Today's briefing
Read-only: the best AI advice of the year
The most regulated industry on earth just decided AI should read everything and touch nothing — a pattern any school, clinic or workshop can copy.
The one that matters
Wall Street puts Claude to work — and refuses to let it touch anything
What happened, in plain words
Anthropic released a version of Claude built for financial advisers. It was picked up by nearly thirty outlets, which tells you the industry took notice. The deployments named alongside it are the largest and most audited names in money management, including Charles Schwab and BlackRock. The detail worth your attention is not the product announcement. It is the fine print: the Schwab arrangement reportedly limits Claude to read-only access. The assistant can look at client data. It cannot change it, move it, or send anything out.
What this actually means for your business
An industry where a single wrong entry triggers a regulator, a lawsuit and a headline sat down, looked hard at AI, and concluded the safe shape is: it reads, a human writes. That is not timidity. That is the most useful deployment pattern anyone has published this year, and it costs you nothing to copy.
Concretely. A 600-student school has fee ledgers, attendance registers, exam records and three years of parent emails. A read-only assistant can answer in seconds: which forty families are more than two months behind on fees, which students crossed the attendance threshold, what did we promise this parent in March. It does not send the reminder, does not touch the ledger, does not edit a report card. The office manager still does that — but starts from an answer instead of a spreadsheet.
A 30-bed clinic can have it read discharge summaries against insurer rules and draft the claim, flagging what is missing before submission. A human presses submit. A workshop with 40 staff can have it read job cards and purchase orders and say which six jobs have slipped past their promised date and why. It does not raise the purchase order.
The pricing is ordinary: business seats for this class of assistant run roughly US$20–30 per person per month. A ten-person back office is a few hundred dollars a month, not a capital project.
What it does not mean, and who is overstating it
It does not mean AI is now trusted to give financial advice, or that these firms handed over judgement. The guardrail is the story, and it will be the first thing dropped from the sales pitch. Expect resellers over the next quarter to cite these names as proof your business should buy an autonomous agent that acts on its own. The same institutions being quoted at you declined to allow exactly that.
Also note how unsettled the vendor side is: Salesforce reportedly began building its own model weeks after signing a Claude deal. Nobody's stack is final. That is an argument against long lock-ins, not against starting.
What a sensible owner does this month
- Pick one repetitive reading-and-summarising task that currently eats a person's afternoon.
- Give the tool exports — a spreadsheet, a folder of PDFs — not live credentials to your accounting or patient system.
- Write down, in one line, who checks the output before it leaves the building.
- Run two weeks with one capable staff member. Count minutes saved and mistakes caught.
- Do not buy seats for everyone, and do not sign past twelve months.
- Ask every vendor one question: what can this write, and who approves it? If they cannot answer plainly, you have your answer.
Also worth knowing
-
Fake dating apps used Claude to scam 25,000 people
Anthropic says operators of fraudulent dating apps used its model to run scams against roughly 25,000 victims. The lesson for you has nothing to do with dating: convincing, personalised, well-written fraud now costs a criminal almost nothing to produce at volume. Reinstate the boring control — any change to supplier bank details gets verified by a phone call to a number you already had, not the one in the email.
Yahoo Tech ↗ -
An AI agent that got it right once may not get it right twice
A widely shared piece argues that agents are judged on impressive one-off demos while their actual reliability across repeated runs goes unmeasured. This is the single most useful idea to hold in your head when a vendor demos something in your office. Before anything runs unattended, make them run the same task twenty times on your real data and count the failures — the demo told you the best case, not the average.
Hugging Face ↗ -
Document software starts plugging itself into Claude, ChatGPT and Copilot
Templafy has wired its document tools into the three main assistants using MCP, the emerging standard for connecting business software to AI. Templafy itself is built for large enterprises and is not for a 40-person firm, but the pattern matters: your existing accounting, HR or ERP vendor is likely to ship a similar connector. Wait for that rather than buying a separate AI product — it will be cheaper and it already knows your data.
manilatimes.net ↗ -
Google ships Gemini 3.8 Live, aimed at real-time voice
Google released new Gemini models built around live interaction and longer reasoning. This is relevant if your business lives on the phone — an admissions desk, an appointment line, a service counter handling multiple languages — because that is where voice AI is improving fastest. For everyone else it is an upgrade to tools you may already pay for, and requires no action.
Google DeepMind ↗ -
Expensify integrates with Anthropic for expense handling
Expense reporting is exactly the sort of dull, rule-heavy, universally hated task that AI handles well. If you have staff who travel and an admin who spends days chasing receipts, this class of integration pays for itself faster than anything strategic. Check whether the accounting tool you already use is announcing the same thing before you switch vendors.
Business Travel Executive ↗ -
Researchers got admin access to an AI company's production code in 25 minutes
Security researchers obtained administrator access to Baseten's production GitHub in under half an hour. Every AI vendor you sign becomes part of your security perimeter, and most of them are young companies moving fast. Ask where your data is stored and who on their side can read it — and while you are at it, check who still has admin rights on your own shared drive and code repos.
Hacker News ↗ -
Anthropic's CEO talks slowdown just before an expected IPO
Commentators are debating what Dario Amodei's suggestion of slowing AI progress means for Anthropic's anticipated listing. Operationally this changes nothing for you this quarter. It is, however, a decent reason to refuse any contract priced on capabilities a vendor promises will arrive next year.
cnbc.com ↗ -
Windows 11 users can finally reclaim the Copilot key
Microsoft is letting users remap the dedicated Copilot key that ships on new laptops. Trivial news, genuinely useful if you have just bought machines for a computer lab or an office and staff keep hitting it by accident. Two-minute fix, one less daily irritation.
Neowin ↗
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.