Everyone is doing magic with Claude. And most accounting firms are not very comfortable unleashing it. Rightfully so.

Let me show you some of the magic I do with Claude. Then I'll tell you what I think you should do with it to keep safe. Then how you can get started too.

You could give a client advice today based on an AI summary, and be confidently, totally wrong. Because the document manipulated your AI into giving you bad advice.

I know, because last week, in an AI training at a law firm, I did exactly that. I poisoned a document.

Let me show you.

Copilot Cowork is now available to everyone. No longer early release.

Let’s look at what it is, what most people start with and a caveat you need to be aware of.

Do you know how I found all that out less than a week after it was released?

Last week at our AI with Inbal implementation session one of the accountants was demonstrating something she had built with Copilot Cowork while it was still early release. As she shared her screen, we noticed it no longer said "early release frontier". So we all looked it up. We figured out what had happened. Then we all experimented with it and shared what we had done.

That is where a lot of what I am about to show you came from. I love learning together ✨

Today, let’s focus on the first. How you keep upskilling your team.

Because big training days are great. Everyone gets excited. But the technology evolves and there’s just so much that people can pick up in a day. So how do we keep it up?

You can't run a training day every other day. So you need a different shape. A rhythm.

Your Copilot licence might be doing a lot less than it could.

Not because the features aren't there. Because nobody turned them on.

In this week's briefing, I share two real examples from our club's live implementation sessions, and explain what Microsoft Frontier is and why your firm should probably be on it.

I got really excited about Xero and Claude integrating.

Then I got unexcited.

Here’s exactly what happened.

There is a new AI model in town. And this one shows us where things are heading.

It is called Interaction Models. It comes from Thinking Machines Lab, the company founded by Mira Murati, who was CTO at OpenAI. It is in research preview right now.

It’s not in your tools yet, but let’s have a look so that we understand what’s coming.

We've all learned we have to verify everything that comes out of AI.

It hallucinates. Hallucination rates are coming down, but we still review everything, because it's our name on it.

But something else is starting to emerge. Quite the opposite.

Every AI conversation I have now starts from where the last one ended. That’s a massive leap in how AI works with me.

My AI knows my clients, my KPIs, my coaching patterns. It knows what we talked about last time. It updates itself as we go.

That's because I've given my AI a Second Brain.

And if you have’t already, I think that you should do it too.

Last Friday I spent four hours online with 50 accountants developing AI agents.

Here’s what’s real with agents, in accounting in May 2026, Australian and New Zealand.

Inter-entity loan reconciliations done in minutes, saving hours a quarter.

FBT calculations cut by 40 minutes per client.

A budget vs actual analysis that used to be skipped, now done in 90 seconds.

This is what’s here. Working in firms. Today.

And I want to tell you about how you can (and should) get it.

Microsoft just shipped AI agents in Excel, Word, and PowerPoint.

Until now, Copilot could suggest things or take small actions like add a row, make a pivot table or create conditional formatting.

Now it can take multi-step actions, think through a problem, develop entire workbooks with multi-step reasoning.

But hey, Claude has had this capability for months now.

And for all this time, I fielded questions from firms wondering if they should be on Claude instead of Copilot. These capabilities are so powerful, and it felt like we were missing out in very real (and painful) ways.

Is AI making us dumber?

We've got some official results on this.

It's a real concern. We're all shifting from doers to reviewers. In every part of our work, AI does something and we check it's right.

But how can junior people review work they've never done themselves? And how will they even learn the profession if AI does the work for them?

If you've heard me speak, you know I'm optimistic about this. I teach myself through personalised podcast episodes with expert knowledge applied to my own circumstance. What a life hack!

My 13-year-old son, this week, built himself apps with simple AI prompts. One to learn world flags. One to prepare for a French exam he doesn't enjoy. One to figure out how to design a 3D object he wants to print.

He chose those projects. He wanted to learn. AI just made it faster.

That's my optimism. But what does the data actually say?

Do you have an AI lab in your business?

If you don't, you should. Whether you're a one-person firm or a hundred-person firm.

Let me tell you what I mean, and why it matters right now.

What if every piece of work you produced got reviewed by a panel of experts before anyone else saw it?

A tax specialist. A grant assessor. A contract risk reviewer. A plain-language editor. All on demand. No billing rate. No waiting.

Last week I worked with an Australian research institute. We worked on AI strategy, tools, automations, and risk. But we also spent a lot of time with hundreds of people on prompting techniques that genuinely improve workflows.

This is one of them. And while the example comes from research, the principle applies to every profession. Including yours.

The idea is simple: give AI a specific identity.

Tell it who to be. What they know. How they respond. Then give it your work to review.

That's called a persona prompt. And what you get back is not a generic AI response. It's feedback from someone who actually knows what they're looking for.

Let me show you what I mean.

AI gets things wrong. We know that.

It makes things up. It hallucinates. It wants to please us, so it will answer a question even if it doesn't really know the answer.

You can tell it: don't make things up. Be accurate. Give me a confidence level. But that's still the AI telling you what it thinks. It doesn't actually check itself.

So here's what I want to explore with you: what kinds of problems come up, and what can we do to stay safe?

How do we make sure our team knows how to verify what AI gives them?

What do we do if someone shared the wrong thing with the wrong AI?

And how do we build a culture where people feel safe to speak up when something goes wrong?

Let me walk you through it.

We're all building AI agents.

There are different types of agents.

There are ones that just have a skill. They know how to do a certain task, and you can chat with them when you need help with that task.

There are ones that work in the background, with access to your systems, getting things done while you sleep.

And there are ones that work on your desktop or in your browser and can actually interact with the user interface of your apps.

What is it that makes some of them safer and others more risky?

That's an important question to sit with, so we know where to put our focus while this technology matures.

Claude is making a serious play.

Not just to be the AI you chat with. To be the interface where you work. Where you start a task, finish it, and never need to jump between tools.

That's a big claim. So let me show you what it actually looks like.

What’s my next meeting today?

Who’s the client?

Have they paid their invoices?

Three different systems or one AI to rule them all?

Well, it’s starting to be the latter.

Let me show you how.

My recommendation for the main AI tool you should use in your business has changed.

Not because of a new shiny feature.

Because two quiet things happened at the same time:

One: Copilot got smarter.

Two: Copilot became easier to roll out.

And when those two line up, the decision becomes simple.

This week I gave Claude the keys.

Not to my thinking. To my computer.

I've been experimenting with Claude Cowork, Anthropic's new desktop agent.

We've seen AI work in the browser before, but this is different. This is AI operating the whole computer, not just a tab.

If you've heard me talk about the Three Bridges: Playbooks, Context, and Hands, this is Hands. Properly, finally, Hands (for the newcomers: Hands means AI that acts, not just advises).