And here we go again.
Every few months a model gives a life-changing experience.
After spending a great amount of time using GPT-6 Astra for building CFO apps and finance tasks. I’m convinced it feels really close to AGI.
The idea of letting GPT-6 use Claude feels amazing.
This is as cool as it gets.
Hang on tight.
Just in a moment, I’d share with you how I am using GPT-6 for CFO and finance-related tasks.
Every CFO has systems that will never integrate with anything. The old ERP module nobody will touch. The bank portal. The state filing site. The reporting tool your predecessor bought that has no API and never will.
I just built a planning application and deployed it to the cloud. Then I connected Claude to it, loaded the actuals, answered a few questions about assumptions, and had a first version of the budget.
(GPT-6 Astra did everything itself.)
The whole thing took no more than an hour.
You can see it here.
The workflow is simple.
Claude loads and maps the actuals straight from the app. Nothing to export.
Claude asks you for the budget assumptions. Growth, headcount, pricing, whatever the model needs. You answer in plain language.
Claude builds the first version of the budget.
You are not typing numbers into cells.
You are having a conversation about assumptions while the model does the assembly. The mechanical part of planning is the part that eats your quarter, and it is the part that was never the valuable bit.
Read on to see how easy it gets with GPT-6 to build a planning app and get finance work done.
The best thing is now you can build your own MCP server
But let me take this moment and explain to CFOs the concept of an MCP server.
Think about your bank feed.
Your accounting system does not make you export transactions every morning and upload a CSV. It connected once, and the balance is simply there, current, whenever you look at it.
MCP does that for AI. Instead of exporting your actuals and pasting them into a chat window, the AI connects to the system and reads what is sitting there right now.
A file goes stale the moment you export it, and then it scatters into email and Slack and somebody’s laptop. A connection does not. And you decide what it can read and whether it can write anything back.
So in what follows, the app holds the data. The MCP server is the doorway. Claude walks through the doorway and works with numbers that are actually current.
Back to the flow.