How to read invoices with AI? Introduction steps from Excel transcription to verification

When automating the reading of invoices, it will be easier to find problems if you try the process of creating a list for posting instead of linking accounting registration or transfer from the beginning. Being able to read the text with OCR and being able to process it correctly as a transaction are two different things. The process of checking the amount, business partner, and deadline with the original remains.

Extracted items to be fixed first

ItemHow to think about savingExamples of things to be careful about
Original IDValue that allows you to return to the file or reception recordOverwriting files with the same name
Invoice numberKeep as stringLoss of leading zero
Business partnersSeparate name and internal business partner IDAbbreviation/branch/former company name
Billing date/payment deadlineOriginal notation and converted formOmitted year, end of month notation
AmountDistinguish between currency, tax excluded, tax amount, and tax includedDiscount, multiple tax rates, foreign currency
Confirmation statusUnconfirmed/Confirmed/RejectedError of assuming blank fields as zero

The extraction tables in this article alone cannot determine qualifying invoice requirements or tax retention requirements. Accountants decide on the items necessary for their company's accounting and tax processing.

Five steps from reading to registration

  1. Attach the original ID at the time of reception and record the storage location and processing status.
  2. Read characters/tables and separate them into defined items.
  3. Mechanically inspect the amount, date, and required items.
  4. The person in charge checks the original and updates the confirmation status.
  5. Import only verified rows into the existing system.

Don't skip matching just because the "confidence" given by the model is high. Whether higher confidence means fewer errors must be verified using actual reports.

The difference in practice is whether you can handle exceptions.

Try cases where the text is too thin, the stamp overlaps with the amount, the details continue on multiple pages, and the total is in two places. Rather than filling in illegible numbers by guessing, the design is designed to indicate the corresponding part of the original document as "confirmation required."

To detect double processing, we combine not only the invoice number but also the business partner and original document information. Since there are some invoices without numbers, we will consider a method to return them as duplicate candidates rather than automatically deleting them.

Separate evaluation by item

Even if the name of the client is correct, if the amount is incorrect, it will cause a big problem in practice. Instead of combining all items into one correct answer rate, errors are counted for each important item such as amount, business partner, deadline, etc. We also measure work time including corrections and verification.

You can record the original ID, extracted values, and confirmation results using CSV for checking extraction results. The sample is fictitious billing information and is not in the format for importing into accounting software.

Notes on passing to Excel

Treat the identification number as a character string and avoid uniformly converting dates and currencies. Also check the output format to ensure that strings extracted from external documents are not executed as formulas. Instead of double-clicking the CSV as is, consider specifying the column type when importing.

FAQ

Can I automate the transfer directly?

Apart from reading, it is also necessary to design things such as payment confirmation, approval, and duplication prevention. First, check the quality of list creation and matching.

Can I use it even if the form format is different?

Must be verified in actual format. Decide how to handle not only known forms but also forms from new business partners.

Before consulting ZETTAi

Please provide a sample invoice format, items you would like to read, and any exceptions you would like to see. Regarding ZETTAi's AI system development/contracted development, we can consult with you regarding confirmation after extraction and the scope of integration with existing systems. The scope of individual support will be confirmed at the time of consultation.

Articles you may also want to read