The Windstorm

A weekly dispatch from the eye of the storm

Week of August 17, 2026

Welcome to the Eye of the Windstorm.

Ninety-five out of every hundred enterprise AI pilots this year returned nothing measurable — and the models were almost never the reason. That is MIT's number, out of their Project NANDA, and I wrote about it last week. I am staying on it, because the reaction since has been worse than the finding: leaders back from summer are reading it as a verdict on the technology. When the researchers went looking for the culprit, they didn't find dumb machines. They found sharp tools bolted onto workflows nobody redesigned, data nobody cleaned, and people nobody taught. Roughly 80% of the work of turning a pilot into a payoff is plumbing and habit — integration, governance, the unglamorous wiring. The intelligence is on tap now. Conducting it is the whole job.

1. The 5% are pulling away, fast

The same week, Salesforce dropped its 2026 Agentic Enterprise Index, and the contrast is the story. The average business on their platform went from 5 active agents in early 2025 to 13 by this April. Time to deploy an agent fell from four days to under two. Those agents now carry six skills each, up from two. In April alone the platform logged 734 million units of agentic work. Retailers who leaned in reported online sales four times higher than those who didn't. The winners here aren't buying smarter software than the 95%. They're wielding it better. Same tools, different conductors.

2. Ten new models in a month, and no virtuoso yet

August brought Google's Gemini 3.7 Flash on the 13th, Alibaba's Qwen3.8 Max on the 2nd and a 27B sibling on the 14th — roughly ten model releases from six labs in thirty days. The reflex is to feel behind. Resist it. You are a child monarch at a table of advisors, each of whom knows a thousand times what you do. Your power was never to become the smartest advisor. It's to ask the sharpest question and to choose whose answer to trust. That winnowing — knowing which tool for which job, and which output to throw out — is the defining skill of the decade. Nobody's a virtuoso yet. That's the opening.

3. The bill for autonomy came due

Quietly, on August 2nd, the high-risk provisions of the EU AI Act became enforceable, with penalties reaching 15 million euros or 3% of global revenue. Meanwhile Gartner expects 40% of enterprise apps to embed task-specific agents by year's end, up from under 5% a year ago. Put those together and the message is plain: we're handing real decisions to agents faster than we're deciding who answers when one gets it wrong. Conducting intelligence isn't only about getting good work out of the orchestra. It's owning the sound it makes. Accountability doesn't delegate.

The Eye

Here is what's actually under this week's noise: the 95% failing and the few doubling their agents are not two stories. They're one. We keep treating AI like it's the electricity — the miracle you plug in. But the miracle was never the current. When factories electrified, forty years of promise became four years of payoff only once someone rewired the building, moved the machines, and taught every worker on the floor where the switch was. The electricity was the easy part. It always is.

The tools this week are genuinely astonishing and almost beside the point. The gap between the winners and the stalled isn't a smarter model — it's a leader willing to do the deeply human, deeply boring work of rewiring how their people think and work around the current. That work doesn't feel like innovation. It feels like teaching, and asking, and cleaning up. But it's the only work that has ever turned a miracle into money. The calm center of this storm isn't a better prompt. It's remembering that the hard part was always us.

Do this with your agent this week

Pick one task you do every week that you secretly dread. Don't hand it to your agent yet. First, write down the exact question you'd ask a brilliant new hire on their first day to get it done — every bit of context, what "good" looks like, what to avoid. Now hand that to your agent. When the answer disappoints, don't blame the model — sharpen the question and run it again. Do that three times. You're not automating a task. You're practicing the one skill that separates the 5% from everyone else.

From the eye of the storm,
Grant