One thing I've been experimenting with recently is how much context an AI model actually needs before it can work effectively with a complex financial model.

My initial instinct was the same as most users: upload the workbook, explain what I wanted, and let Claude (or whatever AI model I’m using) figure out the rest.

The results were reasonable, but they weren't particularly reliable. The model could usually reach an answer, yet I had very little confidence that it had followed the same chain of reasoning I would have used myself. In other words, while we could get from A to B using Claude, the exact logic was unknown. This could lead to inaccurate assumptions and ultimately, unexplained results.

That pushed me to rethink the workflow.

Rather than spending more time refining prompts or choosing the “right” AI model, I've found it much more useful to document the Excel build before introducing AI into the process. Most of that work happens in a Google Doc, not in Claude.

I start by listing the key inputs, but I also try to explain what each one represents from a business perspective. A retention assumption, an enrollment rate, or a growth estimate carries context that doesn't exist in an Excel cell.

From there, I document assumptions, hardcoded values, manual overrides, and any edge cases that could change the output. Finally, I map how the model actually works. Which tabs feed into one another? Which calculations matter? Where are the decision points?

I've started thinking about this as creating an operating manual rather than writing a prompt.

The more I work with large financial models, the more I suspect that the limiting factor isn't the model's intelligence. It's the quality of the context we provide. Dense workbooks are full of hidden assumptions, historical decisions, and business logic that are obvious to the person who built them but invisible to everyone else, including the AI model.

I'm still experimenting with this approach, but so far it has produced outputs that are easier to trust because the model is reasoning within a framework instead of trying to infer one from hundreds of formulas.