The Windstorm

A weekly dispatch from the eye of the storm

Week of November 2, 2026

Welcome to the Eye of the Windstorm.

This week a 164-year-old Swiss bank told you more about where AI is going than any model launch did. In its third-quarter earnings disclosure, UBS reported that its employees now run more than 12,700 personal AI agents in their daily work — with 608 use cases live in production, up 81% in a year, and another 1,090 still in development. Sit with that for a second. The ability to build these agents has been sitting on the shelf, available to every bank on earth, for a good while. UBS is simply the one that did the unglamorous work of wiring the building and teaching thousands of people to flip the switch. The model was never the hard part. The hard part was the people.

1. A bank’s real moat turned out to be a filing system

UBS doesn’t just hand people a chatbot. It says it is also progressing nine large-scale, end-to-end AI initiatives alongside those personal agents. Notice what that is. It isn’t a smarter model; it’s an operating plan. And the number that should stop you isn’t the 608 live use cases — it’s the more than 1,090 still in development. Everybody can generate a thousand ideas for what an agent might do now. The actual skill, the one UBS is quietly getting good at, is the winnowing: deciding which of the 1,090 is worth the trust, the oversight, and a human’s name on the sign-off. That judgment is the whole ballgame, and no vendor sells it.

2. The cost of letting AI drive just got cut in half

Last Monday OpenAI released GPT-6.1 Terra, built not to dazzle on a benchmark but to be fast and cheap at the dull work of operating software — reading a technical drawing, navigating a desktop app, filling in a form. At $1 and $5 per million tokens, it costs roughly half what OpenAI’s earlier models do, and it’s already landing in Codex, ChatGPT Work, and Microsoft’s Foundry. Read the real headline underneath: the thing getting cheap isn’t intelligence, it’s the labor of clicking. Which is wonderful — and completely wasted if you haven’t first decided what, precisely, is worth clicking on. Cheaper hands don’t help a company that hasn’t figured out what it wants built.

3. The agents are getting a manager

SAP spent the week at TechEd tying Joule Work and Joule Studio together — tools that let a company orchestrate agents across its business apps. IBM, the same week, shipped identity and access controls so a company can actually see which agents are acting inside its systems. Three weeks ago Microsoft CEO Satya Nadella argued that frontier models should be treated as insider risks; this week the industry started shipping the filing cabinet to keep track of them. The pattern is unmistakable. The question has stopped being “can the machine do the task” and become “can we manage a workforce of things that can.”

The Eye

Strip away the logos and this was one story told three times. UBS running 12,700 personal agents, Terra halving the price of clicking, SAP handing agents a manager — each one moves the bottleneck to the same place. For a decade the governing question of this whole era was can the machine do it? As of this week, in production, at a real bank, at half last month’s price, the answer is plainly yes. And the moment that question is answered, it is replaced by a far older and more uncomfortable one, the only one left: what, exactly, should it do? That is not a question a model can answer for you. It’s a question you have to ask well — and asking well is a human skill, almost a Socratic one, that most of us have let go soft.

Here is the calm at the center of it. You do not win this by matching the industry’s noise — the launches, the price cuts, the breathless threads. You win it by getting quiet enough to ask a sharp question and clear-eyed enough to recognize a good answer when it comes back. UBS’s advantage isn’t its models; it’s the thousands of people who have to learn to tell a good answer from a plausible one. That’s the work now. The storm is all capability. The eye is judgment.

Do this with your agent this week

Pick one task you do every week that naturally splits into a draft and a review — a client email, a budget note, a first-pass plan. This week, run it bank-style: let the agent be the maker and make yourself the checker. Have it produce the draft, then, before you accept a single word, interrogate it with three hard questions — “what did you assume here, what would make this wrong, what did you leave out?” Notice which muscle you’re actually building. It isn’t prompting. It’s judging.

From the eye of the storm,
Grant