Our current bots can only handle structured data, but we receive thousands of unstructured PDF invoices every month. How do we effectively add Optical Character Recognition (OCR) and Machine Learning to our RPA processes to handle these complex documents without manual intervention?
3 answers
This is the transition from RPA to Intelligent Automation. You should look into "Document Understanding" frameworks. Most major RPA vendors have native AI engines now. For example, UiPath Document Understanding uses a hybrid approach of OCR to read the text and ML models to understand the context (e.g., finding the 'Total Amount' regardless of where it is on the page). Start by training a model on at least 50-100 of your specific invoice types. You will also need a "Human-in-the-Loop" stage initially, where a human validates the bot's confidence levels until the model becomes sufficiently accurate.
What is the average quality of your PDF scans? If the image resolution is low, even the best AI models will struggle. Have you tested different OCR engines like Tesseract versus Google Cloud Vision?
You can use Python scripts within your RPA workflow to call specialized AI APIs. This is often cheaper than using the vendor's "all-in-one" AI packages if you have the coding skills.
That's a great point, Christopher. Using an API for Azure Form Recognizer can be a very cost-effective way to add high-level intelligence to a standard bot.
Most of our invoices are digital-born PDFs, but about 10% are low-res scans from older vendors. I haven't compared OCR engines yet. Does Google Cloud Vision handle handwritten notes better than the built-in RPA engines? We sometimes get manual annotations on the invoices.