Using a New Technology to Solve a Very Old Problem

Identifying a document by type and assigning it to a shipment. Rotating a document scanned at the wrong angle. Responding to requests for documents in order to get invoices approved. Entering data from documents to use in the invoicing process. Removing duplicate copies of the same scanned document. Viewing a document to see if there is a signature on a form. Comparing data on documents to data saved in a computer system to identify discrepancies.

All of these document-related tasks are things we do every day at Daycos while billing and collecting over 400,000 transportation invoices annually on behalf of our customers. And for years, all of these tasks were done manually by Daycos employees. However, thanks to new technological capabilities like machine learning, we have now successfully automated these tasks.

Over the years, we’ve invested heavily in all kinds of technologies to improve document handling. Dedicated fax lines, bar code cover sheets, rooms full of employees scanning, and optical character recognition (OCR)––we have tried it all. And while we saw improvements in our processes, they were painfully slow to arrive. Then we started utilizing machine learning technology and have seen substantial improvements in our document handling process.

While there is a mysterious impression of machine learning as “computers magically learning on their own,” it is really just simple technology at work. When given a large amount of data, a computer can learn to recognize patterns in that data, and then make predictions based on those patterns. For example, in the area of household goods documentation, if you give the computer hundreds of thousands of examples of weight tickets, the computer can eventually look at a document and tell you if it thinks it is a weight ticket or not. And while this is relatively simple to understand, it is not easy to implement. It has taken us many attempts and refinements to get machine learning programs to be able to make accurate predictions, but the payoff has been substantial.

At Daycos, we use machine learning technology in a number of ways to make the invoicing and payment process run more efficiently. When a billing documentation packet is received from a customer, we can identify every document type in the packet, clean up the documents so they are easier to view, remove duplicate copies, re-order the documents to the preferred order, and extract necessary data from certain forms––all before a human even looks at it.

Machine learning technology can be utilized in the collections process as well. If we receive a message that we need to provide billing documents before the invoice can be approved, machine learning technology can help us identify and respond to this situation. Over the years, we have received and collected thousands of messages requesting documentation, so we use those to train the machine learning program to read messages and predict if the message is a request for documentation. If the program is confident in its prediction, then we can automatically send the billing documents off without a human ever being involved.

This technology brings the obvious benefits of being able to handle a higher volume of invoices without additional people. During a recent 24-hour period, our document system processed over 81,000 images, performing various tasks that in the past would have to have been done by employees.

In addition to that increase in capacity, there are other benefits to utilizing machine learning technology in our processes. It speeds up the billing and collections process tremendously, and in the invoicing world, time is money. Even the most efficient humans can’t compete with the speed of a computer doing these tasks. And it also allows us to identify problems much earlier in the billing process. If we know we need certain documents to bill a shipment, our machine learning technology can tell us within minutes of receipt of a billing documentation packet what is missing. Then we can immediately notify the customer of the missing document instead of waiting for it to go through our billing process to be identified, reducing the payment cycle by days.

This technology also allows us to scale our document handling to meet the peak season surge. A human-driven document process can be easy to run efficiently in the winter slow season, but when peak season hits, the handling of paperwork can quickly become a bottleneck. Using technology allows us to keep those processes moving quickly no matter the volume. We can spin up more computer capacity from our servers when the volume requires it.

Using technology to improve processes sounds great when everything works, but the real test is what happens when that technology fails. How can you guarantee accuracy when computers are making decisions that could cause significant problems if they are wrong? This is another area where machine learning provides some real advantages.

When the machine learning program is making a prediction, it also provides a level of confidence in its prediction. We use this confidence level to determine when and where to trust the prediction. If it is a critical element and would cause significant problems if the prediction is wrong, then anything less than 99% confidence gets flagged to be reviewed by an employee. However, if it is not as important, and a mistaken prediction will be caught by an employee in the next step in the process, then we may lower the acceptable confidence level to 80% before we send it for further review.

While the goal of moving towards true computer-to-computer data exchange is getting closer every day, it appears that paper documentation will still be an important part of getting paid for a while. Our development in machine learning has allowed us to gain the advantages of a true computer-to-computer exchange even in today’s document-driven environment. We are turning documents into data and building machine learning around the data, which will prepare Daycos to handle whatever the future holds for transportation payments.