Document Processing
Automation
How the pipeline works
Read the document
Extract the facts
Send for review
You get checked data with the source attached.
Ingestion: we hook into specified inbox(es), upload points and/or shared folders, including access controls.
Documents are diverse, arriving via different routes, in different formats, and different teams will need different elements from each document.
Extraction and validation: we analyse documents using OCR and vision models and extract required fields such as totals, dates, parties, and line items into target format. We apply business rules, such as checking whether totals match, the VAT number is valid, the supplier is known, the PO is open, or the delivery date is in a sensible range. Tools tested include OCR tools, document processing platforms and vision models. We don't have an opinion on the right approach. We look at accuracy, time/cost to process, ease of use and manageability. We correlate confidence scores with actual error rates to understand when documents can be posted automatically.
Human review: where rules indicate the document should be routed to a human, they can be viewed via a review interface showing the original document alongside the extracted elements. Depending on rules, some documents will always go to a human. Errors aren't eliminated but reduced. Medical records have additional requirements including clinical safety assessment, human review processes and end-to-end traceability for clinicians.
Posting: where all checks are OK, extracted information can be posted to destination system(s).
Auditability: we provide an audit trail including the original document, rules applied, approver and downstream postings.
Testing and piloting
We would need some example documents including any exceptions, tables, poor quality scans, mixed document types, unusual layouts. During a pilot, we can process a single document type alongside your current process before integrating with destination systems. We can test performance, agree acceptable error rates, review queues, user access etc. We can assess reviewer effort, failed runs, compute/storage requirements, hosting, overall cost compared to your current process.
Document audit
Send us some examples of your documents and we can analyse them, look at who deals with them today, where they're used, who/what they're fed into, what's being re-typed, exceptions, pain points etc. You can then book a document audit with us to define a pilot that can be rolled out safely. We will need some example documents, an estimate of volumes, examples of previous exceptions, target system(s) to post to etc.
Ownership, hosting, integration
You will be able to see the code, prompts, model(s) etc. We can work on your stack or ours. Integration is highly variable and some destinations may not support automation. Some channels and destinations may be retained, review tools may change. Ultimately, you should be able to control and understand what's happening. You own the code, prompts, models, data, all of it.