Monday, 6:12 a.m., a café emails a bakery: “hey can we do 4 dz plain croissant for fri and the almond ones same as last wk. thx!!” That afternoon a PDF order form comes in from a second café, and on Tuesday a phone photo of a handwritten list from a third. Someone types all of it into the order sheet. That retyping is the data entry worth automating.
AI is good at the reading and poor at flagging its own mistakes. So the AI fills the cells, formulas handle prices and totals, and a short list of cells goes to a person. The sheet above is a few lines from that bakery's week. Three cells needed a person, and each shows a different way this goes wrong.
What AI reads well
It reads documents that don't share a layout. An emailed order, a typed PDF and a vendor invoice with the total in a new corner each month can all land in the same columns. It understands how people write: “4 dz” is 48, “fri” is this Friday, “alm” is almond.
Where AI data entry goes wrong
Handwriting and phone photos. A handwritten “½ dz” comes out as 12, and nothing on the cell says it's a guess. Microsoft's own help page for Excel's picture import says to crop the photo, take it head on, and review the results before inserting them. The AI won't reliably tell you which cells it doubts, so decide in advance: handwritten quantities and amounts always go to a person.
Tables and totals. Line items that wrap onto two lines or run across pages get split or merged. Don't let the AI do the arithmetic. Let a formula multiply and add, and compare its answer with the total printed on the document. When the two disagree, as they do in row 4, a person opens the PDF. Usually a line didn't come through; sometimes the form itself is wrong, and then they call.
Anything that isn't on the page. “Same as last wk” refers to an order the AI never saw. A good setup leaves that cell blank. A bad one guesses.
Check what you already have
No AI at all. If the sender's system can export a CSV, or your customers can order through a Google Forms or Microsoft Forms order form that writes straight to a sheet, there's nothing to read.
Excel. For typed PDFs, Data > Get Data > From File > From PDF pulls tables out without AI, where your version of Excel has it. Microsoft warns that wrapped rows may not come out right. Excel for Microsoft 365 also has Data from Picture, which turns a photo of a table into cells and asks you to review them first.
QuickBooks Online. An admin can set up a forwarding address for the receipts and bills you pay (not the orders you get). It pulls the date, amount and vendor, and you review each one in the Receipts tab.
A parsing service. For orders and forms from many senders, Parseur has a template engine for layouts that never change and an AI engine for ones that do, and exports to Excel and Google Sheets. In October 2026 it was free for 20 pages a month, then $49 a month for 100. Docparser's Starter plan was $39 a month for 100 documents of up to five pages each.
Set it up so mistakes get caught
- Write to a review tab, never straight into the books. Nothing posts until a person has looked at the marked cells.
- Flag by rule. Handwriting, blanks, totals that don't match, and any amount over a limit you choose.
- Keep the source next to the row. Checking a cell takes seconds, not a search through the inbox.
Then count. At ten documents a week, a better order form may be the whole fix. At two hundred, price it out. For bills that have to be matched against approved work, the invoice matching guide walks through one example. When the documents have to land in your order system or QuickBooks with your price list and customer rules applied, I map that out in a roadmap: $1,500 · ~2 weeks · credited in full if you continue. Sometimes the plan is just the $49 parser and a review tab.