Recently, I’ve been bench testing our IDP tool of choice, ABBYY FlexiCapture/Vantage, against other competitor products. At Spectrum, we’re developing a cloud-native Accounts Payable Automation tool, and as the Product Owner, I wanted to ensure that we are using the optimal tool for extracting data from incoming invoices. Traditionally, we’ve used FlexiCapture, but with a move to cloud-native solutions, Vantage now appears to be a more fitting choice.
One of the key performance indicators (KPIs) our customers frequently discuss is touchless processing. We closely monitor success levels through a detailed dashboarding system. To process an order-based invoice touchlessly, it’s essential first to read its contents accurately and then perform three-way matching. This matching relies on receiving accurate invoice data, making the quality of data extraction critically important.
Spectrum has a particular focus on manufacturing and distribution, industries known for their challenging invoice processes. Their invoices often require unit of measure conversion, surcharge recognition, discount calculations, and more, further complicating the capture process.
To benchmark the systems I was testing, I undertook the following steps:
• Created a sufficiently large sample set of invoices with the required fields for capture.
• Configured the systems with identical master data.
• Began with a standard system configuration.
• Tested the invoices over several batches, noting the accuracy of results and utilising the suggested ‘training’ techniques. These techniques, used by end users, involve point and click operations that update the neural networks and learning models to improve the AI.
• Provided feedback to the vendors and implemented any suggestions.
• Repeated the process.
The results were intriguing to analyse. In addition to evaluating touchless processing potential, I assessed overall accuracy at a field level. I measured the percentage of fields captured correctly before and after training, versus the percentage of invoices that could be processed touchlessly.
There were some contrasting results, with certain systems demonstrating high levels of overall field capture but much lower levels of touchless processing. While a high level of field accuracy is beneficial, if users frequently need to correct errors on nearly every invoice, it doesn’t provide the best user experience.
Ultimately, I confirmed that Vantage was the tool to choose, offering the largest touchless processing potential and notable improvements from point-and-click training.