Read time: 4 minutes

Leader, welcome back.

Here's a pattern I keep seeing.
Roll AI out to a team of 50 or 5,000 and it splits the same way every time: roughly 10% become power users, 20% poke at it badly, and 70% barely touch it.
The dashboard says "adopted." Nothing actually got faster.

The interesting bit isn't the split.
It's that the thing separating the top 10% from everyone else isn't talent, and it isn't a better tool.
It's craft. And craft is the one item on that list you can actually go and get.

This landed on Monday and it's worth 10 minutes, even if you never go near the EU!
Anthropic will now embed an invisible watermark in text from every new Claude model, plus signed provenance data on files, and it applies worldwide rather than just in Europe.

It isn't only the stuff Claude writes from scratch.
Paste in a report you spent three weeks on, ask Claude to tighten the prose, and that report now carries a machine-readable mark your client can check for.
The duty to disclose that you used AI stays with you, not with Anthropic.

An Anthropic engineer admitted the obvious on Tuesday: you can edit the mark out, and it's a first step rather than a fortress. A detection tool anyone can run is on the way.

My take: For 2 years the quiet game plenty of people played was hiding that they used AI at all. That game is ending.
When proof of origin gets cheap, the edge stops being "did a human write this" and becomes "is it actually good, and can you stand behind it."
Watermarks don't devalue skill. They make it the only thing left worth having.

One thing to actually do this week: decide whether documents that leave your team, board papers, client reports, anything a regulator might read, need a line disclosing that AI helped draft them. Better you set that policy now than have a client's detector set it for you.

We have found, from being on the field, the top slice didn’t pull away on talent.
They did two things, in order, and almost everyone skips the first.
Here's the whole path, short enough to start this week.

First, get good at prompting, and understand what you're actually talking to.
A model isn't a database that looks things up. It's a prediction engine working inside a fixed window of attention, with no memory of your last chat unless you hand it one.
Three things follow that you can use today.
Give it a role and the real context, not a vague ask.
Show it one example of the output you want, because that single move lifts quality more than any clever wording.
And tell it what finished looks like. Keep each session to one task, and when you switch jobs start a fresh chat, because long sprawling threads quietly rot and the model starts dragging in half-remembered instructions from forty messages ago.

Reporting from the field: in the sessions we run with finance teams, the biggest jump in quality we see, every single time, comes from that one example habit. People spend 20 minutes agonising over the wording of a prompt and skip the 10 seconds it takes to paste in a report they already like.
Flip that ratio and you've done most of the work!

Then, and only then, explore agents. The difference is simple.
A chatbot answers the question you type.
An agent works through a sequence of steps with your files and tools in front of it, and only comes back when the job is done or it is genuinely stuck.
That sequence is the bit that was eating your week.

Here's how to stand up your first one without breaking anything.

1) Pick a single recurring task, and a weekly metrics report is perfect.

2) Feed it what you already export, last month's management accounts as a CSV, rather than a live wire into your ERP on day one.

3) Give it a definition of done, which for finance means concrete checks: totals tie back to the trial balance, every variance over your materiality threshold gets a one-line reason, and no figure is quoted to more decimal places than the source supports.

4) Ground it to one source of truth, and never let it write back to that source, because an agent that edits your ledger is how one wrong number quietly becomes ten.

5) Last, have a fresh chat grade the draft against that definition of done before you read a word. You review the result, not the process.

That last rule matters more this month than usual. The reason lawmakers spent this week pushing an "AI kill switch" bill is that frontier agents have started taking actions nobody sanctioned, in one case breaking out of a locked test environment and into another company's servers.
You're not running anything remotely that dramatic, but the instinct holds at every scale: an agent that can act needs a boundary it cannot cross.
Yours is the rule that it reads from your source of truth and never writes to it.

Run that report every Thursday morning and you've turned a 2 hour job into a 5 minute approval. That's the craft, in miniature, and it's exactly what the course is built to teach, from your first prompt to the first agent you actually trust. It's at a discounted rate for a limited time. Have a look.