Today's briefing
The software that keeps working after you leave
Microsoft rebuilt Copilot around agents that run unattended. The useful part is real, the sales pitch is not, and week one is permissions.
The one that matters
Microsoft's Copilot can now keep working while nobody is at the desk
Microsoft has overhauled Copilot around what it calls Autopilot agents. Until now Copilot was something you typed at: you asked, it drafted, you edited. The new version can be handed a task and keep going on it while the person who assigned it is away from the desk. The same overhaul adds tools that let people who cannot program assemble small apps inside Microsoft 365. Trade press covered the change across several outlets; investor commentary read it as Microsoft finally settling on a plan for Copilot, and one forecast expects paid Copilot seats to double within twelve months. Separately, a channel-industry piece warned that smaller companies need their data permissions in order before switching any of this on. That warning is the most useful thing in the whole announcement.
What it actually changes in your office
Two shifts matter. The first is that the unit of work changes from a sentence to a job. A typing assistant saves a clerk ten minutes. A job that runs unattended can own the whole dull tail of an office week:
- A workshop with 40 staff: chasing unpaid invoices, consolidating timesheets from supervisors who mail them in as photos, turning a customer's spec email into a first-draft quotation.
- A 600-student school: fee reminder letters, attendance exception lists, first-pass sorting of the parent inbox so the principal sees the twelve mails that need a human.
- A 30-bed clinic: drafting discharge summaries from notes, assembling claim paperwork, reconciling which files are missing a signature.
The second shift is less pleasant. An agent working unattended sees exactly what the person who started it can see. If your shared drive is one flat folder everybody can open, an agent asked to summarise salary trends will do it, correctly, for whoever asked. Nothing in the announcement fixes your folder permissions. That is your job, and most small offices have never done it.
What it does not mean
It does not mean fewer people. Nothing announced changes who is accountable for a wrong invoice, a wrong fee notice or a wrong medical summary, which means somebody still reads the output before it leaves the building. The prediction that paid seats will double is a stock analyst's view on Microsoft's revenue, not evidence that any of this works in a 40-person firm; it tells you about share prices, not about your back office. Resellers will tell you agents run your paperwork. They run a template that you have to specify, test and supervise, and specifying it properly is a week of somebody's attention. Expect the usual gaps too: these features arrive first in the larger licence tiers and work best in English. If your records live in Hindi, Marathi or French documents, or mostly in WhatsApp and Tally, much of this is not reachable for you yet. And an announcement is not a switch flipped in your tenant today.
What a sensible owner does this month
- Spend week one on permissions, not prompts. Ask your IT provider for a report of who can see which shared folders and sites, and close the obvious everybody-can-see-everything cases. This is worth doing even if you never buy Copilot.
- Do not buy seats for everyone. Pick the two or three people who type the most: the accounts clerk, the office administrator, the person who writes quotations.
- Test one task for 30 days and measure it. Hours spent on that task before, hours after. If you cannot name the hours, you cannot justify the renewal.
- Run the arithmetic at your own quoted price. Seats times monthly price times twelve, against those measured hours. Ten seats at even a modest per-seat rate is a real annual line item.
- No unreviewed outbound. For the first quarter, nothing an agent writes goes to a customer, parent, patient or tax authority without a human signature.
Also worth knowing
-
Attackers are actively exploiting a critical Zimbra mail server flaw
A flaw in Zimbra lets an attacker run commands on the server by sending a single crafted email, and it is already being used to steal mail. Plenty of schools, clinics and small firms run Zimbra because it was the affordable on-premise mail option; if that is you, patching is this week's job, not next quarter's. It has nothing to do with AI and is still the most urgent item in today's edition.
Ars Technica ↗ -
Claude went down for many users, and plenty of workflows went with it
Anthropic's Claude was unavailable for many users, widely enough that mainstream tech press covered it. Nothing was permanently lost, but if a daily task in your business now runs through one AI provider, you are one outage away from doing it by hand. Write down the manual fallback for each such task now, while nobody is panicking.
TechRadar ↗ -
Anthropic admits it does not know who is liable when an AI agent misbehaves
In an exclusive interview, Anthropic said the legal risk from its agents acting badly is genuinely unsettled. That is unusually candid from a vendor, and you should read it as a statement about your exposure rather than theirs: if a wrong invoice or wrong patient summary goes out on your letterhead, the complaint arrives at your door. Keep a named human signing off anything that touches money, health or a legal deadline.
WTVB ↗ -
Google ships Gemini 4 Argon, and developers spend the day arguing about it
Google released a new top-end model and it dominated technical discussion, with hundreds of comments and coverage across several outlets. Little concrete detail has settled yet, so there is not much worth repeating. For an owner the effect is second-hand and slow: whoever supplies your software gets a better engine in a few months, usually at the same price.
Hacker News ↗ -
OpenAI's developer day: more than twenty announcements, including GPT-6 Astra
OpenAI announced a new flagship model plus a long list of updates to its coding tools, APIs and security features. Nearly all of it is aimed at people who build software, not people who run businesses. The one question worth sending your software vendor: does anything here change what you already pay for, and is anything you depend on being retired?
OpenAI ↗ -
Why consumer AI products keep getting more expensive to run
An analysis of why the big labs have gone quiet on consumer AI: every use costs real money to serve, and the sums do not work at consumer subscription prices. The practical read-through is that free tiers will keep shrinking and per-seat prices will drift upward. Do not build a process your office depends on around a tool that is only affordable because someone is subsidising it this year.
TechCrunch AI ↗ -
A HIPAA-compliant AI connector for clinics, and why the badge may not help you
A vendor has shipped a Microsoft-certified connector that drops its healthcare AI into Power Automate, Power Apps and Copilot Studio, aimed at clinics automating paperwork. Read the fine print: HIPAA is an American law. If your clinic is in Pune or Peterborough, the questions you must actually answer concern India's DPDP Act or Canada's PIPEDA and your province's health privacy rules, and a HIPAA certificate answers none of them.
manilatimes.net ↗ -
A university course publishes what AI actually did to its exams
A computing course released a retrospective on AI-assisted cheating in its assessments, and it drew substantial discussion. If you run a school or a training institute, this is the document your academic head should read instead of the next detection-software pitch. No tool on the market reliably identifies AI writing; the workable fix is in how you set, weight and supervise the work.
Hacker News ↗
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.