Document classification that sorts and routes the pile for you

The hidden cost of sorting by hand

Classification is the step nobody counts. A shared inbox fills with mixed documents; someone opens each one, decides what it is, and forwards it to the right person or drops it in the right folder. It feels trivial per item, but across a full inbox it consumes real time, introduces delay, and goes wrong under pressure, an invoice filed as correspondence, a complaint sitting unread in the wrong queue, an urgent contract lost among receipts. The cost is not just the sorting time, it is everything that waits downstream because a document was mis-routed.

It is also a poor use of people. Deciding whether a document is an invoice or a delivery note is exactly the kind of repetitive judgement that a language model does well and consistently, at any hour, without a backlog. The model reads the content and the structure, not just a filename or a subject line, so it classifies correctly even when the sender mislabels the attachment or dumps three document types into one PDF. Your team stops being a sorting office and starts working the documents that actually need their attention.

• Per-item sorting that quietly consumes real team time
• Mis-routed documents that stall everything waiting downstream
• Errors under load: complaints and urgent items in the wrong queue
• A model reads content, not just filenames, so mislabelled files still sort correctly

How classification and routing work

We start from your categories, the document types you actually deal with and where each one needs to go. The pipeline reads each incoming document, using OCR for scans where needed, and the model classifies it against your categories based on its content and structure. It can also pull a few key fields at the same time, so routing can depend not just on type but on detail: a high-value invoice to a different approver, a complaint about a specific product to the right team, a contract from a key account flagged for priority.

Once classified, the document is routed: filed in the right place, sent to the right person or queue, or handed straight to the matching processing pipeline, an invoice into the invoice flow, an application into form processing. Classification is often the front door to the rest of the document automation we build, the step that decides which downstream pipeline each document enters. And because it is content-based, adding a new category or adjusting a routing rule is configuration, not a rebuild.

• Classifies against your categories from content and structure, not filenames
• Extracts key fields so routing can depend on detail, not just type
• Files, forwards, or hands off to the matching processing pipeline
• Priority and exception routing for high-value or sensitive documents
• New categories and routing rules added as configuration, not a rebuild

Confidence, exceptions, and the human check

Misclassification has a cost, so we build for uncertainty rather than assuming the model is always right. Every classification comes with a confidence score. Documents the model is confident about route automatically; anything ambiguous, a type it has not seen, a document that looks like two things at once, low confidence, is sent to a person to classify instead of being guessed and mis-routed. That human decision is logged, and over time it shows us where to tighten the categories or where a genuinely new document type has appeared.

This keeps the automation honest. A classifier that confidently mis-files ten percent of your documents is worse than one that sorts ninety percent automatically and cleanly hands the hard ten percent to a person. We tune the confidence threshold with you to match how costly a misroute is in your process, tighter where a wrong turn is expensive, looser where it is cheap and speed matters more. You control that balance, and you can watch how the classifier performs because every decision is logged.

Yours to run, extend, and audit

The classifier runs in your accounts and on your infrastructure, reading documents wherever they arrive and routing them into the systems you already use. The categories, the routing rules, and the confidence thresholds are documented and yours to change, so as your business adds document types or reorganises who handles what, your team can adjust the logic without coming back to us. Every classification and every human correction is logged, giving you an audit trail of what was sorted where and why, and the data to keep improving the categories over time. No lock-in, no black box deciding where your documents go.