The Windstorm

A weekly dispatch from the eye of the storm

Week of August 10, 2026

Welcome to the Eye of the Windstorm.

The number everyone is passing around again this week reads like a eulogy for AI: MIT's Project NANDA put it at 95% of organizations deploying generative AI seeing zero measurable return. Zero. Not disappointing — nonexistent. Everyone else is reading that as proof the technology is oversold. Read it again from in here, in the calm. The models are not the thing that failed. The same week that number landed, three labs shipped systems faster and cheaper than anything we had a year ago. The tools work. The tools have never worked better. What 95% of companies could not do is get their people to wield them. That gap — between owning the intelligence and conducting it — is the whole story of this decade, and almost nobody is naming it out loud.

1. The zero-ROI number is a mirror, not a verdict

Sit with the specifics. Ninety-seven percent of executives say they deployed AI agents in the past year. Fifty-two percent of employees already use them. Yet only 29% of companies see significant return from generative AI, 23% from agents, and fewer than one in ten have scaled a single agent into real value. Notice what that pattern rules out. If the models were the bottleneck, the failures would cluster by capability. They don't. They cluster by organization. The technology walked in the front door of nearly every company on earth and then just… sat there. That is not a machine problem. That is a human one, and it is the most expensive one in business right now.

2. Everyone got an army; almost no one learned to command it

Here is where it gets vivid. Workforce interactions with AI agents grew 14x in the first half of this year. But 67% of those agents were built by people with no engineering background — folks in sales, ops, customer success — and 60% of them are over-permissioned, wired with access far beyond what their job requires. SAP now calls agent sprawl a board-level issue. Picture it honestly: you are a child monarch who woke up with a thousand advisors, each a thousand times more capable than you, and no one taught you how to hold court. So you either freeze, or you let them all run loose. Both are what failure looks like. Neither is the advisors' fault.

3. The labs just told on themselves

Watch where the money is actually moving. OpenAI stood up a Deployment Company — more than $4 billion — and is embedding Forward Deployed Engineers physically inside customers' buildings, even buying a consultancy to staff it. Anthropic launched a $1.5 billion services venture with Blackstone and Goldman. Read that plainly: the companies that make the models have concluded the model is no longer the product. The wiring is. When the people who built the electricity start selling the electricians, they are confessing the truth of the era — the hard part was never the current.

The Eye

We have seen this exact movie. Electricity was commercially available in the 1880s, but the factory productivity boom didn't arrive until the 1920s — roughly forty years — because buying the power was never the hard part. Rewiring the building was. Redrawing the floor so machines sat where the work flowed, not where the old steam shaft happened to run. Teaching every worker on the line to flip a switch they'd never trusted. The current was cheap and instant; the human rearrangement took a generation. Our version is the same story on fast-forward — forty years is becoming four — but the shape hasn't changed one bit. The intelligence is on the wall. The buildings are still wired for the old way.

So the winnowing skill of this decade isn't prompting, and it isn't picking the right model. It's judgment under abundance — knowing which of your thousand advisors to ask, what to ask them, and which of their confident answers to throw away. That is a conducting skill, not a coding one, and here is the part that should make you sit up: nobody is a virtuoso yet. Not the labs, not the consultants, not me. The 95% aren't behind because they were slow to buy. They're behind because they thought buying was the job. It never was. The job is learning to flip the switch — and this is the rare moment when everyone starts that lesson roughly even.

Do this with your agent this week

Stop asking your agent for answers and make it ask you questions. Take one decision you've been circling — a hire, a pricing change, a process you keep meaning to fix — and prompt: "Interview me about this decision. Ask one question at a time, and don't offer any advice until you've asked at least eight questions." Then actually answer them out loud. The Socratic method is the superpower nobody's using: the value isn't the machine's answer, it's the questions it drags out of you that you were too busy to ask yourself.

From the eye of the storm,
Grant