What to remember from this issue
Building things got cheap. Saying exactly what you want did not. Most companies still budget as if it were the other way round.
The field note
You probably believe, without ever saying it out loud, that making things is the hard part and saying what you want is the easy part.
For your whole career that was true. Building software took a team and a quarter. Describing what it should do took a meeting. So companies organised themselves around the scarce part: engineering got the budget, the roadmap and the six-month plans. The description got one slide, written in an afternoon by whoever had the meeting room.
That has flipped, and most companies have not noticed. Producing software, content and analysis now costs close to nothing. What did not get cheaper, because no AI can do it for you, is knowing exactly what should be built, why it matters, and how you will tell whether it worked. Making got cheap. Deciding became the bottleneck. The meeting room and the engineering floor have traded places, and the org chart has not heard.
I call the old belief the Production Illusion, and you can watch it fail in public every week now. The pattern is always the same: a capable system does exactly what it was told, and what it was told was wrong. The danger was never a machine that disobeys. It is a machine that obeys a bad instruction perfectly.
This newsletter is written from inside the book I am writing about that flip, The Specification Economy. One field note a week, as the chapters take shape. You are reading the idea of the prologue before the prologue exists. Reply and argue with it. That is what field notes are for.
Try this week
Take your last AI budget. Split it in two: what went to tools and licences, and what went to writing down what the tools should do. For most companies the second number is zero. That zero is the hard part, unpaid.
Seen this week
A customer asked an AI support agent to check a price typo. The agent confirmed the typo instead, politely and in detail.
Who Gives A Crap, the bamboo toilet-paper company, emailed a subscriber that their plan was changing from $66 for 48 rolls to $69.50 for 24. It was a typo: $69.50 covered 48 rolls. The customer, reasonably, asked support to double-check. The AI agent confirmed the mistake in confident prose: “the quantity of rolls in your subscription will be halved, and the price will increase.” The company switched the agent off and sent a correction. The agent did not break. It did exactly what it was given, and what it was given was wrong. (SmartCompany, 8 July 2026)
Someone said it better
“Handing out Claude credits to your employees in buckets is the same thing as giving a 480-volt sparking live line to a factory worker and saying: ‘Hey, it’s electricity. We need electricity to make stuff. Figure it out, make yourself more productive.’ How’s that going to go? That’s not going to go super well.”
Nate B. Jones, on the AI Automation Society podcast, at 35:18. Tools without instructions are not a transformation. They are voltage.
One number
25%. Only one AI project in four delivered the return it promised, in IBM’s survey of 2,000 CEOs across 33 countries. Not for lack of tools: 61% of the same CEOs already run AI agents. The tools were never the problem. Knowing exactly what to build, and how you would tell whether it worked, was. (IBM, 2025 CEO Study)