The task
A shared inbox is a queue of documents in no order, several of them stapled into one PDF. Configure an agent to identify what each document is, split the files that hold more than one and route them to the right queue with the reason recorded. Getting the splitting right matters more than getting the classification right, because a wrongly split file loses a document entirely.
What you provide
The inbound stream
Whatever arrives: scans, email attachments and photographs, frequently with several documents in a single file.
Your document types
The types you route on, with an example or two of each, including the ones that look similar to each other.
Your routing table
The queue or team that receives each type, and the confidence below which a document goes to a person instead.
Give the agent a clear brief
Use these instructions as a starting point, alongside your approved sources and permissions.
Split each file into individual documents and state where each document begins. Classify each one into a supplied type and give the evidence for the classification. Route it using the supplied table. Send anything below the confidence threshold, and anything that fits no type, to the human queue. Do not delete, merge or discard any page, and keep every page traceable to the file it arrived in.
How the work comes together
Split before classifying
Establish the document boundaries first. Classifying a file that holds three documents produces one answer for three things and the other two disappear.
Classify with evidence
Assign a type and record what indicated it. Documents that look alike, such as an order and a delivery note, should show why one was chosen over the other.
Route, or hand over
Apply the routing table, and send low-confidence and unrecognized documents to a person. An unrecognized document reaching a human queue is a success, not a failure.
Illustrative example · fictional data
One file, three documents
An illustrative batch from a fictional shared inbox, routed with a confidence threshold below which a person reviews.
| Document | Basis for the type | Routing |
|---|---|---|
| Pages 1 to 2 of a 5-page scan | Header and totals of a supplier billing document | Accounts payable queue |
| Pages 3 to 4 of the same scan | Signed delivery confirmation | Goods receipting queue |
| Page 5 of the same scan | Fits no supplied type | Human queue for review |
Page 5 is the reason a threshold exists. A confident guess would have filed it somewhere plausible and nobody would have looked again.
The human review
An operations reviewer works the human queue and can correct any routing. Every page stays traceable to the file it arrived in, so a wrong route is recoverable.
Set it up for your team
Test splitting first
Evaluate boundary detection on real multi-document files before you evaluate classification. A splitting error loses a document; a classification error only misdirects one.
Include the confusable types
Supply examples of the types that resemble each other. Those pairs are where the accuracy of the whole queue is decided.
Make the human queue normal
Set the threshold so uncertain documents reach a person routinely. A routing system with no human queue is a system that hides its errors.
Check the result before scaling
Use a batch reflecting your real intake, including multi-document files, poor scans and types that rarely appear. Measure splitting errors, misroutes and the share reaching the human queue, and watch the rare types separately.
Keep the data in the right place
An intake stream is the widest data flow in most operations teams, carrying documents from every part of the business. A fully local deployment keeps the whole stream inside your network.
Explore deployment optionsQuestions & answers
What happens to a document type we did not define?
It goes to the human queue. Forcing an unknown document into the nearest defined type is how documents quietly disappear into the wrong process.
Can it handle several documents in one PDF?
That is the main case it is configured for, and the one to test first. Boundaries are established before any classification is attempted.
Will it delete duplicates it finds?
No. It never deletes or merges pages. Identifying possible duplicates is a separate review with a person deciding.