How the bench above would become a working tool for one finance team.
Today, by hand, for one invoice
- Open the email and save the PDF (2 min, estimate)
- Type the supplier, invoice number, date and amounts into the accounting software (5 min, estimate)
- Re-add the lines and check the VAT (3 min, estimate)
- Find the purchase order and compare it (4 min, estimate)
- Forward it for approval and follow up (3 min, estimate)
About 17 minutes per invoice, estimate.
The design
- Pick up invoices from the inbox or folder where they already arrive.
- A model reads each one into the fixed fields. A field with no source in the document holds the invoice.
- Code checks the supplier and invoice number against every invoice already processed.
- Code runs the arithmetic, VAT, TRN and purchase order checks.
- Code routes the invoice: to the approval queue if every check passed, to review with the reason if not.
- Cleared invoices are entered as bills awaiting approval, with a retry if the write fails.
- Every step goes to an audit log, and a person checks a sample of each week's readings.
Model or code
Uses a model
- Reading the fields from the invoice
Plain code
- Every check, including duplicates
- Routing to approval or review
- The audit log
Three ways it fails
- A scanned invoice the model cannot read
- A required field with no source in the document holds the invoice, with the missing fields named. A misread amount breaks the arithmetic check and holds it too. Nothing is guessed.
- A supplier on two invoices with the same number
- Code holds the second as a possible duplicate, linked to the first, and a person decides: a reissued invoice and a duplicate look the same. The bench above does not run this check.
- A VAT line that is right but on the wrong base
- The VAT check passes if the subtotal itself includes a line that should not be taxed. A client build also totals the lines by the VAT treatment printed on each and holds the invoice when the two bases differ. Where no treatment is printed, the tool cannot tell, and the rule comes from your accountant.
The approval point
Payment. The tool can enter a bill but never pays one. A named person in finance approves each invoice first, and clears the reason on a held one.
Estimate
Running cost per 100 invoices
From Anthropic's published price list, 1 October 2026: Claude Opus 5.5 at USD 4 per million input tokens and USD 20 per million output tokens. Assumes about 2,000 input tokens and 300 output tokens per invoice.
- 100 × 2,000 = 200,000 input tokens = USD 0.80
- 100 × 300 = 30,000 output tokens = USD 0.60
- Total: USD 1.40, about AED 5.14 at 3.6725
This is the model cost only. It excludes hosting, the time your team spends reviewing, and adapten's work.