Guide
How AI receipt scanning works, what it gets wrong, and how to check it
What happens between photographing a receipt and getting a categorised expense: OCR versus AI, how Canadian sales tax is read, and where errors come from.
Last reviewed 5 min readRead as Markdown
An AI receipt scanner turns a photo or PDF into structured data. The image is prepared, then a multimodal AI model reads it the way a person would, returning the vendor, date, subtotal, each tax, total, currency and a category in a fixed format. Unlike older template-based OCR, it copes with any receipt layout, language or faded print. It still makes mistakes on damaged or ambiguous receipts, so every result should be reviewable before it becomes a record.
- Input
- Photo, scan or PDF
- Output
- Vendor, date, subtotal, each tax, total, currency, category
- Main difference from OCR
- Reads meaning, not just characters
- Hardest receipts
- Faded thermal paper, handwriting, crumpled photos
- Canadian detail
- GST, HST, PST and QST read as separate lines
- Safeguard
- Human review before a record is saved
What is the difference between OCR and AI receipt scanning?
Classic OCR turns an image into characters. It can tell you a receipt contains "13.00" but not whether that is the tax, the tip or a line item, so older scanners relied on templates for known store layouts and broke on everything else. Receipts are the worst case for that approach: every till prints a different format, and thermal paper fades.
A multimodal AI model reads the image as a whole. It understands that the number after "HST 13%" is a tax amount, that the largest figure at the bottom is probably the total, and that "TPS" and "TVQ" on a Quebec receipt are the GST and QST. It needs no template, reads French and English, and can be told in plain words what to extract.
What happens when you scan a receipt?
- CaptureYou upload a photo, a scan or a PDF. Multi-page PDFs are rendered page by page into images.
- InstructionThe model is given a prompt built from your own setup: the fields you want, your expense categories with their descriptions, and your projects. This is why categories named after T2125 lines produce T2125-shaped data.
- ExtractionThe model returns the vendor, date, subtotal, each tax, total, currency, category and any custom fields as structured data matching a fixed schema, rather than free text that has to be parsed.
- ConversionA foreign-currency total is converted to CAD at the exchange rate for the transaction date, not the day you scanned it.
- ReviewThe result is shown next to the image so you can confirm or correct any field before it is saved as a transaction.
That is how SimpleLoonie works. It uses AI models from more than one provider, falling back to another automatically if one is unavailable. SimpleLoonie is operated from Ontario under Canada's privacy law, PIPEDA, but the shared cloud service runs on Google Cloud in the United States, uploaded files sit in Cloudflare R2, and receipts are read by AI providers in the United States and the European Union. Every provider is named in the privacy policy. If your records cannot live there, we can set up a private instance deployed in the region you specify, against a database and an AI provider key you own.
How does AI read Canadian sales tax from a receipt?
It reads the tax lines the receipt actually prints — GST, HST, PST, QST or RST — and records each as its own amount. That matters because the taxes are claimed differently: GST and HST come back as input tax credits, QST as an input tax refund from Revenu Québec, and provincial sales tax in British Columbia, Saskatchewan and Manitoba generally not at all.
A good scanner records what the receipt says rather than calculating what the tax should have been. If a receipt shows no tax line, the right result is an empty tax field that you can see, not a rate the software assumed.
When does AI receipt scanning get it wrong?
| Situation | Typical error | What to check |
|---|---|---|
| Faded thermal paper | Digits misread or missing | The total against the card statement |
| Handwritten tip or total | Tip left out or read as a line item | Total paid, including tip |
| Several receipts in one photo | Two receipts merged | One receipt per photo |
| Totals printed twice (subtotal, total, amount tendered) | Wrong figure chosen as the total | Total equals subtotal plus taxes |
| Ambiguous date formats | Day and month swapped | Dates in the future or far past |
| Foreign receipts | Currency assumed to be CAD | The currency field |
Accuracy figures quoted by vendors are measured on their own test sets and vary widely with image quality. The useful test is your own: scan twenty of your worst receipts and check every field. The cheapest safeguard is a habit of photographing receipts on the day, flat and in good light, which removes most of the failures above before they start.
Can an AI receipt scanner extract custom information?
Yes, if it lets you write the instruction. Because the model reads meaning rather than fixed positions, you can ask it for fields no template would anticipate: the licence plate on a fuel receipt, the job number on a hardware store invoice, or the attendee names on a restaurant bill. On SimpleLoonie's Business plan you define these fields with your own prompts, and they are extracted alongside the standard ones.
Frequently asked questions
How accurate is AI receipt scanning?
Accuracy depends mostly on the image. Clear, flat photos of printed receipts are read very reliably; faded thermal paper, handwriting and crumpled receipts are where errors concentrate. Treat any vendor's accuracy figure as a claim about their test set and check a sample of your own receipts.
Is AI receipt scanning better than OCR?
For receipts, generally yes. OCR turns images into characters and needs templates to know which number is which. A multimodal AI model understands the layout, reads any format and language, and returns labelled fields such as tax and total directly.
Can AI read French receipts from Quebec?
Yes. Multimodal models read French and English, and recognise TPS and TVQ as the GST and QST. SimpleLoonie records the two taxes as separate amounts because they are claimed on different returns.
Does the CRA accept receipts processed by AI?
The CRA accepts electronic records, including images of paper receipts, if they are legible, complete and can be produced on request. What matters is that you keep the image, not only the extracted numbers, for six years.
Where does an AI receipt scanner send my receipts?
To whichever AI provider runs the model, which is why it is worth asking. SimpleLoonie names every provider in its privacy policy and states where each one processes data. Organisations that need receipts to stay in a particular place can have a private instance using their own AI provider key.
Sources
- Canada Revenue Agency — Keeping records
- Canada Revenue Agency — GST/HST Memorandum 8-4, Documentary requirements for claiming ITCs
This page is general information about Canadian tax administration, not tax advice, and it was last reviewed on . Rules and rates change. For a decision that matters to your business, check the source above or talk to an accountant.