Papertrail is a document processing workflow for accounting operations. It identifies uploaded documents, extracts relevant fields and prepares structured information for review and downstream accounting systems. The experience connects the original source to the data it produces, so reviewers can check and correct the information before it becomes part of an accounting record.
03 / THE PROBLEM
The document is digital. The work often isn’t.
An invoice or receipt may arrive electronically, but someone still has to identify its type, locate the right fields, enter values, check accuracy and transfer the information into another system. Papertrail needed to reduce that repeated work while keeping human oversight available wherever the document or extracted data left room for uncertainty.
04 / Defining the product
Design for everything after upload.
The workflow could not stop at upload and success. A document moves through identification, extraction, checking and posting, with incomplete information and exceptions along the way. I treated review as a working part of that lifecycle: the person needs to know what requires attention, see the source and correct a field without starting over.
01Automate the certain
Information the system understands confidently can move forward without demanding the same attention as a missing or uncertain field.
02Review the uncertain
Make exceptions identifiable and actionable. A reviewer should be able to focus on the part of the document that needs judgement.
03Keep the source visible
The original document stays connected to its extracted fields. Corrections happen within that context, rather than through another upload or a separate checking task.
05 / Designing the system
Document in. Structure out.
OCR makes the document readable, classification establishes its type, and extraction gives the relevant fields a structure. Confidence guides human review before validated data reaches the accounting system. Keeping these stages connected makes it possible to trace an accounting value back to the document that supplied it.
Treating every extracted field as equally reliable would make reviewers check the whole document again. Treating every field as an exception would remove much of the benefit of automation.
02The decision
Use confidence to focus review on uncertain or incomplete information, with the source available beside the extracted record. Keep correction inside the same document lifecycle.
03The result
The review model gives human judgement a specific place in the automation, instead of requiring a second manual processing workflow.
07 / The solution
Keep the document and its data in view.
JOURNEY 01
AI confidence is an interface problem too.
Extraction is not equally reliable across every document and field. Papertrail distinguishes information ready to continue from information requiring judgement. The review experience keeps the source and extracted fields together, so attention follows the specific uncertainty.
01Missing or uncertain fields
A missing value and a doubtful value are different review tasks. The reviewer needs to see which field requires attention and refer to the source before supplying or correcting information.
02Ready for handoff
A reviewed record is distinct from a record that has reached the accounting system. Extraction success should not stand in for completion of the accounting handoff.
03An exception in the document
Keep the document and the work already completed available during correction. The reviewer should not need to upload it again or rebuild the extracted record.
JOURNEY 02
Correct without starting over.
The correction journey starts with a specific field, not with the entire pipeline. The reviewer checks the source, adjusts the value and validates the record before it continues. This preserves completed work and keeps the accounting value connected to its origin.
01Locate the source
02Correct the field
03Validate the record
04Continue the handoff
Exception recovery / the original document stays attached
08 / Outcome & impact
A review model built around the uncertain field.
My design contribution was a document lifecycle that connects extraction to focused review, correction and accounting handoff. Confidence determines where human attention is needed; the original document supplies the context for resolving it. A corrected field stays within the existing record, and a reviewed record remains distinct from one that has reached the accounting system.
01Review with context
The source document, extracted fields and correction path stay connected through the handoff into structured accounting data.
02What I learned
Automation works better when uncertainty is part of the experience. A confident result and an uncertain result call for different kinds of attention. Designing that difference is as important as designing the successful path, because it determines how useful the automation is to a reviewer.
Evaluate the complete path from document intake to accounting handoff. Successful extraction alone does not establish that the record is accurate or ready to use.
Review time
≤2 min
Median time to resolve exceptions and reach a reviewed record for a standard document.
Field accuracy
≥99%
Reviewed values matching the verified source, including errors missed during review.
Handoff success
≥98%
Reviewed records accepted by the accounting system without correction or reprocessing.