how to stop ai from making mistakes on your own data.

elisabeth hitz · july 14, 2026 · 5 min read

when ai gets a number wrong on your own spreadsheet or contract, the model is rarely the problem. the problem is how you handed it the data. feed it a mangled format, dump too much at once, or leave the columns unlabeled, and you get confident, wrong answers. fix the input and most of those mistakes disappear before they ever reach a client.

this is the part almost nobody talks about, because it is boring. everyone wants to argue about which model is smartest. but on your real business data, accuracy is an input habit long before it is a model choice. here are the three input mistakes that cause most wrong answers, and the fix for each.

why does ai misread numbers that are right there in the file?

three reasons, and they stack. the format is mangled, so the model reconstructs the numbers instead of reading them. the pile is too big, so the one row you care about sinks into the dead middle. or the data is unlabeled, so the model guesses what a column means. each one is fixable in under a minute.

mistake one: you fed it a mangled format

a pdf looks clean to you because your eyes assemble it. to the model, a pdf is often a scattered set of text fragments with no reliable order. when you paste a pdf table, the model does not see rows and columns, it sees a stream it has to guess the structure of. that is where a total lands next to the wrong label, or a figure from one row attaches to another.

the fix: convert the data to a clean format before you ask anything. a markdown table or a csv keeps rows as rows and columns as columns, so there is nothing to reconstruct. this is the whole argument behind feeding claude markdown instead of pdf, and it is the single highest-leverage habit for accuracy on your own files.

mistake two: you dumped too much at once

the instinct is to give it everything and let it find the answer. but the more you paste, the more likely the important row ends up buried in the middle of the pile, which is exactly where models pay the least attention. i wrote a whole piece on how the middle of a long context vanishes. the short version: the start and the end get weighted, the middle gets skimmed.

the fix: feed only the relevant slice. if the question is about last quarter, paste last quarter, not the whole year. if it is about one client, paste that client's rows, not the entire export. a small, targeted input is not just faster, it is more accurate, because there is no dead zone for your key number to hide in.

mistake three: you gave it unlabeled data

you paste three columns of numbers and ask which product is most profitable. but you never told it which column is revenue, which is cost, and which is units. so the model guesses, and a confident guess reads exactly like a correct answer. this is one of the quiet reasons ai makes things up: not because it is lying, but because it filled a gap you left open.

the fix: label everything. give every column a clear header. say what the currency is. state the time period. one line of context up front removes the guessing entirely, because now the model is reading your data instead of inventing an interpretation of it.

the pattern behind all three

notice the through-line. every fix is about the input, not the prompt. clean format, relevant slice, clear labels. do those three things and you have removed most of the ways a model can go wrong on your numbers before you have typed a single instruction.

• mangled format: convert to markdown or csv, never paste raw pdf tables.
• too much dumped: feed only the slice the question needs.
• unlabeled data: add headers, currency, and time period up front.

none of this is glamorous. it is the plumbing. but it is the difference between an answer you can send to a client and one that quietly embarrasses you a week later. for the other half of the job, catching the mistakes that still slip through, read the sibling piece on the 10-second check that catches ai mistakes before they ship.

the three questions people ask

does a better or newer model fix bad input?

no. a smarter model reconstructs a mangled table more convincingly, which can make the error harder to spot, not easier. clean input helps every model. it is the cheaper and more reliable lever.

is markdown really better than just pasting the spreadsheet?

pasting cells directly from a spreadsheet is usually fine, because that keeps the row and column structure. the danger is the pdf or screenshot in between, where the structure gets lost. plain text that preserves rows and columns is what you want, whether that is pasted cells, csv, or markdown.

how much data is too much?

there is no fixed number. the rule is simpler: if you would not expect a person to find the answer quickly in what you pasted, trim it. give the model the slice a sharp assistant would need, and nothing else.