You Feel Faster. The Data Disagrees.
A group of experienced software developers participated in a randomized controlled trial this year.
They used AI coding tools — the category with the strongest adoption, the clearest productivity case, the most uniform agreement that it works.
The result: they took 19 percent more time to complete their tasks with AI than without it.
Also: they reported feeling 20 percent faster.
Read those two numbers together. A 39-point gap between perceived and measured productivity. Among daily users. In the domain where AI is supposed to be working best.
That gap is not a paradox. It’s a measurement problem.
The number you measure shapes the reality you build
The developers weren’t wrong to feel faster. They were measuring the right thing for the wrong system.
AI compressed individual task speed. Any unit of work that could be broken into a prompt got faster. The feeling is accurate, because the execution is faster.
But execution isn’t the thing that makes software ship. Coordination is.
Code reviews that wait three days. Requirements that arrive unclear. Handoffs that lose context. Decisions that need four people and take two weeks. The interval between “done” and “closed.”
AI didn’t touch any of that. And developers who moved faster on execution collided with coordination overhead more frequently — because speed on isolated tasks surfaces the bottlenecks that were always there, just less often.
You’re faster. You hit the wall more often. Net: slower.
The macro version of the same problem
Fortune published a piece last week on the AI Solow Paradox.
In 1987, Robert Solow noted that computers were everywhere except in the productivity statistics. For nearly a decade, IT investment climbed while economic output growth stayed flat. The paradox resolved in the 1990s — not because computers got better, but because companies rebuilt the organizational layer above them. Networked systems. Standardized handoffs. Workflows redesigned around what computers could actually do.
Complementary innovation. The productivity showed up when the coordination layer caught up.
The AI version is structurally identical. Individual productivity gains are real and measurable. They’re not showing up in economic output. The gap between individual speed and organizational throughput is the coordination overhead living between nodes.
AI made every node faster. The pathways between nodes didn’t move. And those pathways are where time actually lives.
What you’re actually measuring
Most productivity measurements capture:
- Time to complete a discrete task
- Output per unit time
- Perceived efficiency — surveys, self-reported, the feeling of flow
None of them capture:
- Time between task completion and loop closure
- Commitments made but not tracked
- Decisions resolved versus decisions deferred
- Coordination overhead that doesn’t appear on any timesheet
The developer study measured time-on-task. The developers measured feelings of speed. Neither measured: did the work close, did the decision land, did the thread complete.
You can be 20 percent faster at every task and have nothing move. If the things that required coordination — and most things do, eventually — are still waiting for the handoff, the approval, the context that was supposed to travel with the deliverable but didn’t, the 20 percent gain is invisible at the output level.
That’s what the randomized trial actually found. Not that AI doesn’t work. That the measurement missed the coordination layer entirely.
Why this matters
If you’re measuring the wrong thing, you optimize for the wrong thing.
More AI tools. More output. Feeling faster. Less awareness of the coordination overhead compounding in the background. Until it surfaces: working longer hours despite identical workloads, more open loops than you started with, lower output quality on the things that actually require judgment, the persistent sense that something important isn’t getting done despite constant activity.
A UC Berkeley longitudinal study followed this pattern for eight months. 67 percent of workers who adopted AI tools in 2025 were working more hours by year-end. Not fewer.
That’s not a productivity failure. That’s what happens when the measurement doesn’t see the cost.
The system is working exactly as designed. The design doesn’t account for coordination.
The right number to measure
Not: how long did this task take?
But: how many open loops closed today, without going through me?
Not: how many outputs did AI produce?
But: how many commitments were made, tracked, and fulfilled?
Not: did I feel faster?
But: is the actual work moving — through the handoffs, the reviews, the decisions, the closures — or just arriving faster at the same bottlenecks?
The Solow Paradox resolved when organizations stopped measuring compute speed and started measuring system throughput. The new version resolves the same way.
The infrastructure that closes loops, tracks commitments, and makes coordination visible changes the number that actually matters. Until it exists, we’re measuring execution and calling it productivity.
The developers in the trial knew something was off. They just couldn’t name it.
That gap has a name. And it’s not in the execution layer.
