Hello fellow keepers of numbers,

No new major models were released this week. * takes deep breath *

Still, some very exciting news in a few areas that are extremely important for accountants. Microsoft deepens its partnership with Anthropic by straight up copy/pasting Claude Cowork into M365. OpenAI launched ChatGPT for Excel, which I demoed in the Put it to Work section. Anthropic also published a labor-market report that’s scary and also complete B.S.?

Finally, I’ve added a new section called Trending News, where I cover a few quick hits because so much is happening each week that I can’t cover it all in detail.

THE LATEST

Microsoft introduced Copilot Cowork, a new “agentic” version of Microsoft 365 Copilot that can plan and execute multi-step work across Outlook, Teams, Excel, and the rest of Microsoft 365.

It uses a context layer called Work IQ to pull signals from emails, meetings, chats, files, and relationships so it can act with similar context to a human working inside Microsoft 365.

Copilot Cowork was built in close collaboration with Anthropic, using the same agentic tech behind Claude Cowork but running inside the Microsoft 365 tenant rather than locally on a device.

According to the blog, tasks can run for minutes or hours while coordinating actions like assembling meeting prep packets, building spreadsheets, drafting documents, and scheduling calendar time from a single request. Users can pause, steer, or approve actions before changes are applied, and all artifacts are saved into OneDrive and SharePoint with normal permissions and sensitivity labels.

Why it’s important for us:

This is likely the most exciting announcement from Microsoft since the AI wave began. And at the risk of turning Microsoft news into Anthropic news, this is also another indicator that Anthropic’s Claude is the industry leader. Microsoft has partnered with Anthropic to use the exact same tools that power Claude Cowork in the M365 stack.

I’m also going to remain cautiously optimistic here because we’ve been burned by Microsoft a few too many times recently. A few things I’ve noticed from some of the demos and info:

Regardless, I’m still really excited for this. I have far more confidence in this tool, specifically because Anthropic is seemingly signing off on it. This should be a massive upgrade in how people use Copilot.

OpenAI introduced ChatGPT for Excel in beta, an Excel add-in that embeds its GPT-5.4 model directly inside workbooks. The tool sits in a panel inside Excel and works against live cells and formulas, so users can describe what they want in plain English and have it build or update models, run scenarios, and generate outputs without leaving the sheet. Calculations still run natively in Excel, preserving cell-level formulas and allowing users to trace assumptions and audit results.

ChatGPT for Excel can also explain formulas in plain language, trace why a particular cell changed, and follow links across tabs to show how assumptions flow through a model. Before it changes anything, it asks for permission and links its explanations to the exact cells it read or edited so users can review or undo changes.

The beta is rolling out to ChatGPT Business, Enterprise, Edu, Teachers, Pro, and Plus users in the U.S., Canada, and Australia, with Google Sheets support planned.

Why it’s important for us:

This has been coming for some time now. Claude in Excel has taken off recently, and it’s really good. ChatGPT for Excel is much needed for OpenAI.

I’ve tested this myself in the Put it to Work section below. Check it out. Overall, not quite as good as Claude in Excel. But still great for those on ChatGPT plans who don’t have Claude.

GPT-5.4 is really smart. And the ChatGPT limits are much friendlier than Claude’s. So there’s still a decent argument for ChatGPT for Excel.

The undo option is an interesting feature. You can undo Claude in Excel edits, but it’s not a native feature for the app. Still, it currently feels like Claude is significantly better at working with M365 files. The experience, formatting, and formulas are all better with Claude at the moment. But this is also new for ChatGPT, so it’ll likely improve quickly.

Anthropic published a new labor-market report introducing “observed exposure,” a metric that measures not just what AI could theoretically do, but what it is actually doing in real workplaces.

Observed exposure measures the share of an occupation’s tasks that are both theoretically automatable and actually appear in work-related Claude usage, with fully automated uses like API pipelines weighted more heavily than cases where AI assists a human. Anthropic finds a large gap between capability and real-world deployment in several areas, including finance.

