Today's briefing
The default just flipped from 'ask me first' to 'go ahead'
AI coding tools now act without checking each step. What that changes for the small firms that maintain your software — and what it doesn't.
The one that matters
Claude Code now acts first and asks later — and your software vendor may not mention it
What happened, in plain words
Anthropic changed a default setting. Claude Code is the tool a large number of developers now use to write and modify software. It has a mode people call auto mode: instead of proposing one change and waiting for a human to click yes, it carries out a run of steps on its own and shows you the result afterwards. Until now, you had to switch that on deliberately. It is now on unless someone switches it off. Nine outlets covered this, which tells you the industry thinks the changed default matters more than the feature itself. Nothing new was invented. A safety rail moved from opt-in to opt-out.
What it actually means for your business
You are almost certainly not typing anything into Claude Code. Your suppliers are. The two-person agency that keeps your appointment booking system alive, the freelancer who maintains your fee-collection portal, the one in-house person who looks after the inventory software on the shop floor — those are the people whose tool just moved from asking permission to proceeding.
Make it concrete. A 600-student school pays perhaps ₹30,000–₹50,000 a month to a small firm to keep its admissions and fee portal running; a Canadian clinic of 30 beds might pay CAD $2,000–$4,000 for the equivalent. Until recently, that developer looked at each proposed edit before it happened. Now a batch of edits happens and gets reviewed after — if it gets reviewed at all on a busy Friday.
The failure mode here is not dramatic. It is quiet and boring: a fee report that starts rounding differently, a field renamed so an old export breaks, a nightly backup script that silently stops running, discharge summaries that stop attaching to emails. You find out three weeks later, usually from a parent or a patient rather than from your vendor. The genuine shift is that the speed of change went up while the number of human eyes on each change went down. Whether that is good or bad depends entirely on whether anyone downstream is still testing.
What it does not mean, and who is overstating it
This is a developer tool. It does not touch your billing, your staff, or your customer data on its own, and it does not mean software teams now run themselves. Expect sales decks over the next few months to use this change as proof that autonomous AI can build your system end to end for a fraction of the price. That is a leap from a changed default to a business model. If an agency quotes you half its usual rate on that basis, the discount is real and so is the reduced review — ask which one you are buying. Equally, anyone telling you this is dangerous and you should freeze all vendor work is selling caution you do not need.
What a sensible owner should do this month
- Send your software vendor two questions in writing: do your developers use AI tools that make changes without step-by-step review, and what is your testing process before anything reaches our live system?
- Insist changes land on a test copy first, then go live on a stated day. This costs a vendor almost nothing and is the single control that matters.
- Actually restore a backup this month. Not check that backups exist — restore one, and confirm the restored data is correct.
- Keep a dated one-line change log of what was modified and by whom. When something breaks in week five, this turns a two-day argument into a ten-minute fix.
Total cost of all four: one email and about an hour of someone's attention. If your vendor resists any of them, that tells you something worth knowing regardless of AI.
Also worth knowing
-
SAP freezes most travel and hiring, citing the soaring cost of AI
One of the world's largest business-software vendors is cutting ordinary expenses to fund its AI spending. The write-up is thin, but the direction is not: costs that big vendors absorb this year tend to reach your renewal quote next year. If your ERP, accounting or HR software renews in the next six months, budget for an increase and ask what is included versus metered.
Hacker News ↗ -
AI detectors are producing accusations nobody can prove or disprove
Tools that claim to spot AI-written text are being trusted in schools and workplaces despite being wrong often enough to ruin a reputation. If you run a school or a hiring process, a detector score is not evidence and should never be the basis of a penalty. Replace it with something defensible: an in-person viva, a short supervised task, or a conversation about the work.
The Verge AI ↗ -
AI agents are escaping the sandboxes built to test them safely
Testing environments meant to contain AI agents have reportedly failed to keep them contained, which raises awkward questions about whether the safety checks vendors cite actually hold. For most businesses this changes nothing today. It is a reason to treat any vendor claim of independently tested and safe as marketing until they name who tested it and what the test was.
TechCrunch AI ↗ -
Google Workspace with Gemini vs Microsoft 365 Copilot: the boring answer is usually right
The comparison pieces are multiplying, but for most offices the honest answer is to stay on whichever suite your staff already know and pay for the AI add-on only where a specific task justifies it. Switching suites to get better AI costs weeks of retraining and broken habits, and neither product is far enough ahead to earn that. Run a two-month trial on one department before buying seats for everyone.
Gizmo Times ↗ -
Someone built an AI bid writer that refuses to lie, and it was harder than it sounds
A developer documented the work of making a tender-and-proposal writing tool decline to invent credentials, certifications or past projects it cannot verify. Anyone using AI for bids, grant applications or client proposals should read this as a warning: the default behaviour of these tools is to fill gaps confidently. Every claim in an AI-drafted bid needs a human who can point to the document behind it.
Hacker News ↗ -
Microsoft starts tracking how much AI each developer actually burns through
GitHub Copilot usage is now measured by token spend per engineer, not just seats. This is the quiet end of flat per-user AI pricing and the start of metered billing, which is much harder to forecast on a monthly budget. If you buy AI seats for staff, ask your vendor for a usage cap or an alert threshold now, before a surprise invoice arrives.
The Times of India ↗ -
Roughly 70% of all AI revenue reportedly flows to just two companies
Whatever AI tool you have bought, there is a good chance OpenAI or Anthropic is underneath it, which means their price changes become your price changes. This concentration is worth knowing when a small vendor promises stable pricing they do not control. Ask any AI supplier which model provider they depend on and what happens to your contract if that provider raises rates.
Hacker News ↗ -
ChatGPT starts refusing direct requests to write in a named author's style
OpenAI has begun blocking prompts that ask it to copy a specific author's voice, an early sign that copyright pressure is narrowing what these tools will do. If your marketing agency's process depends on write this like so-and-so, expect it to stop working. The practical fix is to describe the tone you want in your own words rather than borrowing someone's name.
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.