Scan a receipt

The camera lives on the upload area itself, so there is nothing extra to find: on a phone, tap it and choose "Take a photo"; on a desktop, use the "Scan a receipt" button inside it. It is the fastest way to get one document into your books, and it is the only path that keeps the receipt image on the transaction.

It handles exactly one document at a time, as an image. Everything else goes into the same area as a file.

The scanner uses the same two reading modes as a normal upload. Choose Standard or AI at the top of the Upload page first: until you do, the upload area stays locked and neither dropping a file nor opening the camera does anything.

AI mode also needs the AI-processing permission, just as it does for a dropped file. Whichever mode you choose is the mode that reads the scanned document, so the privacy trade-off is the same one described in "Standard mode vs AI mode".

On a touch device with a camera, tapping the upload area asks what you meant: "Take a photo" or "Choose a file". Tap "Take a photo" and a live viewfinder opens.

Point it at the receipt. The scanner looks for the edges of the document as you move, and an outline snaps onto it once it is found. When the whole receipt is in frame, sharp, and held still for about a second, the scanner takes the photo itself. You do not have to press anything.

A hard receipt — crumpled, low contrast, on a busy surface — may never earn that instant confidence. Hold it steady anyway: after a couple of seconds of the outline sitting still on the page, the scanner takes the photo on its own. If it waits instead of shooting, it tells you why: move closer, fit the whole receipt in frame, or hold still.

The crop is only applied when the scanner genuinely trusts the outline. When it is unsure, it keeps the whole frame instead of guessing — an uncropped photo still reads fine, while a wrong crop could cut off the total.

On a desktop the upload area offers a "Scan a receipt" button, which picks an image file. There is no camera option, because a laptop webcam is a poor way to photograph a receipt.

Pick the image file and the flow continues exactly as it does on a phone: the document is straightened, read, and shown to you for confirmation. "Choose an image" is also available on a phone, for a photo you took earlier.

Both paths straighten the photo before reading it. The perspective and rotation are corrected, the background (the table, the desk) is cropped away, and the lighting is evened out — the grey gradients and soft shadows of a phone photo are removed, so the result looks like a scan rather than a photo.

If there are no edges to find, for example in a screenshot or an already-cropped scan, the whole image is used as it is. Nothing breaks: the document is simply read as it came in.

The straightened image is then read in the mode you chose, and the flow moves to the confirmation screen.

A single receipt becomes a single expense in both modes: the printed grand total, the purchase date, and the merchant from the top of the receipt — never one row per item line. Standard mode reads these deterministically on our own servers and categorizes by merchant; AI mode reads the receipt with the vision model, in any language.

The confirmation screen shows the scanned image on one side and what was read from it on the other: merchant or description, amount, date, and category. All four are editable right there, so a wrong amount or a missing date is a one-line fix instead of a trip to Review.

Anything the reader was not confident about is flagged with a "Worth a check" marker. Look at those fields first.

A scanned document is attached to the transaction it produced. The image appears in the receipt column in Review, and you can open it full size from there.

This is the difference from dropping a receipt photo on the drag-and-drop area, which reads the numbers but does not keep the picture. With that path you had to attach the image by hand afterwards.

The scanner finds the document by its edges, so the edges have to be visible. Almost every disappointing scan comes down to one of the points below.