We want to automate our Accounts Payable, but many invoices are scanned PDFs with varied layouts. Standard OCR isn't accurate enough to extract line items. Has anyone successfully integrated AI Center or Document Understanding models into their RPA flows? How do you handle the "Human-in-the-loop" requirement for low-confidence extractions?
3 answers
Document Understanding (DU) is far superior to standard OCR because it uses Machine Learning to "understand" the context of a document rather than just reading text. For your invoice project, you’ll need a Taxonomy to define fields like Invoice Number, Date, and Line Items. You can use the "Intelligent Form Extractor" for structured data and the "ML Extractor" for semi-structured invoices. For the "Human-in-the-loop" part, use the Action Center. When the bot encounters a confidence score below your threshold (e.g., 80%), it creates a task in Action Center. A human then validates the data, and once submitted, the bot resumes the workflow to post the data to your ERP.
What is the average volume of invoices you process monthly? The licensing for AI units can get expensive if your volume is very high.
We used Amazon Textract via an API call within our Blue Prism bot and found it much more accurate for line-item extraction than native OCR.
Interesting approach, Linda! We actually found that using the "Validation Station" in DU helped our ML model "learn" from human corrections over time, improving accuracy.
Steven, we do about 5,000 invoices. Is that cost-prohibitive for DU? At 5,000, you are in the sweet spot where the ROI usually justifies the AI Unit cost. To save money, use "Public Endpoints" for common documents like invoices instead of training your own model from scratch. Only move to a custom-trained model in AI Center if your invoices are highly specialized or in a language that the pre-trained models don't support well.