AI Productivity Doesn't Mean What I Think It Means

I used to think AI productivity meant doing less. Automate the mechanical work, and the human effort shrinks toward zero. That was the promise, and for a long time I believed it.

The reality is stranger. AI does automate the mechanical work. But it does not shrink the human effort. It raises the ceiling of what the same amount of work produces.

Over the past three years, I have tested nearly every approach to writing with AI. I tried prompt templates, fine-tuned models, elaborate system instructions, & multi-agent debate pipelines. Most produced bland prose because they treated writing as one-shot generation.

The architecture that finally works is a closed-loop taste flywheel. Before drafting, the agent ingests background research alongside a local memory bank of learned house rules. It produces a draft in seconds, captures every human critique during review, & logs those lessons back into permanent files for the next essay.

The result changed my workflow, but not in the way most expect.

When an AI system improves, the common assumption is that human edits will fall toward zero. The data across ten published essays shows the opposite.

A week ago, drafting a data-heavy post required 47 draft revisions. Yesterday, an essay on agent lifespans took 3. Yet granular sentence edits remained flat at 130 per post. Across the entire dataset, line-level diffs show no statistically significant decline, hovering around an average of 140 edits per piece.

The time savings redirected attention. Editing did not vanish; it moved up the value chain.

Early in an agent deployment, almost all human effort goes into structural triage: throwing out academic literature reviews, relocating the thesis to the first paragraph, & fixing broken narrative spines. Once local taste memory stabilizes, structural churn collapses. The agent lands the thesis on the first or second attempt.

Freed from fixing broken arguments, human attention concentrates entirely on line-level craft. The ratio of granular edits per draft version surged from 4.4 to 43.3.

Chess shows the same pattern at a different scale. When engines arrived, the common assumption was that human chess would stagnate. The machine would do the thinking, and the human’s edge would erode. The data tells a different story.

AI did not flatten the distribution of chess strength. It reshaped it, and the ceiling rose. In 1979 exactly one player in the world was rated 2700 or higher.¹ Today the top thirty players average nearly 2,750, a cutoff that once belonged to a single man now marking the crowd behind the very best.² The number of grandmasters has gone from 524 in 1993 to roughly 1,750 today.³ The pool of elite talent did not shrink as the machine got stronger. It grew more than threefold. The tool that threatened to make human skill irrelevant instead made more of it possible, and free and universal.

Nor did players spend more hours to reach it. Viktor Korchnoi, among the hardest-working champions of the pre-computer era, trained as long as a tournament game ran, five hours a day, and noted the modern discipline takes four.⁴ Studying is faster. What once took two weeks of gathering and a month of preparation now takes half an hour in a database.⁴ The same or fewer hours produced a far higher ceiling of skill, not because players worked harder, but because the tool they trained with got better. The efficiency gain redirected the work into a higher ceiling.

This is the pattern I see in my own editing. The agent moved my edits from structural triage to rhetorical precision, from fixing broken spines to loading the loaded pistols.

In Words Like Loaded Pistols, Sam Leith observes that words are not passive containers for information. They are loaded mechanisms aimed at a reader’s mind.⁵ Rhetoric is the hidden machinery that makes an argument land.

When software removes the mechanical chore of building structural scaffolding, it frees the writer to focus on the aim: stripping adverbs, tuning sentence rhythm, sharpening antitheses, & cutting decorative clauses.

Economists have a name for the broader pattern. The Jevons Paradox holds that increasing the efficiency of a resource increases its consumption rather than lowering it.⁶ But the more interesting question is not about consumption. It is about the bar.

AI does not just save time. It raises the ceiling of what the same effort produces. The mechanical work collapses, and the discretionary effort moves up the value chain. The question is whether we accept the pre-AI baseline and stop there, or push for the higher-quality output that AI now makes possible.

The measure of a mature agent harness is not the elimination of human edits. It is elevating the writer from a structural mechanic to a sharpshooter.