How AI Reads a Receipt or a Lease — and Why You Should Still Check It

You take a photo of a receipt, and a few seconds later the merchant, date and total appear in neat fields. Or you upload a lease and the tenant name, rent and start date fill themselves in. It can feel like magic, but it is not. Understanding roughly how it works helps you know when to trust it and when to take a second look.

Step one: the photo or the PDF

Everything starts with the document itself. A receipt might come from your phone camera, your photo gallery or a PDF emailed by the merchant. A lease is usually a PDF, sometimes a scan of a printed and signed copy.

The quality of this input matters more than anything else. A sharp, well-lit, flat photo gives the AI a clean starting point. A crumpled receipt photographed at an angle in a dim car gives it far less to work with.

PDFs that were created digitally often contain real text the software can read directly. Scanned PDFs and photos are just images, so the text has to be recognized from the pixels first.

Step two: the model reads and extracts fields

Modern document AI does not just turn the image into a wall of text. It looks at the words, their positions on the page and the relationships between them, then tries to fill in specific fields.

For a receipt, that usually means:

• Merchant name
• Date of purchase
• Subtotal, tax and total
• Currency
• Card type and the last four digits of the card
• Individual line items, where they are legible

For a lease, the fields are different:

• Tenant name or names
• Monthly rent
• Start and end dates
• Security deposit

The model makes these decisions based on patterns. It has learned that the largest number near the word “Total” at the bottom of a receipt is probably the total, and that a date near “Commencement” in a lease is probably the start date. Most of the time those patterns hold. Sometimes they do not.

Where it goes wrong

AI reading errors are rarely random. They tend to cluster around a few predictable situations.

Faded thermal paper

Many receipts are printed on thermal paper that fades with heat, light and time. A receipt left on a dashboard for a week may be half invisible. If a human would struggle to read it, the AI will too, and it may guess.

Handwriting

Handwritten tips, totals changed by hand, or notes added to a lease are much harder to read than printed text. Handwritten numbers are a common source of errors.

Unusual layouts

Receipts from small businesses, restaurants with split bills, or foreign merchants may not follow the patterns the model expects. Leases written from scratch rather than from a standard template can put key terms in unexpected places.

Number and date formats

Formats vary by country and by habit, and some are genuinely ambiguous.

What is printed	One reading	Another reading
1.234,56	1,234.56 in the style used in much of Europe	Misread as 1.23 or as 123,456
1,234.56	1,234.56 in North American style	Misread as 1.23456 by a European-style reader
03/04/2026	March 4, 2026	April 3, 2026

The AI can often use context, such as the currency or the merchant’s country, to pick the right reading. But when the clues are weak, it has to make a choice, and it may choose wrong.

Why human review still matters

The goal of document AI is not to replace you. It is to do the typing so you only have to confirm. A quick review takes a few seconds and catches the mistakes that matter most.

When you review, focus on:

• The total and tax. These affect your records and any reimbursement or bookkeeping downstream.
• The date. Especially if the receipt uses a day-month format or the print is faded.
• The merchant. Abbreviated or oddly printed names are common.
• For leases, the dollar amounts and dates. Rent, deposit and the lease term are the details you will rely on later.

This is not legal or tax advice. If a receipt or lease is important for your taxes or a legal matter, keep the original and check with a qualified professional.

Corrections should stick

Here is a detail many people never think about. Suppose you correct a total from $48.00 to $84.00 because the AI swapped the digits. Later, the app re-reads the receipt, perhaps after an update. If the new read simply overwrites your fields, your correction is gone and the wrong value is back.

A good system treats your correction as the final word. It should mark the field as edited by you and never replace it with a later automatic read. Your review is only worth the effort if it is respected.

Privacy: keep only what is needed

Receipts and leases contain personal information, and some of it is sensitive. A sensible approach is to store only what you actually need.

• Card numbers. The last four digits are enough to tell your cards apart. There is no reason for an expense app to keep the full number.
• Lease details. Keep the terms you need to manage the tenancy, and be thoughtful about who can see the document.
• Original images. Keeping the image is useful as a record, but store it somewhere secure and delete what you no longer need.

When evaluating any app that reads your documents, ask what it keeps and why.

How SpendFlow and TryRentGo handle this

SpendFlow uses AI to read the merchant, date, subtotal, tax, total, currency, card type, last four digits and line items from receipts captured by camera, gallery or PDF. You review the details before saving, your corrections are marked and never overwritten by a re-read, and if a read fails the receipt is still saved. Only the last four card digits are ever stored. For landlords, TryRentGo lets you upload a lease and have AI fill in the tenant, rent, dates and deposit for you to check. You can learn more in our expense tracking industry guide.

The short version

Document AI is very good at the boring part: turning a picture into structured fields. It is not perfect, and its mistakes are predictable. Take clear photos, give every result a quick look, and use tools that respect your corrections and keep only what they need.