Last week we talked about the arc almost everyone goes through with AI, and the 70% rule. This week is the most common concrete version of that problem: you have an Excel file, you drop it into Claude, and you type something like “go fix all the formulas.” Sometimes it works. Often it does not, and it does not tell you which one you got.
What is actually happening when you upload a workbook
The model is not looking at your spreadsheet the way you look at it. You see a clean grid of tables and formulas. Underneath, a workbook is a large structured document with its own markup, and the model has to run scripts to read that document and more scripts to write a new one back out.
If the file is small, that round trip is fine. Two plus two equals four, one sheet, a handful of formulas, no problem. But as you add sheets, formulas, named ranges and cross references, that underlying markup grows fast. What looks like a modest file to you is an enormous amount of text to the model.
Which runs you straight into the context problem from last week. The workbook eats the budget. Things fall off the back. And because the model does not announce that it has lost the second half of sheet three, you get an answer that looks complete.
Excel is a database wearing a friendly coat
Databases exist for a reason. They are strict about what goes where, they do not let you type a note into a column that is supposed to hold a number, and they are genuinely better at this job. They are also not very pleasant to use, which is why almost nobody runs their small business out of one.
Excel is the friendly version. It gives you a database you can actually see and touch, and that tradeoff is why it has run small business operations for thirty years. But the friendliness is a layer over the top, and when you hand the file to an AI you are handing over the whole layer, not the data you care about.
Where this gets expensive
Now combine the two ideas. The file is too big for the model to hold reliably. And the answer comes back at 70%, in a confident voice, with no marker for which part is the bad 30%.
With prose, 70% is a usable draft. With arithmetic, 70% is a wrong number that looks exactly like a right number. If you are pulling that number into a decision about pricing, or headcount, or whether a product line is carrying its weight, you can run six months on it before anything surfaces.
And you will not be able to hand the consequences back. The terms of service on every one of these tools say plainly that you should not rely on the output for business or legal decisions. They have good lawyers. That protection is real and it points in one direction.
A wrong number in a confident voice is more expensive than no number at all.
The structure that actually works
The fix is not a better prompt. It is to stop asking the model to be the calculator, and start asking it to build the calculator.
1. Let it read the sheet once, to extract the logic
Point it at the workbook and ask what the formulas are doing, in plain language. Not to compute anything. Just to describe the rules. This is a reading task, which is the kind of thing it is good at.
2. Turn the rules into scripts
Take the logic and express it as a small set of scripts. A formula that says two plus two equals four becomes something closer to x plus two equals y. Everybody did this in school. The point of the variable is that the rule stops being tangled up with any particular number.
3. Feed it variables, not files
Once the rules live in scripts, the only thing that moves through the model is the variables. That is a tiny amount of context instead of an entire workbook, which means the whole class of problems we started with simply stops applying.
4. Verify against real data before you trust it
This is the step people skip. Take two or three cases where you already know the correct answer, cold, from the real world. Run them through. If the structure reproduces those exactly, you have a tool. If it does not, you have found the bug now instead of in November.
Do not skip this because the first output looked right. Looking right is the failure mode.
What you end up with
Something better than a fixed spreadsheet. You end up with a small, boring, repeatable thing that produces the same answer every time you run it, and you can hand it to someone else without also handing over a tribal explanation of which cells not to touch.
- Stop uploading large workbooks and asking for a fix.
- Extract the logic once, in plain language.
- Express it as scripts with variables.
- Verify against cases where you already know the answer.
- Then run it, and keep spot checking.
You are not working with a person. You are working with something that is a genius and an idiot at once, that does not know your business, and that is extremely effective when you point it at the right half of the job.
Next week: how to actually roll this out to a team without the whole thing stalling.
