Hello fellow keepers of numbers,

Happy Halloween! Here are a few scaries:

Maybe the last one only scares me.

THE LATEST

Anthropic launched Claude for Excel in beta as a research preview, adding a sidebar to Microsoft Excel where Claude can read, analyze, modify, and create Excel workbooks. The integration provides full transparency about Claude's actions by tracking and explaining changes while letting users navigate directly to referenced cells.

The beta is currently very limited, with Anthropic collecting feedback from 1,000 initial users before a broader rollout. Claude for Excel works alongside existing Microsoft 365 integrations where Claude can already create and edit Excel files, search emails, and access Teams conversations.

Why it’s important for us:

Microsoft just introduced Agent Mode in Excel and Anthropic has already responded with Claude for Excel. That was fast. This is part of another rollout of Claude for financial services, but this is a major announcement.

This is currently a very limited beta, so it’ll likely be some time before this officially rolls out to all users. The beta is specially trained for financial services, so it appears as though there may be a future where you can deploy a Claude agent into Excel that has specific expertise.

If you’ve followed my previous posts, you’ll know that Claude is currently my preferred AI model. It’s really good at coding tasks, which makes it a nice fit to assist Excel users in building all types of schedules, formulas, and deliverables.

There’s an exciting near-term future where we can ask Claude to help us create a file from scratch, quickly do our own review, and then deploy the Claude in Excel tool to make edits. Both working together in the same file.

First version will probably be best suited for formulas, debugging, and Q&A with your file. I doubt it’ll have very strong design and formatting capabilities. It’s also TBD how seamless it’ll be to build an Excel file alongside an agent while you both work in the file.

PwC has abandoned its 2021 pledge to add 100,000 employees by mid-2026. Their 2025 Annual Review reveals a $1.5 billion investment in AI capabilities and quietly excludes their previous hiring targets. Global chair Mohamed Kande said the firm has upskilled over 315,000 employees in AI, which is boosting the productivity of PwC staff.

PwC has been deploying AI tools across operations. As part of the initiatives, PwC has made a major workforce shift. The firm is coming off of recent job cuts earlier in the year where they let go of approximately 1,500 U.S. employees in audit and tax.

Other Big Four firms have also made similar cuts. KPMG recently cut 195 U.S. audit roles and Deloitte made cuts in the UK consulting division this month.

Why it’s important for us:

PwC is pouring money into AI tools and AI agents. At the same time, they’re lowering hiring targets, cutting staff, and refocusing long-term goals for their employees. Some might argue this is in part due to the current economic climate, and maybe we’ll see a bounce back. But on the surface, it appears that AI is already playing a major role in staffing decisions for larger firms.

Even if AI isn’t directly taking jobs right now, the investment in AI is impacting staffing decisions. It’s clear this will be a trend that continues.

The main takeaways for me are:

Microsoft rolled out updates to Copilot focused on personalization and cross-platform integration. The memory feature lets users ask Copilot to remember preferences, project details, and recurring tasks that get recalled in future conversations. Users can view, edit, or delete stored memories at any time. Microsoft is also rolling out the ability to reference past conversations to reduce repeated context.

Connectors link Copilot to OneDrive, Outlook, Gmail, Google Drive, and Google Calendar after explicit user consent. Users can search for documents, emails, and calendar events across multiple accounts using natural language.

Another update includes Copilot Mode in Microsoft Edge. Copilot Mode can understand tabs, summarize information, and perform actions with user permission.

Why it’s important for us:

Microsoft is again late to the game on a lot of these things, but I guess better late than never (looking at you, Apple). Similar to Claude’s recent rollout of memory, the same feature in Copilot is big news for users. It’s becoming slightly easier to get useful information from AI models if you use and “train” it enough.

I’m not sure how connectors to Microsoft’s own products like Outlook and OneDrive took this long, but at least they’re here now. And with some third party connections, which is nice. Notably, the announcement failed to mention SharePoint.

Lastly, if you’re one of the seven people using Microsoft Edge, Copilot Mode sounds useful. It’s Microsoft’s version of the AI browser. Similar security issues should be considered like other AI browsers that launched recently. At least the data remains under your Microsoft umbrella, so maybe there’s some additional protection compared to Perplexity’s Comet or OpenAI’s Atlas. Nonetheless, it’s still premature from a security and usefulness standpoint.

PUT IT TO WORK

In the Loom below, I show you how to use ChatGPT to perform actions on your behalf using n8n.

ChatGPT is great for brainstorming, researching, and chatting. But then I see a lot of people take that information and copy/paste it somewhere else, or use it to write their own email, or go do the task themselves. I want to help you think of ways you can combine the power of AI models with automation software to perform useful actions.

In the example, I provide a use case where you use ChatGPT to call an n8n workflow that performs tax research, drafts an email, and sends a Slack message to your team - all without leaving ChatGPT.

Using ChatGPT to perform actions via n8n - Watch Video

WEEKLY RANDOM

During a recent livestream, OpenAI published their internal roadmap. Interestingly, they outlined their current estimated timeline for automated AI research.

They expect to have an “automated AI research intern” by September 2026, and “automated AI research” by March 2028.

OpenAI is transparently providing their timeline for self-improving AI. I suspect that an “automated AI research intern” means that the AI models perform their own AI research with some human review required. But by March 2028, they expect no human review necessary because AI models will be smart enough to handle the entirety of the research process.

These specific dates don’t have a direct impact on most of us. They’re also likely to change. But self-improving AI is an important inflection point because the rate at which AI models and tools improve will be exponential when an unimaginably intelligent AI is teaching itself.

Maybe OpenAI is bold-faced lying to us to continue receiving billions of dollars of investment. Maybe they’re being incredibly optimistic with their timelines. But it’s still something interesting to monitor. And it’s nice to finally see some transparency from the company.