Skip to content

Invoice Inbox Agent

From the email inbox to the bank transfer. The agent sorts incoming mail, reads each invoice, checks it against 18 rules and hands it to a person for approval.

Role
Concept, UX/UI, build
Stack
Next.js, TypeScript, pdf.js, zod
For
Small businesses in Austria & Germany
Year
2026
Invoice Inbox Agent overview with open payables, time saved and a cash out forecast

What it does

Small businesses get invoices by email in every format you can think of. Someone opens each one, types the numbers into accounting, checks the VAT, hopefully notices when a bank account changed and plans the payment.

The agent takes over the reading and checking. A person still makes every decision, but now with the original PDF on one side, the extracted fields on the other and a clear warning wherever something doesn't add up.

The numbers

  • 18

    automatic checks on every invoice

  • 6 min

    saved per invoice, estimate

  • $0.55

    model cost per 1,000 invoices, estimate

Minutes per invoice

  • Typing it in by hand7 min
  • Checking what the agent read1 min

At 30 invoices a month that's about 3 hours back. At 300 it's close to a working week.

How it works

The language model only reads. Everything that decides something is plain code you can test, and a person signs off. Uploads are read right in the browser, nothing leaves the device.

  1. InputInboxEmails and PDF uploads come in
  2. AIReadSorts mail and extracts 20+ fields with a confidence score
  3. CodeCheck18 rules for math, tax, fraud and process
  4. PersonApproveReview side by side, second approval above €1,000
  5. OutputPayPayment run, SEPA XML and accountant CSV
InputAICodePersonOutput
  • [ Next.js ]
  • [ TypeScript ]
  • [ pdf.js ]
  • [ zod ]
  • [ Zustand ]
  • [ IndexedDB ]
  • [ React PDF ]

Screens

1 / 3
Simulated inbox where the agent sorts incoming emails
The inbox. The agent sorts invoices, reminders, credit notes and spam, and flags phishing.
Review queue listing invoices with flagged problems first
Review queue. Problems first, each with a clear reason and a fix.
AI quality page with accuracy and cost figures
AI quality. Accuracy against labelled invoices, cost per run and a model comparison.

What it brings

  • Hours back every month

    Typing numbers is gone. What's left is a quick look and a click.

  • Fraud gets caught

    A changed bank account blocks the invoice until someone confirms it by phone.

  • Clean handover

    Exports for the accountant and the bank, plus a full log of who approved what.

  • Tiny running cost

    Reading a thousand invoices costs less than a coffee at current model prices.

Good to know

In this version the language model is simulated. It replays recorded answers so anyone can try it without an account or a key. The 18 checks, the PDF reading for uploads, the exports and the whole review flow are real code. Going live means a real model call, a mailbox connection through Microsoft Graph, a database and a login.

→

Watch Inventory