Describe the result for one document

The task of “recognizing bills” is too broad to be understandable. Start with a specific result: extract the number, date, supplier, items and total, compare them with the directory and prepare a draft entry. Determine which fields are required, which fields can be left blank, and who checks for discrepancies. Save the original file next to the result so that the employee can quickly find the disputed value.

  • Select the document type and acceptable input formats.
  • Describe fields, units of measurement, and rounding rules.
  • Indicate directories of suppliers, products and departments.
  • Determine the result of processing and the conditions for transfer for verification.

Separate extraction and verification

Readable text does not mean that all details are defined correctly. The document number may be confused with the order number, and the amount with the subtotal. After extraction, check the format of the fields, the connections between the lines and the total, and the match with the directories. If the system provides a confidence rating, use it in conjunction with these checks and test results. A high score alone is not enough to automatically accept a document.

Example of checks after recognition
FieldReviewReason for stopping
DateThe value exists and refers to the required attributeIt is unclear whether this is the date of the invoice or the order
SupplierDetails are compared with a valid cardSeveral suitable organizations found
Positions and totalQuantity, units and calculation agreed uponLines are incomplete or the total is not explained
Document numberChecked for similar entryIt is possible to repeat a previously processed file

Make the exception queue convenient

The verification queue is needed for documents that have not passed the agreed upon conditions. Show the reason for the stop, the recognized value, and the location in the source file. Allow the employee to correct the field and continue processing without re-uploading the entire document. Distinguish between poor scan quality, unknown item, and integration error: these problems have different owners and solutions.

DEMONSTRATION EXAMPLE

The request states “10 boxes,” but the directory stores the product in pieces. Automation should not silently write down the quantity 10 as ten pieces. It shows the original string and asks for the conversion factor. After confirmation, the result is saved along with the selected rule.

Maintain a link to the original file

For each record, it is useful to know which file it came from, when it was processed, and which fields the employee changed. Re-uploading the same file and a new document with a similar number require different checks. Before recording in the accounting system, define the rules for recognizing repetitions, checking access and recovering from failure. The first version can be limited to the preparation of drafts with confirmation by the person in charge.

Include correction time in the calculation

Compare the full cycle: receiving a file, extracting, checking, correcting and transmitting the result. Add different layouts, multi-page documents, fuzzy scans, and blank fields to your test sample. See what errors are repeated and how many documents require manual parsing. It makes sense to scale the script after the team understands the structure of exceptions and knows how to handle them without losing data. If the path from the file to the verified entry is already agreed upon,business process automationhelps to combine field extraction, manual verification and result transfer into one route.

Apply to your task

Tell us which files you process manually. We will define the required fields, checks and human review for the first pilot.

Discuss document processing →
← Return to materials