ADP is the document AI of CP4BA: in Document Processing Designer (part of Business Automation Studio) you define document types (invoice, claim form, ID) and their fields, teach the classifier with samples (machine learning), and let deep learning models extract fields; you add validation rules and data clean-up. The project is deployed to the document processing runtime, which also provides the verification and validation user interfaces.
Integration paths with workflow:
- Case / workflow solution with document processing - the document processing project is linked to a case or workflow solution in Studio; new documents (uploaded in a coach or a Workplace page, dropped into a monitored folder, received by e-mail through a capture channel) go through classification and extraction; a verification human task is created only when the confidence is below the threshold or a rule fails; the extracted data lands in case properties / process variables, and the classified document is filed in the case folder in Content Platform Engine.
- From a plain BPD - a service flow calls the document processing REST API (submit a document, poll or receive the result JSON), maps the fields to a business object, and continues; useful when the process starts outside a case.
BPD "Claim intake"
[Upload claim documents] (coach with Document List, folder = tw.local.claimFolderId)
[Process documents] (document processing step: project "Claims", input = documents of the folder)
-> per document: type, fields with confidence, validation results
[Needs verification?] gateway: any field confidence < 0.85 or a validation error
yes -> [Verify extraction] (the ADP verification task: the user sees the page image and the fields side by side, corrects and confirms)
[Map fields] (script)
tw.local.claim.amount = tw.local.docResult.fields["TotalAmount"].value;
tw.local.claim.policyNo = tw.local.docResult.fields["PolicyNumber"].value;
tw.local.claim.docType = tw.local.docResult.documentType;Practical points: train with real, varied samples (at least a few dozen per type; a few hundred for deep learning extraction quality); set thresholds per field (amounts and identifiers strict, free text lenient); keep the raw extraction result in the process for audits; corrections made in the verification UI can be fed back as training data (a periodic re-training cycle owned by the business); the document processing runtime scales separately and can use GPUs for the deep learning models; and migrations from Datacap keep the same idea (taxonomy, rules) but the models are rebuilt in the Designer.
References