We're measuring AI productivity the wrong way
When we measure AI productivity, we tend to ask one question: how much time did this save me? That’s the wrong way to look at it. The one that actually matters is how much time did this save everyone. It sounds like a nuance, but it changes the whole picture.
The hidden cost shift
For centuries, writing had a built-in quality filter. It was expensive. Writing an email, a report, or a proposal took time, and every extra paragraph cost you effort. So authors edited themselves. The cost of communicating was paid primarily by the person doing the writing.
Language models changed that overnight. Today I can turn a handful of bullet points into three polished pages in seconds. The cost of producing words went to near-0. But the cost of reading them didn’t move at all. We took the burden off the author and quietly handed it to the audience. In fact it increased, because the more we produce, the more there is to consume.
The productivity illusion
Say I use AI to generate a design document. Before AI, I’d spend two hours writing it, and ten colleagues would spend ten minutes each reading it. Total human effort: three hours and forty minutes.
Now I spend ten minutes generating it. I don’t bother editing, because it already looks good. The document comes out four times longer, so those same ten colleagues now spend twenty-five minutes each getting through it. My productivity went up twelvefold. The organization’s productivity went down.
Measure only my time and AI looks like a miracle. Measure everyone’s time and it made us slower.
We’ve seen this before
The same thing is happening in software engineering. Submitting a pull request used to represent real effort. Someone had spent hours understanding, thinking, implementing, testing, and polishing. Now AI can produce a convincing implementation in minutes, so opening the PR is cheap. Reviewing it isn’t.
The maintainer still has to answer the questions AI can’t. Is this actually correct? Does it fit our architecture? What edge cases did it miss? Did the author understand the change, or just accept the output? The friction just moved onto the reviewer’s desk.
Or think about stand-ups
We’ve all seen the 500-word daily update that says almost nothing. It documents every command run, every file touched, every meeting attended, and still leaves the only questions anyone cares about unanswered. What changed? What’s blocked? Is there anything I need to do? The update optimized for activity, not for understanding.
The Bond principle
There’s a scene in Spectre where Bond orders a vodka martini at a health clinic bar. The bar doesn’t serve alcohol, so the barman hands him a prolytic digestive enzyme shake instead. Bond takes one look at it: “Do me a favor, will you? Throw that down the toilet. Cut out the middleman.”
Increasingly, that’s what we do with AI. Someone feeds bullet points into a model to expand them into polished prose. The recipient pastes that prose into another model and asks it to summarize the whole thing back into bullet points. We built an expensive pipeline whose input and output are nearly identical. All we added is noise.
The wrong metric
This is the same mistake I keep seeing when people evaluate AI coding agents. They obsess over token usage, but by itself it’s meaningless. An agent that burns ten times more tokens but finishes the work in one attempt can be far cheaper than a “token-efficient” one that needs constant retries and hand-holding. So the number that matters isn’t resource consumption. What we should be looking at instead is the total cost of getting the task done.
Communication works the same way. Writing- and reading time is our token count: the resource everyone consumes. The completed task is the decision or the shared understanding the communication exists to produce. So the number that matters isn’t how quickly I could write this, or even how long it takes to read. It’s the total human time it takes everyone to reach the decision. Time to decision is our task cost.
The real bottleneck
AI didn’t eliminate the cost of communication. It redistributed it. Generation is becoming free. Evaluation is becoming the expensive part: reading, reviewing, verifying, summarizing, choosing. Those are the scarce resources now, and that changes what our job actually is.
The best AI users are the ones who remove the most. Generating a wall of text is easy. The real skill, and the real courtesy, is leaving everyone else with as little to read as possible.