Split a Bill by Taking a Photo (Receipt OCR Guide)

To split a bill by taking a photo, open the app, snap a clear photo of the receipt, and let the AI read the line items, prices, tax, and tip into an editable draft. You then review what it read, correct anything it misheard, assign items to people, and split. The important distinction: the AI only reads the receipt, it never does the math. Once the numbers are on screen, all the arithmetic (per-person totals, proportional tax, tip, rounding) runs in verified Python code, not in the language model. That split of labour is the whole point: OCR is fast but fuzzy, so a human confirms the numbers, and deterministic code does the sums so the totals are always exact and reproducible.

The snap-review-split flow, step by step

The photo-split flow is three moves: snap, review, split.

Snap. Open a new bill, tap the camera, and photograph the receipt. Get the whole thing in frame: top (merchant + date) to bottom (subtotal, tax, tip, total). A flat, well-lit shot on a dark surface reads far better than a crumpled one under a dim table lamp.

Review. The AI returns a draft: each line item with its price, plus subtotal, tax, and any service charge or tip it found. This is the step people skip and shouldn't. Scan the draft against the paper. Did it read $18.00 as $78.00? Did it merge two lines? Fix it inline before you split: editing a draft takes five seconds, unpicking a wrong split later takes five minutes.

Split. Choose how to divide it. Equal split for a shared pizza night; by-items when everyone had different dishes; by-shares when two people split one plate. Assign items to people by tapping, and the per-person totals appear instantly, with tax and tip spread proportionally across what each person actually ordered.

Start to finish, a clean receipt is under a minute. The review step is the tax you pay for not typing twelve line items by hand.

Why the AI reads but never does the math

This is the core design rule, and it is worth understanding because it is why the totals are trustworthy.

Language models are excellent at reading messy text: a faded thermal receipt, a handwritten tip, a foreign-language menu. They are genuinely bad at arithmetic. Ask a model to add fifteen prices and apply 8.875% tax proportionally and it will sometimes be off by a few cents, or a few dollars, in ways you cannot predict.

So the app draws a hard line. The AI's only job is understanding: turn the pixels of a receipt into structured numbers (this item costs this much, tax is this, tip is this). The moment those numbers exist, they are handed to verified calculation code that does every sum. Per-person totals, proportional tax allocation, tip distribution, rounding to the cent so the parts always add back to the whole: all of it is deterministic code, run the same way every time.

The practical payoff: the same receipt always produces the same split. There is no run-to-run drift, no hallucinated total, no 'the AI was having a bad day' rounding error. If you and a friend split the identical bill, you get identical numbers. The AI got you the numbers fast; the code makes sure they are right.

What the OCR actually captures

From a receipt photo, the draft typically pulls:

Line items and prices: each dish or product with its amount, so you can assign per item rather than dividing the lump sum.

Subtotal, tax, and service charge: read as separate figures so tax can be spread proportionally instead of split flat. If two people split a $120 dinner and one ordered $80 of it, they should carry two-thirds of the tax, not half. Proportional allocation handles that automatically.

Tip / gratuity: whether it is a printed suggested amount you wrote in, or an auto-added service charge on a large table.

Merchant and date: used to name the bill and set its date, so your history and reports stay tidy without extra typing.

What OCR is good at: printed thermal receipts, clear grocery tapes, standard restaurant checks. What it struggles with: heavy creases through the numbers, glare washing out a column, handwriting, and receipts photographed at a steep angle. When in doubt, retake the photo flat: a second snap is cheaper than correcting six line items.

Review is not optional: here's what to check

The review step exists precisely because OCR is fuzzy. Spend ten seconds here and the split is bulletproof.

Check the total first. If the draft's total matches the printed total on the receipt, the line items are almost certainly right and you can move on. If it is off, the mismatch tells you where to look.

Watch for digit misreads. The classic ones: a 1 read as a 7, a 3 as an 8, a decimal point in the wrong place turning $4.50 into $45.00. These jump out when the total does not match.

Watch for merged or dropped lines. Two items printed close together can read as one; a faint line can be missed entirely. Count the items against the paper.

Confirm tax and tip landed in the right fields. Occasionally a service charge gets read as a line item. Move it to the tax/tip field so it spreads proportionally rather than being assigned to one person.

Everything is editable inline: tap a number, correct it, done. You are not re-entering the receipt; you are proofreading it. That is the deal photo-splitting offers: the machine does the tedious transcription, you do the two-second sanity check.

Photographing a receipt for a clean read

A better photo means fewer corrections. A few habits that consistently help:

Flatten it. Smooth the receipt against a table before shooting. Creases that run through the price column are the number-one cause of misreads.

Contrast the background. A pale receipt on a dark table gives the edges something to sit against. A white receipt on a white tablecloth is the hard case.

Light it evenly. Avoid a single harsh light that blows out one side and shadows the other. Diffuse indoor light or daylight is ideal; kill the glare from an overhead spot.

Get the whole thing in frame, straight-on. Merchant line at the top, total at the bottom, shot from directly above rather than at an angle. A skewed photo distorts the columns.

Long receipts: if a grocery tape runs off the frame, get as much as you can and correct any missed lines in review. A cleaner single shot beats a blurry attempt to cram it all in.

None of this is fussy: it is the difference between a draft you glance at and approve versus one you have to fix line by line.

When you can't take a photo: talk or type instead

The photo is one input, not the only one. The same understand-then-calculate engine also takes voice and plain text, which matters when there is no receipt to snap.

Voice. Record something like 'dinner was 84 dollars, I paid, split between me, Sam, and Priya' and the AI transcribes it into a draft bill you review the same way. Useful when you are walking out of the restaurant and the receipt is already in the bin.

Chat / text. Type the same thing in plain language. Handy for a bill with no paper trail: a cash split, a reimbursement, 'Alex owes me 30 for the concert ticket.'

In every case the pattern is identical: the AI turns your input into structured numbers, you confirm them, and verified code does the math. Photo is the fastest path for an itemised restaurant or grocery bill; voice and text cover everything else. And because it all lands in the same bill and the same running balances, you can mix them freely: snap the dinner receipt tonight, type the taxi split tomorrow.

After the split: balances, settle-up, and history

A photo-split bill is not a dead end: it flows into the rest of the app the same as any manually entered bill.

Running balances update immediately. Once you split, each person's share posts to the group balance. Nobody has to re-enter anything; the receipt you photographed is now part of who-owes-whom.

Settle up in a tap. When it is time to square up, per-share settle-up links let people pay their exact amount, no 'wait, how much did I owe again?'

Everything is exportable. The parsed bill, its line items, and its split live in your history and can be exported to CSV for records, reimbursements, or expense reports. The photo did the data entry; the export gives you the paper trail back, structured.

Multi-currency travels too. Photograph a receipt abroad and the amount is captured in its own currency with the exchange rate locked at the time of the bill, so a later rate swing never quietly rewrites what someone owes.

The photo is the fast on-ramp. What makes it worth doing is that the numbers it captures behave exactly like every other number in the app: verified, balanced, and permanent.