when you can trust ai with numbers (and when you can't).

elisabeth hitz · july 14, 2026 · 4 min read

ai is reliable with numbers that are structured text, and risky with numbers that live inside a picture. a markdown table, a csv, or cells pasted straight from a spreadsheet, it reads those. a pdf, a screenshot, or a scan, it reconstructs those, and reconstruction is where numbers get misread. the whole trust question comes down to one thing: is the model reading your number, or rebuilding it?

this is not about which model you use or how you phrase the prompt. it is about the shape the number arrived in. get that right and the model is dependable. get it wrong and it is guessing with a straight face.

why does the format decide whether you can trust it?

a number in plain text is unambiguous. the characters "1,240" are exactly that, nothing to interpret. but a number inside a pdf or an image is not text the model can just read, it is something the model has to extract from a layout, matching a figure to a row and a column by inference. that extraction step is where a 3 becomes an 8, or a total attaches to the wrong label. no reading step, no reconstruction, no reconstruction error. that is the entire mechanism.

the reliability ladder

think of your input formats as a ladder. the higher up, the more you can trust the number without a second look. the lower down, the more you must verify.

• pasted text and cells (most reliable). numbers you copy straight out of a spreadsheet or type in keep their structure. nothing to reconstruct.
• csv and markdown tables. still plain text, still structured, rows stay rows and columns stay columns. this is the sweet spot for feeding data you care about.
• clean, digital pdf. workable, but the model is now extracting from a layout. usually fine, sometimes not. verify the load-bearing figures.
• scanned pdf or screenshot (least reliable). the number is a picture of a number. the model is reading pixels and guessing characters. treat every figure as unconfirmed until you check it.

the practical read: the higher you can get your data on this ladder before you ask a question, the less can go wrong. this is the same reason i keep telling people to give claude markdown instead of pdf, you are climbing the ladder before you start.

at the risky end, always verify

sometimes you cannot climb the ladder. the client sent a scanned invoice, the number only exists as a screenshot, that is the reality you have. fine. you can still use it, you just change how much you trust it. at the bottom of the ladder, every figure is a claim until you confirm it against the source.

this is where the fast check earns its keep. spot-check the two or three numbers the decision rests on, ask the model which line each came from, and sanity-check the totals. i broke that whole habit down in the sibling piece on the 10-second check that catches ai mistakes before they ship. the ladder tells you when to worry. the check tells you what to do about it.

the one habit that covers most of it

before you ask an ai anything about a number, ask yourself one question: is this number text, or is it a picture of text? if it is text, you are on solid ground. if it is a picture, you are in reconstruction territory and you verify. that single question, asked before you rely on the answer, prevents most of the number errors people get burned by, and it is a big part of why the ai usually is not the thing that failed, the format you fed it was.

the three questions people ask

can't modern ai just read a pdf now?

it can read one, and it still reconstructs the structure to do it. reading improved, the reconstruction step did not disappear. a clean digital pdf is usually fine, a scan or a screenshot still deserves a check on every load-bearing figure.

is a screenshot of a spreadsheet ever okay?

for a rough look, sure. for anything you will act on or send, no. you already have the spreadsheet, so paste the cells instead of the picture. you are one copy-paste away from the top of the ladder, take it.

does trusting the number depend on how important it is?

yes, and that is the right instinct. a ballpark figure for your own eyes can ride lower on the ladder. a number going into a client invoice or a decision gets climbed as high as possible and verified regardless. match the effort to the stakes.