Document Intake Agent for Accounting Firms: Invoices to Draft Bills

Illustrative scenario: one supplier invoice arrives in a client's inbox

This is an invented example, not a client result. Picture a small bookkeeping practice and a fictional client, a cafe, that forwards supplier invoices to a dedicated address. One morning a PDF arrives from a made-up wholesaler. Here is what happens to it.

Received

A workflow notices the email, saves the attachment to the client's folder, and records the sender, time and file name. The email body is kept but treated as untrusted text.

Read

The agent sends the file to a language model that accepts PDFs. Anthropic's documentation says Claude can work with the text, pictures, charts and tables in a PDF. The model returns a fixed set of fields: supplier name, invoice number, issue date, due date, currency, net total, tax total, gross total and line descriptions. Anything it cannot see clearly is returned as empty, not guessed.

Checked

Ordinary code, not the model, runs the tests. Net plus tax must equal gross. The supplier must match a contact in the client's ledger. The combination of supplier, invoice number and amount must not already exist. The date must fall in a believable range.

Held or queued

If every test passes, the document joins the review queue marked clean. If the supplier's bank details on the invoice differ from those on file, the agent stops, flags the document, and tells the bookkeeper to confirm the change by calling a number already on record, never one printed on the invoice.

Reviewed

The bookkeeper sees the document beside the extracted values, with each field linked to its location on the page. They correct anything wrong and approve or reject.

Posted

On approval, the workflow creates a draft bill in the ledger with the PDF attached and a suggested account code from the supplier's previous bills. A person still releases it in the ledger.

Logged

One record ties the source file, extracted values, checks, corrections and approver to the resulting entry. Our AI document processing agent page covers the general service.

How does it decide what to trust?

The model reads; fixed rules judge. That split matters because a language model can sound confident while being wrong, and arithmetic is better done by arithmetic. Four checks do most of the work:

• Totals reconcile. A document whose net and tax do not add up to the gross figure goes straight to a person with the mismatch shown.
• Known supplier. A new supplier is never created automatically. It is proposed, with the document attached, for a person to approve.
• No duplicate. Matching on supplier, number and amount catches the same invoice sent twice or forwarded by two staff.
• Nothing changed that moves money. Bank details, payment terms and remittance addresses that differ from the file are always held.

You can add client-specific rules, such as a ceiling above which a second person must approve. Our post on testing and evaluating AI agents explains how we would check extraction accuracy on your documents before relying on any of it.

What does the agent do alone, what needs approval, and what is never automated?

Task	Alone	Needs approval	Never automated
Collecting and filing documents	Yes
Extracting fields and running checks	Yes
Suggesting an account code	Suggests only	Reviewer confirms
Creating a draft entry	After review	Bookkeeper approves
Adding a new supplier		Yes
Acting on changed bank details		Callback by a person
Deciding tax or VAT treatment of an unusual item			Yes, a qualified person decides
Releasing or scheduling payments			Yes
Filing returns or contacting tax authorities			Yes
Deleting documents or ledger entries			Yes

Which accounting software can it write to?

Only software that gives outside tools access. From Xero's published OpenAPI description we confirmed that its Invoices endpoint covers purchase bills as well as sales invoices, that a draft status exists, and that files can be attached to an invoice or bill through an attachments endpoint that needs its own attachments permission. We have not verified QuickBooks Online's bill and attachment endpoints against Intuit's official documentation, so we would read those before scoping a QuickBooks build. For other packages, if there is an API or webhook a workflow in n8n or Make can usually connect it.

Where there is no usable API, the fallback is an import file or a spreadsheet that the bookkeeper uploads. It is slower, and we would say so rather than imply live posting. Related reading: our invoice automation for small business page covers a simpler, single-business case.

What happens to client confidentiality?

Client paperwork is confidential, so decide the data route before the first document. This is general information, not professional or legal advice; check your engagement letters, your professional body's guidance and your data protection obligations.

• Read the model provider's data-handling terms for the plan you would use, and choose a plan whose retention suits your clients.
• Keep each client's documents separated, with an access list that matches who works on that client.
• Set a retention period for extracted text and logs, and delete on schedule.
• Tell clients in your terms that automated tools help prepare entries and that a person reviews each one.

Our post on AI and data privacy goes through the wider checklist.

How would success be measured?

No figures are promised. We would measure before and after on the same clients:

• Handling time per document, from receipt to approved draft.
• Field accuracy per field, found by comparing extraction with reviewer corrections.
• Clean-pass rate and exception rate, with reasons for each exception.
• Duplicates caught and duplicates missed, the latter found at month-end review.
• Reviewer correction rate, which should fall as client rules improve.
• Turnaround time for a client's month-end pack.

To discuss your own mix of clients and software, contact us.