, maintainability or defect reduction. If AI makes you slower per task but reduces long-term complexity, the headline number misses the bigger shift.

A Realistic Path Forward

The broader data reinforces caution. MIT research shows only a small percentage of AI deployments deliver rapid revenue acceleration. Economist Daron Acemoglu estimates only a modest share of tasks benefit meaningfully from automation. These aren’t anti-AI arguments. They’re calibration points. AI is augmentation, not autopilot.

For development teams, that means:

• Temper near-term ROI expectations
• Invest in explicit context and documentation
• Deploy AI strategically, not universally
• Measure quality and maintainability, not just velocity

The 20% slowdown is real. It’s documented. It matters. But it’s a snapshot, not a destiny.

We are in the awkward middle phase of adoption: tools still maturing, workflows still adapting, expectations still inflated. Over the next 12–24 months, the durable AI coding platforms will separate from the noise.

The revolution may still come. It just won’t arrive evenly, instantly or without friction.

And that’s exactly why this conversation matters.

—

What’s your experience with AI-assisted coding? Are you seeing productivity gains, slowdowns, or something more nuanced? I’m genuinely curious how different teams and organizations are navigating this transition.

—

About this article: This piece was inspired by Sasha Rogelberg’s Fortune article “Experienced software developers assumed AI would save them a chunk of time. But in one experiment, their tasks took 20% longer” (January 5, 2026), which reports on research from Model Evaluation and Threat Research (METR).