Today's briefing
The machine suggested it. Who signed off?
A thin but serious Pentagon admission about overreliance on AI carries a lesson every clinic, school and workshop should apply before its next rollout.
The one that matters
The Pentagon says it trusted AI too much — the same failure fits a clinic or a school
What happened, in plain words
The Pentagon has said that overreliance on artificial intelligence contributed to a missile strike that hit a school in Iran. That sentence is the substance of the report, and it is close to all of it. The account comes from a single outlet and drew a very large amount of discussion, which is not the same as being heavily corroborated. We do not know from the available material which systems were involved, what they recommended, or exactly where in the chain a person was supposed to object and did not. So we will not pretend to know.
We are leading with it anyway, because the shape of the failure is the shape of the failure that will eventually show up in a 30-bed clinic, a 600-student school or a workshop with 40 staff. Not the AI was wrong. That happens daily and is survivable. The dangerous version is nobody had been assigned to be wrong. The output arrived looking finished, it moved through a process that used to contain a pause, and the pause had quietly been removed because the tool was usually right.
What it actually means for your business
Every AI rollout in a small business removes a moment of friction, and that friction was doing work you were not tracking. A school that uses an AI tool to draft fee-defaulter lists and send notices has removed the clerk who recognised a family name and said "wait, they paid in cash last Tuesday." A clinic that runs consultation audio into an AI notes tool has removed the second reading where a dosage that looked odd got questioned. A workshop that lets a tool generate reorder quantities from last year's consumption has removed the storeman who knew one big customer left in March.
These tools are cheap — roughly ₹1,500 to ₹3,000 per user per month in India, roughly CAD 20 to 40 per user per month in Canada. The cost of the tool is never the exposure. The exposure is one wrong legal-sounding letter to a parent, one wrong dose in a chart, one wrong ₹4 lakh of stock sitting in your godown.
What it does not mean, and who is overstating it
It does not mean AI is uniquely dangerous. A wrong formula in a spreadsheet has closed more businesses than any model will this year. It does not mean you should pause your plans. Two groups will overstate this story in opposite directions. The first is the vendor selling "autonomous" agents that "run your back office" — ask them, plainly, who is accountable when the agent is confidently wrong, and watch the answer get vague. The second is the consultant who will arrive within a fortnight selling an AI risk audit or a governance package priced off this headline. You do not need either. What you need fits on one page.
What a sensible owner should do this month
- Write the one-page list: decisions where AI output may be used as a draft, and decisions where it may not. Money going out, medication, employment, legal notices and anything touching safety go in the second column.
- Put a human name on every AI-assisted output that leaves your premises. Not a department — a name. The person who signs is the person who reads.
- Before you trust a tool, run it against 20 past cases where you already know the right answer. Count the misses. If the vendor discourages this, that is the review.
- Ask each vendor one question in writing before renewal: when this is wrong, how do I find out before my customer does? File the answer.
- Keep the boring records. Thirty days of what the tool suggested and who approved it costs nothing and settles arguments.
Also worth knowing
-
Anthropic releases Claude Opus 5.5, its new top-end model
Anthropic shipped a new flagship model and roughly 59 outlets wrote it up, which tells you more about press interest than about your operations. For most businesses this changes nothing you must act on: whatever assistant or vendor tool you already pay for will pick it up without you doing anything. The one sensible use of the news is to retry a task you abandoned six months ago because the output was not good enough.
marketscreener.com ↗ -
OpenAI launches GPT-6 Sol and Luna, split by capability and cost
OpenAI released two models aimed at everyday work with different balances of capability and price. The useful signal is the split itself: the industry now expects you to route cheap, high-volume work to the cheap model and reserve the expensive one for hard problems. If your software vendor runs every request through the top tier and bills you for it, that is a conversation worth having at renewal.
OpenAI ↗ -
Microsoft plans deeper Copilot discounts alongside an AI 'super app'
The Information reports Microsoft will increase Copilot discounts as it launches a consumer-facing AI app. Read the discounting for what it usually is: take-up below what was hoped, which means the price you were quoted is negotiable. If you are on Microsoft 365 and have been pushed toward a multi-year Copilot add-on, wait a quarter and ask again.
TradingView ↗ -
Video ad tools aimed at small businesses ship on OpenAI's newest model
Higgsfield AI added video features built for small-business ad creation, using GPT-6 Astra. This is one of the few items today with a direct line to a shop, showroom or clinic that currently pays a videographer for festival and product clips. Test it on one campaign before you cancel anyone's retainer — these tools produce footage competently, but they do not know what your customers need to hear.
OpenAI ↗ -
OpenAI expands free learning paths for staff, managers and teachers
New structured courses aimed at employees, leaders, educators and students, with completion credentials. This is quietly the cheapest real move on today's list: two admin staff who can write a clear prompt will get you more than a third subscription will. If you run a school or a clinic with a front office, put one person through it this month and have them teach the others.
OpenAI ↗ -
An argument that AI has no judgment — and that your staff may lose theirs
A widely-read essay making the case that these tools supply answers but not wisdom, and that leaning on them erodes the judgment of the people who do. Worth twenty minutes because the real risk in a small firm is not a wrong output, it is a junior who never develops the instinct to notice one. Keep some work deliberately manual for the people you are training.
Hacker News ↗ -
A team rebuilt its testing pipeline because AI wrote code faster than it could check it
An engineering team found that AI-assisted coding flooded their automated test systems and forced an expensive rework. If you employ developers or retain a software agency, expect the same imbalance: more code arriving faster, with review and testing capacity unchanged. Budget for the checking, not just the writing — the bottleneck has moved.
Hacker News ↗ -
A chatbot's bitcoin 'prediction' gets written up as a forecast
Someone asked Microsoft Copilot where bitcoin goes by 2027 and the answer was reported as a prediction. It is not one; a language model produces plausible text, not a forecast, and has no stake in being right. Keep this as your template for what to ignore — and when a vendor demos AI 'predicting' your demand or your defaults, ask what data it was trained on and what its error rate was.
Cryptonews ↗
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.