What's the one job you'd hand to AI tomorrow if you could? The VAT return chase, the receipts pile, the SOP nobody's written. Hit reply and tell us your biggest time-sink — we build these issues around what you're actually stuck on.

Here's the quiet thing about the last two years of AI. Almost everything sold to accountants has been the same shape: a box you type a question into, and it types an answer back. Useful. We've all written a client email that way. But it's still you doing the work — you're just getting a faster second opinion.

Claude Cowork is a different shape. You don't ask it a question. You point it at a folder of client files and give it a job. It reads every document, does the work, and hands you something finished — a summary, a draft, a spreadsheet, a list of what's missing.

That's the shift worth caring about. The chatbot makes you a faster typist. This makes you a reviewer. The work arrives done, and your job becomes checking it — which is exactly the part of the job you're actually qualified for.

It's not magic, and it's not for everything. But the first time you watch it read a year's worth of a client's paperwork in the time it takes to make a coffee, the penny drops. So this week we're walking you through what it is, where it fits, and the one question every sensible accountant asks before they touch it: is this safe on client data?

Let's get into it.

Anthropic continues to expand what Cowork can reach — more file types, more connected tools. Why it matters: the gap between "demo" and "does my actual filing" is closing faster than most firms have noticed.

Early-adopter practices are starting to share what they've automated — not influencers, working accountants. Why it matters: the use cases that survive contact with a real practice are the ones worth copying.

Guidance on using AI responsibly in a firm is becoming part of the mainstream conversation, not a fringe one. Why it matters: "we weren't sure if we were allowed" is running out of road as an excuse — and as a risk.

We'll keep these short. If one matters to you, tell us and we'll dig in.

Claude Cowork is a desktop app — you install it on your computer, the same as you'd install Xero's app or Excel. That single detail is the whole story.

Claude in a browser tab (claude.ai) and ChatGPT both live in a window, walled off from your machine. You copy something in, you copy the answer out. Helpful, but you're still the courier carrying files back and forth.

Cowork sits on your desktop and — with your permission — can read a folder you point it at. So instead of pasting one document, you say: "here's the Anderson Ltd folder, tell me what's in it." It opens every file, reads them, and works.

It's a junior, not a partner. It will occasionally get something wrong with total confidence, so everything it produces is a draft you check. And — this is the big one — which version you use matters enormously for client data. More on that just below, because it's the question we get asked more than any other.

A question we get asked a lot — and the right one to ask before anything else. The short version: it can be, but only if you set it up properly. Two rules.

The free and personal Claude plans are fine for general, anonymised work. But for real client files you want a Team or Enterprise plan, where your data isn't used to train the models and you get the controls a firm needs. Treat the plan you're on as a data-protection decision, because that's what it is. Check the current terms before you load anything sensitive.

There's a browser feature that lets Claude see and act inside your web browser. Powerful — and exactly the sort of broad access you don't want switched on while you're handling client information. Turn it off unless you have a specific, controlled reason to use it.

Once you're set up, point Cowork at a single client folder and try this. It's the gentlest possible first job — nothing changes, it just reads and reports back:

You'll get a map of the client in under a minute: what you hold, what's outstanding, what to chase. Do it once on a folder you know well, so you can sanity-check the answer. That's how trust gets built — one verified result at a time.

Got a question you want answered? Reply and ask. The ones we get asked most become a whole section — this week's came from a reader who was nervous about client data. Next week's could be yours.