All cases

Automation · Receipt and warranty app

Kassabon.aiReceipt, warranty and claim assistant, live at kassabon.ai

A receipt photo becomes a purchase record with a warranty date, reminders and a claim letter the user sends.

Service
Product design and full buildAI automation →AI agents →
Year
2026first commit April 2026
Status
Livekassabon.ai, free and premium tier
View website
A phone held above a kitchen table shows a purchase in Kassabon.ai with its warranty details, the paper receipt lying below it.A purchase and its warranty, next to the receipt it came from

Verified system facts

Counted in the codebase, September 2026

4layersExtraction pipeline: document parse, deterministic field extraction, model enrichment, decision log. Vision OCR runs first and feeds it
30 / 14 / 3daysWarranty reminders before cover ends, plus one the day after, moved out of the user's quiet hours
2languagesDutch and English, with consumer-law copy kept explicitly Dutch
DraftfirstClaim letters are concepts the user reviews and sends; the system sends nothing on its own

The facts above are counts in the repository. We do not publish recognition accuracy or claim outcomes: we have no measured figure we would defend.

The problem

Receipts fade, warranties are forgotten, and when something breaks the burden of proof lands on the person who kept the box.

  • Thermal receipts fade long before the warranty ends
  • Nobody remembers when cover expires until it already has
  • A claim under Dutch consumer law needs the right facts in the right order
  • A generic AI would happily invent an amount or a date

The system

Deterministic first, AI second. The model fills gaps and never overrules a locked field.

An uploaded receipt is read by vision OCR and passed through a four-layer pipeline. Deterministic parsing extracts merchant, amount, date, invoice and order numbers and locks anything it is sure about. The model only fills the fields that are still empty or uncertain and may not touch a locked value; every decision is logged, so we can see what was deterministic, what was enriched and what the user had to correct. Each purchase gets a warranty end date and a reminder chain, and the claim engine composes a letter, framed in Dutch consumer law, as a concept the user reviews. If the enrichment call fails, deterministic parsing still produces a usable record and the letter falls back to a Dutch template. Reading the receipt itself does depend on the vision model.

  1. UploadA photo or PDF of the receipt goes to the user's own storage folder.
  2. ExtractOCR, then deterministic parsing; high-confidence values are locked.
  3. EnrichThe model fills empty or uncertain fields only; locked fields stay untouched.
  4. Review and sendThe user confirms the record; a claim letter stays a draft until they send it.Human handoff
A woman in a mustard jumper photographing a crumpled receipt with her phone while receipts fly around her
Photographing a receipt is enough: Kassabon.ai reads out the warranty date.

The result

One place for every purchase,
and a claim that is ready before you are angry.

  • Purchase records with warranty dates and a reminder chain before cover ends
  • A claim concept framed in Dutch consumer law, ready for review
  • Deterministic parsing stands on its own: if the enrichment call fails, the record and the claim letter are still produced
  • Dutch and English; a free tier and a premium tier, billed through Stripe

From the Lab

“Kassabon.ai is what the Lab builds best: a small, sharp job done properly. Read the receipt deterministically first, let the model fill the gaps, and never send anything on someone's behalf without them looking at it.”

Yourworkflow,oursystem

We build systems that quietly take over work. One conversation is enough to see where it starts for you.