How to use Claude safely as a CFO
I spoke to a CFO who asked me if Claude is safe for my finance data. My answer surprised her.
I spoke to a CFO who asked me whether Claude was safe for her data.
Right question, wrong place.
Because here’s what actually happens.
Someone exports the P&L into Excel, drops it into Claude, and gets a good answer. Now two files exist. Next month, two more.
Count where they end up.
Slack. Email, forwarded twice. A laptop that goes home every night. A hotel PC during a board trip. WhatsApp, because that’s how your team works when something is urgent.
And at least one pasted into a free personal ChatGPT account, which trains on it.
Nobody made a bad decision.
Everyone was trying to work faster.
But think about what’s in those files. Next quarter’s numbers before they’re announced. The restructuring scenario with names attached. The commentary you wrote for the board and then quietly softened.
You can’t say where any of it sits right now or who has a copy.
So, to answer her directly: yes, Claude is safe.
Team and Enterprise sit outside model training entirely.
The model was never the risk. The files are.
Connect Claude through an MCP server with real authentication, and none of this happens.
Users read only what they’re permitted to read.
Claude writes back into the system, not into a file. Nothing gets exported.
One version, because there’s one source.
The connection is the security control, not the policy asking people to stop emailing spreadsheets.
Most CFOs worry about the wrong layer.
They ask whether the model is safe while their financials sit in 14 spreadsheets and nobody knows which one is current.
This letter is how to configure Claude properly.
Five layers, in order, ending with the one that removes the files entirely.
Let’s dive in.