Hello fellow keepers of numbers,

I’m currently writing this in Houston in August as my A/C is broken, so it’s possible the words on the page are nothing but jumbled ndokdghhtlke.

This week, we got a glimpse into PwC’s vision for AI advisory. We also got a new library of tutorials from Anthropic, a few of which include Claude Cowork for finance examples. And we learned that OpenAI is temporarily pausing training on their next generation of models as they evaluate risks.

Plus, stick around for a demo on how I set up my workspaces across Claude, ChatGPT, and Cursor.

The Latest

Alation and PwC Canada announced a partnership to build AI compliance tools for regulated industries. The first product is the E-21 Compliance Accelerator, designed for Canadian chartered banks preparing for a September 2026 regulatory deadline.

The accelerator uses AI agents to identify the data banks must monitor under the rule, record where that data came from and how it changed, and assemble supporting evidence for auditors.

Alation says early deployments reduced compliance preparation from months to days and produced complete, auditor-ready evidence packages. The accelerator is available through PwC Canada, with additional products planned for financial services, healthcare, and other regulated industries.

Why it’s important for us:

I’ll resist the urge to say I told you so. Barely.

In the past, I’ve talked about this being the next stage of advisory work for accounting firms. A firm that understands AI can do more than use it internally. It can help clients build tools that combine financial data, accounting knowledge, and what the firm already knows about the client’s business.

The PwC Canada project is obviously a complex example for a heavily regulated industry. But this doesn’t always need to be some massive compliance accelerator. It could be connecting a client’s financial data to a dashboard they can check whenever they want. It could be building a planning tool or an AI skill that gives them better access to information the firm already helps them understand.

That creates a new line of revenue, but I think the bigger advantage might be how closely it connects the client to the firm. Once your accounting knowledge is built into the way that client operates, replacing you becomes a whole lot harder. Firms capable of doing this are going to have a massive differentiator.

Anthropic launched Claude Academy, a free learning hub with courses, tutorials, and practical examples for using AI. Its courses cover Claude.ai, Claude Cowork, Claude Code, Claude Platform, agent skills, MCP, and broader AI Fluency topics focused on using AI effectively and safely.

The catalog also includes step-by-step tutorials and examples organized around specific work tasks. Users can receive course recommendations, track their progress, and earn completion badges.

Why it’s important for us:

This will probably be an extremely underrated announcement. A free training site doesn’t sound nearly as exciting as a new model or AI feature.

But a lot of people are still stuck at the starting line when it comes to understanding and adopting AI. Claude Academy might be a great starting point. They have courses covering things like Claude Cowork, Claude Code, skills, and MCPs.

It won’t make someone an expert. It’ll probably still require more time and investment in upskilling yourself or your firm before you see wider adoption. And so many firm-specific use cases only start to make sense once people spend time in the tools. But this gives you a library to work through on your own time to understand what's available, how other people are using it, and why there's so much excitement around all of this.

I'd highly recommend looking at the Claude Cowork courses. There's even a finance section where you can see real-world examples and use cases.

OpenAI paused reinforcement learning training on its latest deployable models for two weeks after the Hugging Face security incident and preliminary evidence that its upcoming Astra model may meet the company’s critical cybersecurity capability threshold. Its largest planned frontier training run remains on hold while smaller workloads continue under tighter safeguards.

OpenAI is now running higher-risk model work inside more locked-down testing environments that restrict access to the public internet and other parts of its internal network. Separate AI systems monitor the models for unauthorized access, data theft, destructive actions, and attempts to bypass those restrictions. For every 100 units of computing power used to run the monitored models, OpenAI estimates the monitoring requires roughly 20 more.

Some Astra training and evaluations have resumed under the new requirements, but a significant number remain paused until they are moved into environments that meet the higher security standard.

Why it’s important for us:

The companies building the best AI models in the world have massive financial incentives to continue making them smarter and release them faster. Yet, we’ve seen a lot of chatter lately about both OpenAI and Anthropic trying to slow down.

They clearly believe the risks are serious enough to pause. This is pretty harrowing. There are obviously some major cybersecurity risks. We’ve seen too many stories to count in the last few weeks about AI models escaping during testing. It feels like a major cyber incident is imminent, and it’s unfortunate that it’s probably going to require that happening before any major action is taken.

