Classification decides the type of each document, splits multi-page packets into the documents they actually contain, and routes each one to the team or queue that owns it. It runs on your hardware, against a taxonomy you define, with a confidence score on every decision.
Accuracy is measured field-by-field against a human-adjudicated ground truth on a stated document mix. Classification performance against your own taxonomy is measured separately during the Proof of Value, on your documents.
The hard part is almost never telling an invoice from a passport. It is the 40-page PDF that contains eleven documents, three of which are photographs of the same page.
Each document is assigned a class from your taxonomy using layout geometry, template match, reading order and text together, rather than keyword matching that breaks on a reworded header.
A single scanned PDF containing a whole submission is divided at real document boundaries, so downstream extraction sees eight documents instead of one confused one.
Class plus extracted context decides the destination: a team, a workflow, a topic, a folder or an API callback. Routing rules are yours to write and version.
Accept, review and reject bands are set per class, because misrouting a brochure and misrouting a medical report are not the same mistake.
Anything that does not match is labelled unknown rather than forced into the nearest class. Recurring unknowns are clustered so you can add a class deliberately.
Arabic-only, English-only and bilingual versions of the same form resolve to the same class, and a packet that mixes scripts page by page splits correctly.
One PDF arrives from a broker. It contains everything, in no particular order, scanned at three different qualities. This is what happens to it.
Pages are deskewed, dewarped and contrast-corrected, then segmented into blocks and tables with reading order resolved. Classification never works from raw pixels or from a flat text dump.
Layout discontinuity, header and footer change, reference-number change, orientation flip, script change and template change are scored at every page break. The packet is cut where the evidence agrees.
The 40 pages resolve into an Emirates ID copy, a driving licence, a police report, two repair invoices, a medical report, a claim form and a policy schedule — each with a class and a confidence score.
Seven documents clear their accept thresholds. The eighth, a faxed medical annexe, lands at 0.71 against a 0.95 threshold and goes to triage with its two candidate classes and the evidence for each.
Invoices go to the motor assessment queue, the medical report to the clinical review queue under stricter access control, identity documents to KYC verification. Each destination is your rule, not a default.
Because each fragment is now a known document type, identification runs the right schema against the right document, and mapping has clean entities to link across the packet.
If a claim packet is treated as one document, the invoice schema runs against the police report, the medical report inherits the wrong access controls, and the extracted total belongs to no document in particular. Every error after that point is really this error, arriving late.
Classification is cheap insurance against every downstream capability being confidently wrong.
What you can define without a model change, and what the service returns for every decision it makes.
Taxonomy design is part of the Proof of Value. We build it against your real intake mix and measure classification performance on your documents, with the threshold agreed in writing up front.
Cloud classifiers work. The question is what it costs to change your mind about your taxonomy, and whether a medical report has to leave the building to be recognised as one.
Where documents are classified
Multi-page packet splitting
Adding a new document type
Reclassifying 2M archived documents
Behaviour on an unseen type
Arabic and bilingual documents
True air-gapped operation
Comparison reflects publicly documented behaviour of metered document-processing products as of 2026, including per-page pricing and enterprise-gated self-hosting. Verify against current vendor documentation before making a decision.
What comes up when the intake queue is the bottleneck and nobody trusts an automatic label yet.
Boundaries are detected from evidence, not from page counts. Layout change, header and footer discontinuity, reference-number change, orientation shift, script change and form-template change are combined into a boundary score per page break. A 40-page claim packet becomes eight separate documents with a confidence score on each split.
