Today's briefing
Your AI Finally Remembers You — Now Decide What It Should Know
Claude's chat and Cowork app now share memory. That saves a school office real hours a week — and makes shared logins a problem to fix this month.
The one that matters
Your AI assistant now remembers your business between sessions — and that cuts both ways
What happened
Anthropic has joined memory across its two products. What you tell Claude in a chat now carries into Cowork, its work app, and what the app learns carries back. Until now every conversation started cold: you re-explained your business, your customers and your rules from scratch, every single time. That re-explaining is gone. Several outlets covered the change this week, and it is worth ninety seconds of your attention even if you have never opened either product.
What this actually changes for you
The cost of using AI at work has never really been the subscription. It has been the briefing. Consider a 600-student school where the office manager drafts fee reminders, circulars to parents and transport-route notices. Every draft began with the same preamble: we are a CBSE school, the term ends in March, fees are payable in three instalments, do not sound threatening, sign it from the principal's office. Five minutes of typing before any useful work started, ten or fifteen times a week. With memory, that is said once.
The same shape turns up everywhere:
- A 30-bed clinic writing discharge instructions in English and the local language, in the same plain wording the doctors have already agreed on.
- A workshop with 40 staff producing quotations, where the assistant already knows your margin bands, your standard delivery promise, and the three things you never put in writing.
- A two-person agency answering tenders, where the boilerplate about company history and past clients stops being pasted out of a five-year-old document.
Rough cost, since nobody quotes it plainly: the paid tiers these features live in run about US$20-30 per person per month, roughly ₹1,800-2,600 or C$28-42. For one office manager that is less than an hour of their time. For fifteen staff it is a real line on the P&L and deserves a decision rather than a drift.
What it does not mean, and who will oversell it
Memory is not integration. The assistant remembers what you told it. It still does not see your Tally or QuickBooks ledger, your school ERP or your appointment book unless someone deliberately connects them. It is also not a system of record: you cannot audit it the way you audit a shared drive, and it is not the place for student files, patient identifiers, salary data or bank details, whatever a salesperson implies about security.
Two groups will overstate this. Consultants will fold a memory feature into a pitch for 'agentic transformation' with a six-figure number attached. And anyone selling AI into a regulated setting — a clinic, a school, a payroll bureau — tends to skip past the flip side: once it remembers, it keeps remembering. A wrong price you mentioned in March comes back in September, quietly, inside a quotation nobody re-checked.
What a sensible owner does this month
- Write a one-page house brief — legal name, tax registration, tone, fee or price heads, the things you never say — and give it to the assistant once. That single hour is most of the benefit.
- Give every staff member their own login. Shared accounts were always sloppy; with shared memory they mean one person's context, drafts and errors now belong to everybody.
- Put a short do-not-tell list on the office wall: no patient names, no student records, no salaries, no bank details. Then tell whoever handles your privacy obligations what you have deployed.
- Pick one repeated task, run it for four weeks, and count the minutes saved. If the saving is not obvious by week four, cancel the seats and lose nothing.
- Diarise a quarterly clean-out of what it remembers, and delete anything out of date.
That is the honest size of it: a genuine reduction in a daily annoyance, one new housekeeping duty, and no reason whatsoever to change your headcount.
Also worth knowing
-
Claude went down again, and Anthropic says the outage is fixed
Anthropic confirmed and resolved another outage, picked up by eight outlets. If a paid AI tool now sits inside a daily task — fee reminders, shift rosters, quote drafts — assume a few hours of unavailability a couple of times a year and keep the manual version of that task written down. Any tool you cannot survive one morning without is a tool you adopted too fast.
PCMag Australia ↗ -
Stanford study finds AI is hitting entry-level jobs hardest
The finding points at the bottom rung — the roles where juniors learn the work by doing volume of it. It describes labour markets at large, not your ten-person office, and the write-up is thin on which countries and sectors it actually covers. Still worth reading before you decide not to replace your next junior, because somebody has to be able to check the machine's output in three years.
Hacker News ↗ -
A telco starts giving Copilot away free with a business plan
E& in the UAE is bundling Microsoft Copilot into a business package at no extra cost. Expect the same move from carriers, banks and accounting-software vendors in India and Canada, where AI seats become a retention sweetener. Before you sign a per-seat AI contract this quarter, ask the providers you already pay what they are about to bundle in.
Zawya ↗ -
OpenAI adds an admin plugin for whoever pays the ChatGPT bill
It lets the account owner see workspace usage, manage members and permissions, and adjust limits. Only interesting if you already have staff on ChatGPT at work — but if you have ten or more seats and no idea who uses what, this is the afternoon to find out. Unmanaged seats are how a small experiment quietly becomes a five-figure monthly line item.
OpenAI ↗ -
Bain and Anthropic team up to sell enterprise AI transformation
A large consultancy and a model vendor agreeing to sell together tells you exactly what the next twelve months of sales calls will sound like. Nothing in it changes what a 40-person workshop can do today. When the pitch lands, ask for one business your size that ran the thing, and what it cost after the pilot ended.
PR Newswire ↗ -
OpenAI shuts down a Russian influence operation using its tools
Accounts were banned for promoting a fake Israel-based think tank and a flattering 'sovereignty' index. The useful lesson here is not geopolitics: institutional-looking research is now cheap to manufacture at volume. If an unsolicited white paper, industry report or glowing case study arrives, check that the organisation behind it existed two years ago.
OpenAI ↗ -
OpenAI shows first results from its own inference chip
Jalapeño is OpenAI's custom chip for running models, claimed to be faster and more power-efficient than what it uses now. For you this is a cost story on a two-year delay: cheaper serving eventually means cheaper subscriptions and higher usage limits. Nothing to do about it today.
OpenAI ↗ -
Apple introduces the M6 and M5 Ultra chips
New silicon, heavily discussed by developers, and nothing that changes an AI workflow this quarter. It matters only if you were about to buy machines for a design, video or accounts team — in which case wait for the new models to ship and for last year's to drop in price. Faster hardware has never fixed anyone's process.
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.