› Synthetic invoice reconciliation lab
Invoice reconciliation checks whether what was billed matches what was ordered and received. Try a controlled three-way match using fictional data: rules compare the invoice, purchase order and goods receipt; AI explains the evidence; a person makes the final decision. This is a custom-automation capability demo, not an accounting SaaS product.
› Interactive capability demo
The documents are fixed, fictional examples. The server accepts only the scenario identifier—never a file, supplier record or confidential business document.
Synthetic scenario
Selected scenario
Exact three-way match
Fictional document · demonstration only
| Description | Qty | Unit price | Total |
|---|---|---|---|
| Industrial sensor kit | 10 | €120.00 | €1,200.00 |
Complete the security verification to run the server workflow.
One-hour review token · 24-hour private retention · no uploads · no production data
› Control sequence
Only one of three server-owned synthetic scenarios is accepted. There is no public upload path.
Supplier, PO, currency, quantity, price, total and duplicate rules determine the workflow state.
Gemini translates the fixed rule evidence into a concise reviewer summary; it cannot change the result.
A short-lived hashed token authorizes approve, request-correction or reject actions.
Approval creates a clearly marked demo draft instead of writing to an accounting platform.
Each transition is recorded and the synthetic private record is automatically eligible for deletion after 24 hours.
› Safety boundaries
The endpoint accepts a scenario ID only. Names, amounts and documents are created server-side and are explicitly fictional.
Money, quantity and duplicate checks are deterministic. AI produces an explanation, never an approval decision.
Turnstile, same-origin checks, request-size limits and a separate rate-limit bucket protect the public workflow.
Review tokens are random and stored only as hashes. Sessions expire after one hour and records after 24 hours.
› Questions
Extraction reads fields such as supplier, invoice number, quantity and price. Reconciliation compares those fields with independent order, receipt and ledger records. A correctly read invoice can still be duplicated or bill more than was received. This lab starts with structured synthetic fields; it does not test OCR on uploaded files.
No. Public arbitrary uploads would require malware scanning, document-isolation controls, stronger retention tooling and client-specific privacy terms. This demo deliberately accepts only synthetic scenarios.
No. Deterministic checks create the pass, review or block state. Gemini only explains those results. A human records the final decision.
A custom project can connect an approved inbox or document store, purchase orders, receipt records and accounting software when supported APIs or imports are available. We first verify supplier identifiers, tolerances, permissions and review ownership. Posting and payment require separate scope and controls. The public demo creates a non-posting draft only.
Touch time per invoice, straight-through match rate, exception reasons, duplicate prevention, correction cycle time, approval latency and reconciliation accuracy against a reviewed sample.
Describe the documents, systems, exception rules and approval boundary. We will assess where deterministic automation and supervised AI can safely help.
Discuss invoice automation