AI invoice approval: Claude reads the invoice, code checks the totals and duplicates

AI invoice approval: Claude reads the invoice, code checks the totals and duplicates#

The process looks much the same in most small companies. An invoice arrives by email, someone saves it to a folder, retypes the amounts into a spreadsheet, checks it was not paid already and asks the boss to approve. Retyping takes the most time; approving without checking carries the most risk.

This n8n workflow splits the work so the model only does what it is good at: reading the document.

The model reads, it does not add up#

The file from Drive goes to the Anthropic API as a document block (PDF) or an image block (photo). The prompt asks for JSON only and forbids guessing:

If a field is not visible, use null. Never guess a plausible-looking number.

If the response cannot be parsed, the invoice does not crash the workflow. It gets an extraction_failed flag and goes straight to a person. An unreadable invoice is exactly the case someone should see.

Three checks the model does not do#

const diff = Math.abs((sub + tax) - tot);
math_ok = diff < 0.02; // rounding tolerance

Subtotal plus tax must equal the total to within two cents. A mismatch means either a reading error or an error on the invoice. Either way there is no auto-approval, whatever the amount.

The second check is the duplicate: the workflow looks the invoice number up in a Google Sheets ledger. The third is the amount threshold (500 by default).

A person clicks a link#

The message to the approver goes to a webhook (Slack, Teams, anything that accepts JSON with a text field) and carries the reason plus two links: approve and reject. The links point at the Wait node's $execution.resumeUrl, so the workflow simply waits for the click.

After the decision, the invoice is written to the ledger with the check results, the approval type and the running spend with that vendor, and the file moves to a "processed" or "flagged" folder.

Limitations#

Duplicates are checked on the invoice number alone. Two vendors that both send invoice "1/11/2026" produce a false duplicate. A safer key is the number plus the vendor's tax ID or name.

The Google Drive trigger polls the folder every minute. Several files dropped at once arrive in one run, while the later steps are written for one invoice at a time. For bulk uploads, process files one by one.

The threshold compares the gross amount regardless of currency.

In Poland, since 2026 most business-to-business invoices go through KSeF as structured XML. Those do not need a model to read them. This workflow fits foreign invoices, receipts, bills from outside KSeF and scans.

When it does not make sense#

With a dozen invoices a month, manual approval is faster than maintaining another automation. The workflow starts paying off with hundreds of documents and several people in the approval chain. If invoices should flow into accounting software, with approval permissions, history and deputies, that is a process automation project, not a single workflow. We use the same split between model and code when analysing supplier price lists.