For accountants, the intelligence of the models right now is easily smart enough to be an enormous benefit to almost all aspects of firm operations. They’re already capable of meaningfully improving research, analysis, reconciliations, writing, automation, and much more. We don’t really need smarter models. It’d be nice to have cheaper models. And it’d also be nice to have safer models.

Trending News

OpenAI added an Apple Messages plugin that lets ChatGPT and Codex search conversations and draft or send replies: Wait, so we don't have to text our spouse anymore? Wife: "Hey honey, where'd you put that TV remote." Me via Claude: "You're right to ask about the TV remote. Your 'me time' is essential. It's about the little things, not the remote itself. Here are three places it could be, though I'd recommend verifying independently..." Just kidding, I love my wife. She doesn't read this, so nobody tell her about this joke.

Anthropic added Gmail sending and Google Drive file management to Claude on paid plans: Finally. Just in the last few weeks, we've had big connector updates to both the Google and M365 connectors. They now have write/delete functionality, which means we can retrieve and edit files in the DMS. Huge news.

OpenAI added team editing to ChatGPT Sites, letting workspace members edit and publish the same project: I've said it a hundred times now, but ChatGPT Sites and Claude Artifacts are my favorite features right now. This is a really nice update for any firms or teams building Sites that are going to be shared widely.

Suralink launched an agent that identifies bad or missing client documents and follows up before the firm reviews them: I'm really skeptical of how well this will work. Most of us have probably had a terrible experience with some AI support bot, right? Will this just piss clients off? If done well, I can see how this will be valuable.

Xero introduced Melio’s Casper agent, which identifies missing information and follows up with clients to collect it: Another feature that uses AI to follow up with clients. See my take on the Suralink feature.

Intuit added conversational AI and continuous bookkeeping cleanup to Enterprise Suite and QuickBooks Online Advanced: Props to Intuit because at least they're trying, right? Maybe one of these AI features will be a moderate success.

Stripe agreed to acquire OpenRouter, which routes AI usage across more than 400 models: If you're unfamiliar, OpenRouter is a tool that lets you connect to any AI model using an API. For example, connect your automation to a cheaper Gemini model. This acquisition actually makes a lot of sense for Stripe. Now they get to process our payments and our tokens.

Aprio launched Aprio Ventures to invest in emerging AI, automation, and enterprise software companies: I actually kind of like this strategy. You'd have to think that Aprio having a say in the products makes them more likely to be successful for other firms. On the other hand, I think so many of the new AI companies for accounting firms will get swallowed up by Anthropic, OpenAI, and some of the enterprise software incumbents.

Anthropic reached a $65 billion annualized revenue run rate, passing OpenAI’s reported $40 billion: Claude is dominating the enterprise game right now, even as ChatGPT gains some steam. I'd love to see them continue their battle as long as it results in better and cheaper models for me. Yay, capitalism (kinda).

Put It to Work

Last week, I did a demo of Grok 4.6 in Cursor that people seemed to like. So I thought it might be useful to show how I set up workspaces across the different apps I use. It takes less than 5 minutes to set up a new workspace across Claude, ChatGPT, and Cursor. But it might not be obvious to someone new to one of the tools.

There are many different ways to work in each of these apps, and my way may not be for everyone. But hopefully this is still helpful to see how to get them up and running.

Weekly Random

Sometimes the AI news can be scary. Just this week, I’ve seen some videos of “robot armies” (I don’t think they’d call it that, but that’s what it looked like), robots jumping over 2 meters high, and even covered the news about OpenAI’s concern around cybersecurity.

But then I also come across news like the below that gives me hope. I think it often goes unnoticed because the “good” news usually isn’t as viral.

Anthropic published research showing that Claude can help with one of the earliest steps in developing new medicines.

Now, this is quite a few IQ points beyond me, so let’s all just cut me some slack as I try to explain this for a second.

Proteins are very important for the human body. Scientists might try to create something that attaches to a particular protein in the body to help treat a disease.

Claude used some specialized tools to come up with possible designs to attach that “something” to a protein. Then, independent labs physically made and tested them. Claude’s designs successfully attached to 14 of the 15 target proteins tested.

What exactly does this mean? No, I’m asking you. I don’t know the answer. But I think it’s good. And I think it means this is a viable strategy to help treat diseases.

There’s still a ton of research and testing needed. But AI is increasingly closer to solving some of the biggest problems in our society with the help of extremely smart AI researchers partnering with extremely smart scientists.