Why it’s important for us:

If you follow the AI space and have been on any social media app the last few days, you’ve probably seen the graphic above. I have mixed feelings because I think it’s important to understand the gap between capabilities and active deployment. But some of this also seems like complete nonsense.

There’s certainly a gap right now between AI capabilities and how people in accounting and finance use them. Anyone consistently using Claude Code, Claude Cowork, Codex app, or the Agents SDK via API knows how impressive the agents are right now. They can truly handle a significant portion of current workflows within accounting and finance functions and accounting firms.

However… can they handle 90-95% as the Anthropic chart suggests? There’s just no chance. Even if they went entirely based upon whether tasks are ‘theoretically’ possible, which is entirely judgment, I find it impossible to conclude that we’d be around 90% or more. There are so many nuances and complexities to all things accounting and finance.

It’s also unclear what defines a task that falls under ‘theoretical AI coverage.’ If the human has to approve 1-2 steps within the task, is that AI coverage? If the human has to go grab the data and supply it, is that AI coverage? If the human has to train the AI models on specific nuances and continually tweak them, is that AI coverage?

I’m obviously a proponent for these tools and Anthropic, but I don’t think this article does justice to the state of AI. Even if it’s somewhat true, it’s going to be immediately written off by the people who fall within the gap. So I’m not entirely sure what the article’s goal was. I appreciate when AI companies and reputable people try to keep the public apprised on AI risks, but this doesn’t feel like that.

Regardless, I do think there’s some truth to this, even if the exact numbers aren’t entirely accurate. I hope the real takeaway is that there’s so much opportunity for accounting firms and accounting/finance teams to use AI to accelerate their work.

PUT IT TO WORK

I demoed ChatGPT for Excel and provided some thoughts compared to my demo and usage of Claude in Excel.

I asked ChatGPT to read a PDF bank statement and add it to a new tab. I also prompted it to create an executive summary presentation using an income statement. Check out the results.

ChatGPT for Excel Demo - Watch Video

TRENDING NEWS

This is a new section this week because I can’t cover even close to all of the interesting news. I’ll include this in future weeks when necessary, which may turn out to be every week.

Claude now has a 1M token context window for Sonnet 4.6 and Opus 4.6: This is great news for Claude users. It means you can upload more files, instructions, and go back and forth with Claude longer before quality starts degrading.

Perplexity Computer is now available for Pro plan users: Autonomous agent that you can message to complete tasks on your behalf, built with slightly more security than OpenClaw. It’s also available on mobile via the Perplexity iOS app now. I’ll be attempting to demo this in the coming weeks.

Claudenow builds interactive charts and diagrams in chat: The demo visuals look very impressive. This could be useful for anyone who is a visual learner or needs to see a chart without building some complex Artifact.

WEEKLY RANDOM

I wrote a blog post a few weeks ago. If you haven’t checked it out, I’d appreciate it if you gave it a read. I make exactly zero dollars for it, but I did spend quite a bit of time putting thought and effort into the post. Here’s a quick summary of what it’s about:

For decades, corporate America has followed the same blueprint. Start a business, grow, hire, grow more, hire more, repeat. That cycle built empires, the Fortune 500, and the backbone of our economy. AI is starting to change that.

Whether you wholeheartedly believe the preceding paragraph or you’re a skeptic (or somewhere in between), the market is now reacting as if it were true. Large companies are beginning to lay off their workforce and blame AI. Check out this interesting tracker I found recently.

So the question is ‘what changed?’ Because it sure as hell isn’t us going into ChatGPT and asking it to draft an email and then copying and pasting it into Outlook or Gmail.

Something important changed recently. Models got really good at using tools (i.e., browsing the web, reading documents, running scripts) and learned how to run in a loop. These are ‘agentic AI models’. Different than the chatbot.

A chatbot is like asking a coworker a question. You ask a question and (usually) get an answer. An agent is like delegating a task. They go do it, solve problems they encounter along the way, and come back when it’s done. Essentially, models learned how to follow through.

I cover this in much more detail in my blog post called AI Has Changed. Feel free to check it out to learn a bit more about this, plus how people are using these ‘agentic AI models’, and how you can start (or improve) using them.