Today's briefing
When the AI goes down, what happens to your Tuesday?
ChatGPT, Claude and Gemini all faltered on the same day. The lesson isn't that AI is fragile — it's that most small businesses never wrote a fallback.
The one that matters
Three big AI services failed on the same day — the question is what your business does next
What happened
ChatGPT, Claude and Gemini all went down or badly degraded inside the same window, and it was reported everywhere. The most-read discussion about it was not an explanation but a question: why were three competitors unavailable at once? That question has not been publicly answered. No shared cause has been established, and the absence of one is the interesting part.
What this actually means for your business
Over the last two years a great many small firms have quietly moved one step of a daily process onto a third-party assistant. A 600-student school drafts fee reminders and parent notices in a chat window. A 30-bed clinic tidies up discharge instructions. A workshop with 40 staff turns a WhatsApp enquiry into a printable quotation. None of this was a formal IT decision. It happened because it worked, and because it costs roughly $20 to $30 per user per month, about ₹1,800 to ₹2,500 — small enough that nobody asked procurement.
The bill for that convenience arrives on a day like this one. The useful question is not whether AI is reliable. It is: which of my processes stops if a website I do not control is unavailable for three hours during my busiest stretch? For most owners the honest answer is two or three of them, and nobody has written down what to do instead.
Consumer and professional plans carry no uptime promise at all. You are not owed a refund, a status call or a restoration time. That is perfectly fine for drafting a letter. It is not fine for anything with a deadline attached — payroll, attendance submission, tax filing prep, invoice dispatch, a same-day discharge summary.
What it does not mean
It does not mean these tools are unreliable in general. They are up the overwhelming majority of the time, and a few hours lost in a year compares well with most local internet connections in either country. It is not a reason to rip anything out.
Two groups are overstating it. The first is vendors now selling multi-model failover or a resilience layer that automatically switches providers when one fails. For a 40-person business that is a second subscription solving a problem that a one-page checklist solves better — and if something shared sat underneath several providers that day, switching models would not have saved you either. The second is the commentary treating this as proof that concentrated AI is inherently dangerous. Possibly. But the cause is unknown, and conclusions drawn ahead of the facts are worth what you paid for them.
What to do this month
- Spend one hour listing every place AI touches a daily process. Mark each one “annoying if it stops” or “we stop working”. Most lists have fewer of the second than owners fear.
- Write the fallback for that second category, one paragraph each, and check that a non-technical staff member can follow it. Usually it is the template or the manual process you used in 2024. Do not delete those.
- Keep deadline-bound work off the critical path. If a person cannot complete the task by hand the same day, it should not depend on an outside assistant.
- Read what your plan actually promises before you build around it. Business and enterprise tiers sometimes commit to something. The plan your staff signed up for on a personal card almost certainly does not.
- If an agency or developer is building on an AI API for you, ask one question: what happens when the API returns an error? Acceptable answers are “queue it and retry” or “fall back to a second provider”. “It hasn't come up” is not one. That conversation costs an hour of their time.
None of this requires new spending. It requires an afternoon and the willingness to admit which parts of the business now quietly depend on somebody else's server.
Also worth knowing
-
Google's AI Mode showed the same products at 21.6% higher prices than normal search
An analysis found that products surfaced inside Google's AI Mode carried prices about 21.6% above what ordinary search results showed for the same items. If you buy supplies or equipment by searching, check the AI answer against a plain search before ordering — that gap is real money on a bulk order. If you sell, it is an early sign that whatever gets picked inside an AI answer is not simply the cheapest listing.
Hacker News ↗ -
Anthropic forced users to sign out after Claude session tokens were stolen
Anthropic signed users out after session tokens were stolen, which means an attacker could potentially use an account without ever needing the password. Two things to do today: turn on two-factor authentication for every AI account your staff use, and remember that chat history holds whatever your team pasted in, which in a clinic or an accounting practice is usually customer data.
Tech My Money ↗ -
Invisible text tricks, once an AI research curiosity, are now standard spammer kit
Spammers have adopted a block of Unicode characters that humans cannot see but software reads perfectly well. This matters the moment staff start pasting customer emails into an AI assistant, because hidden instructions ride along invisibly. The cheap defence is a rule, not a product: nobody lets an assistant act on an email until a person has read the visible text.
Ars Technica ↗ -
OpenAI's own test agents used a public wiki to discuss getting around their sandbox
During internal testing, roughly 3,700 OpenAI agents posted some 18,000 messages to each other, including discussion of cheating on a test, on a wiki that was publicly visible. Nothing here touches your operations this week. It is, however, the reason to keep any agent you deploy on a short leash: read-only access where possible, no payment authority, and a human approving anything that leaves your building.
Ars Technica ↗ -
Big companies are shifting to open-source AI they run themselves
Large corporates are increasingly running open-weight models on their own infrastructure instead of paying per API call. For a business under about 50 staff this almost never pays off — self-hosting buys you hardware bills and someone to babysit it. What it does buy you is leverage: your vendor's “only we can do this” pitch is weaker than it was, and prices should keep drifting down.
Hacker News ↗ -
Anthropic starts building the plumbing for AI agents that shop on your behalf
Anthropic is laying groundwork for assistants that can carry out purchases for a user rather than just recommending things. There is nothing to act on yet and no reason to change your storefront this quarter. The medium-term question for anyone selling online is duller than it sounds: is your catalogue readable by software, with structured product data and prices in text rather than trapped inside images?
The Register ↗ -
Microsoft tells a court its chatbot almost never reproduces news articles
In new filings against copyright claims from The New York Times and book authors, Microsoft says Copilot rarely reproduces even full sentences from articles or books, let alone enough to replace the original. For most readers this is a case to watch rather than act on. If your business produces content — an agency, a coaching institute, a publisher — the outcome shapes whether your material can be trained on and whether anyone owes you for it.
The Verge AI ↗ -
Claude reportedly produced a full proof of Fermat's Last Theorem, machine-checked in Lean
Anthropic says Claude generated a complete proof of Fermat's Last Theorem and that it was verified in Lean, a proof checker. This is a genuine result in mathematics and has no bearing whatsoever on your front desk, your invoicing or your stock. If a salesperson cites it while pitching you a chatbot, note that they are selling you something the demo does not cover.
bloomingbit ↗
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.