Today's briefing
Who Will Actually Install Your AI — And What They'll Charge
Anthropic is spending $100M to train 10,000 people to deploy AI, starting at McKinsey and Deloitte — here's what that means when the proposal reaches you.
The one that matters
Anthropic will spend $100M training 10,000 people to install AI inside companies
What happened
Anthropic, the company behind Claude, has committed $100 million to training roughly 10,000 engineers to deploy AI inside working organisations. The first cohorts come from McKinsey and Deloitte. These are not research scientists or model builders — they are the people whose job is to take an AI system and connect it to an organisation that already has software, staff, rules and bad habits. Split evenly, that is about $10,000 of training per person, which tells you how much skill the company thinks is currently missing.
What it actually means for your business
The useful signal is not the money. It is the admission underneath it: the hard part is no longer the model, it is the installation. Everyone selling AI has spent two years demonstrating capability. This is the first large bet that capability is not the bottleneck — plumbing is.
Picture a 600-student school. The model can already draft fee-reminder messages in two languages, summarise a long parent complaint, and sort applications by missing documents. What nobody has done is connect it to the school's admissions software, decide which of forty reminders a human reads before it sends, and own the consequences when a sibling discount is applied wrongly. That connecting work is one competent person for a fortnight, and it is the part that actually costs money. The same is true of a 30-bed clinic trying to pre-fill insurance pre-authorisation forms, or a workshop with 40 staff trying to get purchase orders out of email and into its books.
The second signal is who gets trained first. Starting with two of the largest consultancies means the route into mid-size business runs through firms that bill by the day. Expect, within the next year, a polished "AI transformation" proposal to land on your desk carrying a newly badged consultant and a six-figure number.
What it does not mean
- It is not a new product. Nothing you can buy this week became better or cheaper because of this announcement.
- Certification is not an outcome. Someone who has completed a deployment course knows the tooling. They do not know your purchase cycle, your tax filing, or why your senior technician refuses to use the tablet.
- 10,000 people is a small number. Spread across global enterprise clients, this is a workforce aimed squarely at the top of the market. If you employ 40 people, nobody is coming for you — which is fine, because you do not need them.
The overstatement to watch for is the implication that because large firms are staffing up, you are falling behind. A business of this size usually gets its entire return from two or three narrow, text-shaped jobs. That is a week of work with a clear owner, not a programme.
What a sensible owner does this month
- Write down the three tasks that eat the most staff hours and are mostly text: quotations, fee reminders, pre-authorisation forms, purchase-order entry, first-line WhatsApp replies. Put hours per week next to each.
- Pick the one with the lowest cost of being wrong and run it by hand for two weeks with an off-the-shelf assistant — typically $20–30 (roughly ₹1,800–2,500) per user per month — with a staff member approving every output before it leaves the building.
- Name an internal owner, and make it someone who already knows the system the work lives in: the office manager who runs the admissions software, not a new hire with AI in their title.
- If a consulting pitch arrives, ask three things — who maintains this when it breaks, what does month thirteen cost, and show me this running at a business my size. Then ask for a fixed-scope pilot with a spending ceiling instead of a programme.
- If none of your three tasks survives that test, do nothing this month. That is a legitimate answer and it costs you nothing.
Also worth knowing
-
Microsoft adds controls over which AI agents can run in your Copilot account
Microsoft has added governance and usage controls for agents inside Copilot, meaning an administrator can now limit who creates agents and what they are allowed to touch. If your business runs on Microsoft 365, this is the rare item on today's list that is free and worth an hour: ask whoever holds your admin login to review those settings before someone in accounts wires an agent into your invoices unsupervised.
MSSP Alert ↗ -
An AI agent reported the job finished; the database showed otherwise
A write-up of a familiar failure: an agent declared a task complete while the underlying records said it had not happened. The detail is thin, but the discipline it implies is not — never accept an agent's own summary as proof of work. If you automate attendance, stock or invoice entry, your acceptance test is the register or the ledger, not the chat transcript.
Hugging Face ↗ -
Claude can now operate Android apps — one reviewer put it against Gemini
Anthropic says Claude can now work inside Android apps, and a reviewer compared it with Gemini on the same tasks. This matters more in India than in Canada, because a great deal of small-business operations genuinely live on a phone — payments, field staff, supplier chats. Treat it as early: worth ten minutes of curiosity on a spare handset, not worth giving app-level control over your banking or payments.
MakeUseOf ↗ -
AI agents are moving into text messaging, where your customers already are
A roundup of assistants that now live inside text messages, from general-purpose helpers to ones built for families, travel and work. Messaging is where customer contact already happens for most shops, clinics and agencies, so this is the channel where agents will reach your business first. Reasonable to trial for appointment reminders and order status; not for pricing, medical questions or anything a human would have to apologise for later.
TechCrunch AI ↗ -
A US sentence was quashed after an AI-generated video of the victim was shown in court
A court overturned a killer's sentence because an AI-generated video of the victim had been played during proceedings. The signal for ordinary businesses: synthetic media is starting to be treated as something that taints a record rather than illustrates it. If you ever produce AI-made images or reconstructions for an insurance claim, a staff dispute or a legal matter, keep the originals and label anything generated.
Hacker News ↗ -
A Linux distribution has banned AI-generated code from much of its codebase
The makers of Pop!_OS barred AI-generated code from large parts of their project, citing maintainability and provenance. You do not need a position on the policy, but if you commission custom software you now have a fair question for your developer: how much of this was model-generated, and who reviewed it line by line. The liability for code you cannot maintain lands on you, not on the tool.
Hacker News ↗ -
Barclays expands its use of Claude across engineering, markets and user support
A large bank has widened an AI deployment into customer support and trading-adjacent teams. It is useful evidence if a cautious board or family partner is asking whether this technology clears serious procurement — it does. It is not a template: a bank has a compliance department, an audit trail and a team to switch things off, and your 40-person operation has none of those, so copy the caution rather than the scope.
Crowdfund Insider ↗ -
The OpenAI staffer who wrote its model safety reports has resigned and gone public
David Robinson, who prepared the safety reports accompanying OpenAI's major model releases, has resigned and published an editorial criticising the company's internal culture. It changes nothing about what the tools do for you tomorrow, but it is a reminder that a vendor's safety documentation is marketing, not an audit. Keep your own human check on anything that faces a customer or moves money.
The Verge AI ↗
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.