Read time: 4 minutes

Leader, welcome back.

This week AI got measurably better at building the next AI. The headline number is eye-catching, but it is not the one leaders should circle.
Here is what happened, and where it quietly moved the bottleneck inside your team.

On 4th June, Anthropic published a report called "When AI builds itself."
The number that stopped people: in May, Claude wrote more than 80% of the code Anthropic shipped into its own systems! Just over a year ago, that figure was close to zero. That is the whole story in one line.
The curve is bending, fast.

Claude Fable 5, which also landed this week, is that capability put in your hands.
It is built for long, autonomous work, priced at twice Opus 4.8, and included on Pro, Max and Team plans through 22 June.

Look past the "it costs double" headline. On long, hard jobs it finishes in far fewer tokens, so the cost per finished piece of work often lands level or lower. Price the task, not the token.

Here is why it matters even if the pace slows tomorrow.
The cost of producing knowledge work is collapsing. Once the machine handles the doing, your constraint is no longer how fast you can produce a forecast, a memo, or a model. It is how fast you can check it, trust it, and decide on it.
A 100 person firm starts to operate like a far larger one, because each person now directs a stack of agents beneath them. The scarce skill just moved from doing the work to judging it, and that is a much harder thing to hire for.

The tell is that Anthropic did not write this as a victory lap. They openly weigh the risk of losing control of systems that build their own successors, and say they would back easing off the frontier, but only if rivals verifiably did the same. They made that case the same fortnight they filed confidentially to go public! The people closest to it are not sure they can keep pace with what they are building, and they put that in writing. That is the real headline this week, not the 80%.

The quieter story: your advisors got Claude before you did.

Over the past few weeks, KPMG embedded Claude into the platform it uses to deliver client work and was named Anthropic's preferred partner for private equity.

PwC stood up a Claude-native finance practice aimed at the Office of the CFO.
Read that plainly.
The tax advisor, the diligence team and the audit partner now read your numbers with the same tool you do, sometimes before the meeting.

For 30 years the finance seat sold information advantage. That moat is being repriced fast. The response is not to defend the function.
It is to be the person who builds the system the numbers flow through, not the one who just reports them.

If your team is still poking at AI rather than running on it, the gap is almost never the tool. It is the setup.

Get the order right and it compounds. Here is ours:

Pick one painful job, not "use AI more."
Most rollouts stall because they aim at everything at once.
Choose a single repetitive, time-sucking workflow, the monthly pack, contract first-reviews, board prep, and make AI do that one thing properly before you widen out.

Give it your context in one place.
AI is generic until it knows your business. Put your real material, the templates, the policies, the past work, the house tone, into one project the model can draw on, and it starts answering like someone who has worked with you for years instead of a stranger off the street.

Match the model to the job.
The smartest model can cost five times the cheapest, yet most teams leave the priciest one running for tasks as small as renaming files. Default to the mid-tier model and send only genuinely hard work up a level. It is the fastest money you will save.

Keep a human on the judgment.
This is the thread running through this whole edition. Let the model do the producing, but you own the deciding: which output to trust, which to bin, what to put your name to. The teams that get burned are the ones who skip that because the draft looked confident.

Set the guardrails on day one.
Agree who can use what, and check your data settings before you scale, not after. Get it signed off with whoever owns risk early, so AI never becomes the thing legal finds out about later.

None of this needs a transformation programme. It needs the right order and someone who has done it before. That is the work we do with organisations, taking them from scattered experiments to a setup that actually runs.

If that is where you are, reply to this email with the one job you would hand over first, and I will tell you how I would approach it.

The IPO window finance was told to avoid is open again.
"Stay private as long as you can" was the advice for a decade. It just expired. Anthropic filed a confidential draft S-1 with the SEC on 1 June, on a revenue run-rate around $47 billion, up from about $10 billion a year ago, at a valuation close to $965 billion. It got there ahead of OpenAI, which is lining up its own.
If your company is private, the listing conversation is coming, and what settles it is not appetite. It is whether your numbers and controls are ready to be read in public.

Run the 60-second privacy check.
Worried your chats are training someone's model?
On ChatGPT it is Settings, then Data Controls, switch off "Improve the model for everyone."
On Claude it is Settings, then Privacy.
On consumer Copilot it is Account, then Privacy.
The part that matters for your team: Team, Enterprise and API tiers generally are not trained on at all, so your real exposure is people using personal accounts for work. Solve that with a policy, not with panic!

Claude now lives inside Copilot. If your team runs on Microsoft 365, you can now pick Claude Opus 4.8 as the model inside Copilot, with admin controls over which models are switched on. Handy if you want the right model for the task without leaving the tools your team already works in.