Small Local Models for Accounting Document Classification

One file, two kinds of document

A supplier sends a four-page PDF. Pages 1 and 2 carry the heading „Invoice”, and pages 3 and 4 are a „Delivery confirmation” signed by the recipient. Invoices go to accounting review, and confirmations to the purchasing queue. If the system labels the whole file „invoice”, the delivery confirmation lands in accounting and the person who checks deliveries never sees it.

A useful result sets the boundary between pages 2 and 3, labels both parts and routes each to its queue. When the heading on page 3 is blurred, the system reports the uncertainty and passes those pages to manual review, noting which pages the problem concerns.

Where accounting document classification ends

The model recognises the type, and the workflow routes the document to the agreed queue. The scope ends before tax assessment, posting and the accounting entry. The company approves the list of types, the routing rules and the person for uncertain cases, and the model does not add categories by guesswork. In Poland, invoices have arrived through KSeF as structured data since 1 February 2026, and across the EU, cross-border B2B sales move to digital reporting based on e-invoices from 1 July 2030. Classification therefore makes most sense for the rest of the correspondence: confirmations, reports, contracts and attachments.

An average accuracy over the whole set hides a frequent error on one type of document. Until the test is passed, do not extend the workflow to further types.

What classification costs locally and through an API

A local model does not mean the whole flow stays on site. Microsoft describes Foundry Local as processing prompts and answers on the device, and the model files are downloaded on first use. Check the whole flow: where OCR runs, where temporary files go, what the logs record and which systems receive the document. Run the trial on the hardware that will later do the work, with a batch from the end of the month.

When a simpler solution is enough

If each PDF contains one document with a fixed layout, rules or a ready module in the document workflow system are enough. First check whether your current system already recognises types and routes them to the right people. A local model is worth comparing when varied layouts and content keep rules from recognising documents reliably. When the company allows a hosted service, compare it with the local option on the same documents and under the same access rules, then choose the cheaper option that passes the test.

We build classification as part of private LLMs, and the invoice workflow is described in cost invoice automation. For the free process review, bring a mixed PDF you are allowed to show, the list of types and queues, and the number of documents in a normal and a peak month.