A fractional CFO uploaded a real three-statement financial model into Claude and got back a ten-slide board presentation. He never opened PowerPoint.
That demo, posted by @YourCFOGuy on X, is the version of the AI productivity story that actually lands. Not a chatbot answering finance trivia. A working software experience that reads your spreadsheet, builds the analysis, and renders the deck.
What actually shipped
Anthropic released Claude as a native add-in for Excel and PowerPoint, with PowerPoint launched as research preview in February 2026 and the March 2026 update introducing shared context across the two apps, reusable Skills, and MCP connectors for major financial-data providers. As of April 13, Word also has a native add-in. Claude is now sitting inside the trio of applications that drive most knowledge work in the Western economy.
It is not a chatbot sidebar. It reads the spreadsheet's existing structure, including formulas and dependencies, and edits in place. Anthropic recommends Opus for complex financial work. You can connect a PowerPoint file to an Excel file inside the plugin and edit both simultaneously, with the conversation history persisting between them. No copy and paste. No re-uploading.
What "Skills" do
Skills are the most quietly important part of the rollout. They are reusable instruction sets stored as markdown files. You point Claude at your scenario-analysis template, your client report format, your brand style guide, your model roll-forward conventions. The Skill captures the pattern. Every subsequent run uses it. Anthropic demoed a Scenario Analysis Skill that handles beat-miss analysis, model roll-forward, and research-note generation in three steps from a single prompt.
The MCP connectors matter for the same reason. The integrations cover S&P Global, LSEG, Daloopa, PitchBook, Moody's, and FactSet. The "garbage in, garbage out" floor on enterprise AI workflows has always been data access. If your AI cannot read your authoritative source of fundamentals, it cannot replace the analyst checking those fundamentals. With these connectors live, that floor moves.
The distribution play
The strategic shape here is the part worth sitting with. Anthropic spent the last three years selling itself as the safety-conscious frontier-model lab whose go-to-market was the API. Developers build on top, end users get Claude through whatever app the developer ships. That is a slow, indirect distribution channel.
Microsoft 365 has more than 400 million paid commercial seats. Anthropic just walked into Microsoft Copilot's home turf, on Microsoft's own surface, with a plugin that does the work Copilot is supposed to do. Microsoft's edge is integration depth across the broader Microsoft ecosystem (Teams, Outlook, OneDrive, the org graph). Claude's edge is the thing it has always been good at: large-context reasoning over long, structured documents, with cleaner financial-modeling behavior in operator testing.
This is not a courtroom-style winner-take-all fight. It is a hand-to-hand product fight that will be settled over time by which tool finance professionals open in the morning. But for the first time, Claude's distribution layer is the same as Copilot's. Customers do not have to choose a different app. They choose a different add-in inside the app they already pay for.
The thing not enough people are saying
A reasonable productivity argument here is that Claude lets a fractional CFO do in two hours what used to take twelve. That is real. The harder argument is that the fractional CFO's value proposition has always rested on two things: judgment about what to put in the model, and craftwork in turning the model into a deck a board will accept. The first is durable. The second was billable hours.
Andrew Yeung's much-shared observation from a flight to Austin is that economy-class passengers were working on PowerPoint, Excel, and Outlook, while business-class passengers were using Claude Code and Replit. The class divide on that flight is not really about productivity tools. It is about which work has already been automated and which work has not. PowerPoint and Excel just got automated. The next class of passengers up is also vulnerable.
The question is not whether AI does the work. It does. The question is what the work is, once the artifact stops being the bottleneck.
If AI can build your CFO presentation in two minutes, what exactly are you being paid to do?
Originally published as an Instagram carousel on @recul.ai.