Today's briefing
The assistant just got permission to press send
Claude can now send email from Gmail without asking each time. The capability is real; handing it your main inbox this month is not the sensible move.
The one that matters
Your AI assistant can now press send in Gmail without asking you first
What happened
9to5Google reports that Claude, Anthropic's assistant, can now send email from Gmail — and that it can do so without stopping for your approval on each message. The write-up is short and there is little more detail than that, so take the specifics lightly. The change underneath it is not slight at all: an assistant that used to write a draft for a human to read and send has been given permission to do the sending itself, outward, to a real person, under your name and your business's address.
What this actually means for your business
Every AI tool most owners have used so far has been a suggestion machine. It wrote the fee reminder, the quotation, the discharge summary — and a person read it and pressed send. That human step was doing far more work than anyone credited it for. It caught the wrong recipient, the wrong amount, and the sentence that promised a delivery date you cannot meet.
Take a 30-bed clinic with one shared front-desk inbox. An assistant that can send is genuinely useful there: appointment confirmations, "your report is ready", follow-up reminders after discharge. It is also one loose instruction away from mailing a patient's result to the wrong address, which in most places is a disclosure you must report, not a typo you can apologise for. A 600-student school has the same shape — reminders to parents are close to the perfect job for this, and fee-arrears letters sent to the wrong family are a perfect way to spend a week on the phone. A workshop with 40 staff sends purchase orders and price confirmations all day; a supplier who receives a number by email treats it as a commitment, whether a person or a piece of software typed it.
The useful property and the risky property are the same one. An agent is only worth having because it does not stop to ask. Cost is not the constraint here — assistant subscriptions of this kind typically run roughly ₹1,500–₹8,000 or CAD $25–$100 per user per month. Your exposure is not on the invoice. It is in the mailbox.
What it does not mean, and who is overstating it
It does not mean AI now runs your correspondence, and it does not mean anyone should connect the owner's inbox this week. It does not replace a receptionist or an accounts clerk, both of whom do a dozen things of which email is only the visible edge. The people overstating this are the consultants and resellers selling "autonomous agents": the demo answers a customer email in four seconds and never shows you the audit trail, the undo path, or the month it wrote to the wrong parent. Vendor case studies about AI clearing years of engineering work are being quoted at owners as though writing software tests and talking to customers carried the same risk. They do not.
What a sensible owner does this month
- Check what already has access. In Google Workspace, review connected apps and third-party access on your staff accounts. Do this even if you approved nothing — staff connect tools themselves.
- If you want to try it, use a side mailbox. A notices or reminders address. Never the owner's account, and never the one customers reply to.
- Write one sentence of policy. Which mail may leave the building without a human pressing send? A sensible default: routine internal reminders yes; anything touching money, health, legal wording or a promise, no.
- Be able to answer "what did it send, and to whom". Turn on retention and audit logging before you need it for a patient, a parent or a supplier.
Budget required: nothing. This is about two hours of administration. For most businesses the honest answer this month is "not on the main inbox yet" — but sending agents will be ordinary within a year, and these habits are far cheaper to build now than to retrofit after the first bad send.
Also worth knowing
-
Anthropic's revenue run rate passes $65 billion, according to a source
Reuters reports the figure on a single source, so hold it loosely. You buy nothing differently because of it, but it does tell you the money behind AI assistants now comes from company budgets rather than experiments. Expect the tools your staff use to keep changing price, features and terms every few months — plan contracts for churn, not stability.
Reuters ↗ -
OpenAI halves the price of one of its main models
The cost of running GPT-5.6 Sol was cut by 50%. If a vendor built you a chatbot, document reader or call summariser and bills you a markup on "AI usage", that line on your invoice should fall too — ask them, in writing. Falling prices are the most dependable pattern in this business, so avoid multi-year deals priced on today's costs.
Hacker News ↗ -
Google bought a collapsed airline's data at auction
Spirit's data went under the hammer after the airline failed, and Google bought it. The lesson for you is blunt: the data you hand a supplier outlives the supplier, and in insolvency it becomes an asset sold to whoever bids. Read what your scheduling, billing or CRM contract says happens to your customer records if the vendor shuts down — most owners never have.
Hacker News ↗ -
Asana says it replaced a testing system in two weeks with AI, for about $12,000
OpenAI reports Asana used Codex to redo work it had estimated at five years. Look at the shape of the task before the number: a narrowly defined, mechanically checkable engineering chore, inside a software company with engineers to verify the output. It is not evidence that AI can do five years of your business, but it is good evidence that AI is strong at bulk, boring, verifiable work.
OpenAI ↗ -
A fake think tank was reportedly set up to influence what chatbots say
Reporting says Israel created a bogus research outfit in a likely attempt to get its material into the sources AI chatbots draw on. Whether or not it worked, the method is cheap enough that others will copy it. Treat a chatbot's confident summary of anything contested — a competitor, a regulation, a supplier's reputation — as a lead to verify, never as a finding.
Hacker News ↗ -
Efforts to make AI text detectable are already being worked around
Anthropic wants AI-written text to be easier to spot, and Business Insider reports developers are building ways around it. The practical takeaway: any policy you write that depends on detecting AI writing — in admissions, hiring or assessment — will fail. Judge work by what someone can demonstrate in person or under supervision instead.
Business Insider ↗ -
ChatGPT gets a teen version with parental controls
OpenAI launched a teen-focused product with built-in protections, healthy-use features and parent controls. If you run a school, it is worth an hour with whoever handles your IT — but this is the vendor describing its own safeguards. Ask what is retained, who can see it, and whether it meets your local privacy rules before it reaches a classroom.
OpenAI ↗ -
Nobody has a clear picture of how people actually use AI
MIT Technology Review's argument is that even the companies building these tools lack solid data on real day-to-day use. Keep that in your pocket the next time a sales deck quotes adoption or productivity percentages at you. The honest move is to measure it yourself, on one task in your own business, and trust only those numbers.
MIT Technology Review ↗
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.