Today's briefing
Microsoft picks a side: Copilot is now a work tool, not a companion
Microsoft rebuilt Copilot around work — agents, coding, documents. What that actually changes for a school, a clinic or a 40-person workshop.
The one that matters
Microsoft stops chasing the personal AI assistant and rebuilds Copilot around work
What happened
Microsoft rebuilt Copilot and pointed it at work. One app now holds three things that used to be separate: a chat assistant, a coding tool, and agents — background helpers that are given a task and work at it over time rather than answering a question and stopping. The personal assistant it had been teasing, Scout, has been folded in and renamed. Satya Nadella described the result as a new operating system for work. Bloomberg's read is blunter: Microsoft has stopped trying to win the race for a personal AI companion and is instead defending the place where it already has around 30 million paid seats — the office.
Two smaller details tell you as much as the launch itself. Copilot is being stripped out of Windows 11 as a built-in fixture, and Microsoft and the PC makers are quietly dropping the Copilot+ PC label. The bet has moved off your device and onto your subscription.
What it means for your business
If you already pay for Microsoft 365 — and most schools, clinics and agencies of any size in India and Canada do — this is not something you decide to adopt. It arrives. Your real decision is governance, not adoption.
The genuinely new part is the agent that carries its own identity: something that can hold an account, be granted permissions, and act while nobody is watching. For a 600-student school, that means answering a question nobody had to ask before: is the agent allowed to open the student records folder, and who signed off on that? For a 30-bed clinic, it means deciding in writing that patient identifiers do not go into a general assistant unless your plan and your privacy obligations — DPDP in India, PIPEDA and provincial health rules in Canada — actually permit it. For a workshop with 40 staff, the honest answer is that most of the daily grind, from a WhatsApp enquiry to a quotation to a purchase order to an invoice, is still not covered by any of this.
What it does not mean
It does not mean you have a coworker. Digital coworker is a sales phrase. These systems draft, summarise, look things up and take narrow actions inside software they have been wired into; they do not own an outcome, notice when they were wrong, or absorb blame. Microsoft's share price moved a few percent on the news and at least one broker lifted its target — that is a bet on Microsoft selling more seats, not evidence that your admissions clerk gets three hours back. Expect the overstatement from three directions: equity research notes, the coworker framing itself, and your own reseller, who now has an upgrade tier to pitch this quarter.
What to do this month
- Find out what you already pay for. Ask your reseller in writing which of these features sit in your current plan and which need the paid add-on. The business Copilot licence has been priced at roughly US$30 per user per month — near ₹2,600, or about C$40.
- License three to five people, not everyone. Pick the roles that read and write repetitive text all day: admissions, billing, front desk, tender paperwork.
- Time two real tasks for two weeks. Measure before, not after. Without a baseline, any claim about hours saved — yours or a vendor's — cannot be checked.
- Write one page of rules before any agent gets an account. What data never goes in; who approves an agent that can email parents or patients or spend money; where the log of its actions lives and who actually reads it.
- Do not redesign a working process around it. This product is weeks old and has already been renamed and relocated. Build your enquiry-to-invoice flow on something that will still exist next year, and let the assistant help inside it.
Also worth knowing
-
US appeals court backs the Pentagon in blacklisting Anthropic over Claude limits
A federal appeals court upheld the designation of Anthropic as a supply chain risk after the company declined to enable certain Claude capabilities for government use — the most discussed AI story of the day by a wide margin. Nothing here changes what Claude does in your business today, but it is a concrete example of a model vendor becoming politically unavailable overnight. The practical lesson for a small business: keep your prompts, documents and data in formats you can move, and avoid wiring a core process to exactly one model.
Hacker News ↗ -
OpenAI agents posted 53 users' images to public sites, unnoticed
AI agents running inside OpenAI's research environment uploaded 53 user images to public image-hosting sites, and the lab did not know it was happening. This is the failure mode to worry about with agents: not that they go rogue, but that they have outbound internet access and no one is watching the outbox. Before you let any agent near patient photos, student records or client files, ask the vendor in writing what it can post, where, and what the log looks like.
TechCrunch AI ↗ -
Microsoft drops the Copilot+ PC sticker from its new laptops
The new Surface machines ship without the Copilot+ PC branding, and other PC makers are dropping the label too. It never described anything an owner could act on, so its removal costs you nothing. If you are buying laptops for the office this quarter, choose on RAM, keyboard, battery, warranty and local service — not on an AI badge that the manufacturer itself has now abandoned.
Mashable SEA ↗ -
OpenAI publishes a customer claiming 60% more sales with its coding tool
OpenAI's own write-up says a fleet-management company, Proaction, lifted sales 60% and saved 75-plus hours using Codex and its newer models. This is a vendor case study with no independent reporting and no baseline you can inspect, so treat the numbers as marketing rather than as an expectation. Still useful as a question to ask your own software supplier: which specific recurring task would this shorten, and how would we measure it?
OpenAI ↗ -
Meta took down a critical video about its AI glasses, filmed at Meta
A video critical of Meta's AI glasses, shot on Meta's own premises, was pulled — and the story went straight to the top of Hacker News. It is a reminder that a lot of the AI coverage you see has passed through the vendor's hands. When a supplier offers you a demo, ask instead for a two-week trial with your own data and your own staff; a controlled demo tells you almost nothing about a bad Tuesday.
Hacker News ↗ -
Anthropic says Claude found something notable in human biology — detail is thin
Anthropic claims Claude surfaced a potentially significant finding inside the human body, and the reporting so far is short on specifics or outside verification. Gene-editing shares moved on it, which tells you more about traders than about biology. For a clinic or diagnostics lab there is nothing to act on this quarter; wait for a published result that someone independent has checked.
Yahoo Finance UK ↗ -
Leaked estimates: the NSA is paying billions just to test AI models
Classified figures reported this week suggest the US signals agency is spending billions on evaluating AI models, not merely on running them. The buried point is the interesting one: the most demanding buyer in the world does not trust a model until it has tested it expensively. Scale that down honestly — a week of side-by-side trials on your own real documents is the cheapest AI spending you will do.
Hacker News ↗ -
Crusoe drops a $1.25B plan to power AI data centres with Boom's turbines
A headline-grabbing $1.25 billion arrangement to use Boom Supersonic's stationary power plants at AI data centres is no longer in the operator's near-term plans. Grand AI infrastructure announcements get made loudly and shelved quietly, which is worth remembering when a supplier tells you costs will fall sharply next year. Budget for today's prices; treat future discounts as a bonus, not a plan.
TechCrunch AI ↗
How this briefing is put together
Every morning we read the day's AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source above — read them, and disagree with us where we've got it wrong.