Read time: 5 minutes

Leader, leader, welcome back.

Two things happened this week that look like tech news and are actually governance news.

Microsoft's Cowork agent went live for everyone, and the most capable AI model on the planet got switched off three days after it launched.
Here is what both mean for you.

Microsoft Copilot Cowork came out of preview this week and went generally available.

Quick version for anyone who has not met it yet:
Cowork is an AI agent that lives inside Microsoft 365. You give it a whole multi-step job, it goes off and works across your apps, files and the web on its own, and it comes back when the lot is finished.

That is the part worth slowing down on. This is not another chatbot.
You hand it the job, it runs end to end, and it hands back the finished thing.
Not a draft you then fix for an hour. The thing, done.

I have used it a lot, and it has changed how I work.
A task that used to eat a full day was finished in about an hour last week, and I was getting other work done while it ran in the background. Since I got access, my use of normal Copilot has shrunk to quick questions, meeting summaries and editing docs. Everything heavier goes to Cowork.

That is the shape of work it is built for.
Long, multi-step jobs you would normally block out half a day for.

A few worth standing up first:
- An agent that researches a company, market or competitor and hands back a one-page briefing with the headline points up top
- One that turns the week's meeting transcripts into a clean list of what was promised, what is still open and who owes what by when
- An agent that drafts a first version of a proposal or report from your own templates and past work.

The bigger prizes, checking a stack of contracts against your policy and flagging every deviation, or building the board narrative straight from your numbers, are real too, but they touch your most sensitive data, so hold those until the question below is settled.

So the new skill is not prompting. It is judgment. Knowing which work belongs to an agent and which does not. You would not send a Ferrari to the shop for a pint of milk. Quick stuff stays in normal Copilot. Save Cowork for the heavy lifting.

Two things to settle before you switch it on, though.

First, where your data goes. Cowork currently runs on Anthropic's models, which sit outside the EU data boundary and process on US infrastructure. For EU tenants those models are off by default, and the moment you enable them your data leaves the boundary. If you are in a regulated function, that is a decision to make before the first task runs, not after. Microsoft has said an in-house model, Cowork 1, is coming, which may change the picture. Until it does, I would keep the most sensitive work, your contracts and your due diligence, parked, and point Cowork at lower-risk jobs first.

Second, the money. A Copilot licence on its own will not run it. Cowork is metered, billed per task, so you need pay-as-you-go or a prepaid commitment in place.
If you were in the Frontier preview, you have a grace period until 1 July.
The good news is the spend controls are real, set at tenant, group and user level, and it is off by default. This does not have to run away from you if admins set it up properly.

On 9 June, Anthropic shipped Claude Fable 5, the most capable model it had ever made public, topping the standard benchmarks. Three days later it was gone.

The US government issued an export-control directive on national security grounds, barring access for any foreign national, including Anthropic's own staff.
Since you cannot reliably check who is a foreign national at a login screen, Anthropic pulled Fable 5 and its locked-down sibling Mythos 5 for everyone, and rerouted traffic to the older Opus 4.8. As far as anyone can tell, it is the first time a government has done this to an AI model.

The lesson is not about Anthropic. A model that was approved and in production on Tuesday was gone by Friday, for reasons nobody inside your building had any part in or power over.

You already run the rest of the business by the opposite rule. You would never push treasury through one bank with no backup, or single-source a critical part with no alternative. An AI model in a core workflow is no different. Keep a second model approved and working. Know which of your processes could move to a fallback tomorrow, and which would simply stop.

The cleanest way to think about it came from the developer Shadcn, who said to treat intelligence as borrowed. Use the best model while you have it to make durable plans, specs and notes, then execute later with something cheaper, or something you control. Capability is now the easy part. Every frontier model is extraordinary, and the most extraordinary one just vanished over a long weekend.

If you have ever been handed a 40-page report and told to get across it by morning, this one is for you.
NotebookLM, Google's research tool, will turn your own documents into a mind map you can actually navigate.

Drop your sources into a notebook, choose the Mind Map option, and give it a few seconds. It builds a branching map of the whole thing in the Studio panel.
You open or collapse any branch with the arrows on each node, and download the lot as an image when you are done.

The clever part is what happens when you click a single node. Instead of a vague summary, it pulls the exact passage from your source that the node came from and explains it. So you get the big picture and the receipt for it in the same view, which is the bit most AI summaries quietly skip.

Where it earns its place: board packs, long contracts, a stack of policy documents, anything you need to understand fast and still be able to trace back to the source.

It is the difference between an AI that tells you what a document says and one that shows you where it said it.

A price war may be coming. The Wall Street Journal reports OpenAI is weighing cuts to its token prices ahead of expected cuts from Anthropic.
Welcome news if you have watched AI spend climb. And loyalty is thin in this market: Ramp's data shows the heaviest AI adopters run several vendors at once, not one. That is the don't-marry-a-model lesson with a receipt!

For the first time in World Cup history, every team has the same data.
FIFA and Lenovo have given all 48 sides Football AI Pro, a tactical assistant trained on more than 2,000 metrics and petabytes of football data, from individual sprint speeds to how a defence closes a passing lane. It reads opponent patterns in real time and flags tactical tweaks and substitution windows off live fatigue markers.

What it kills is the old edge, where only wealthy federations could afford analytics on that scale. A debutant nation now has the same depth as a perennial powerhouse, so matches turn on ingenuity, not on who spent more.
That is this whole edition in a sentence: when the tool is equal, the only edge left is who uses it best.