When Citizens Bank wanted to import data from the financial statements of its commercial borrowers, it turned to intelligent document processing (IDP) to automate aspects of the process, which previously relied on manual keying. Intelligent document processing leverages artificial intelligence (AI) technologies to read and process documents.

The $176 billion bank already used a leading third-party financial statement software solution, said Vinay Jha, chief data officer at Citizens. But it relied heavily on manual keystrokes to spread those financial statements into the software tool.

“Citizens began to look seriously at both fintechs and other vendors that could automate the financial statement spreading solution in order to reallocate higher paid resources back onto more value-added tasks as well as to improve the efficiency of the [end-to-end] statement spreading function,” Jha told Bank Automation News.

There is no shortage of providers when it comes to IDP solutions, according to the Everest Group, an IT research firm. IDP typically leverages a combination of technologies, such as optical character recognition, machine learning, robotic process automation (RPA) or deep learning technologies to extract data from unstructured forms, such as invoices or contracts.

It’s a crowded space, Everest Research Analyst Anil Vijayan told BAN. The research firm ranks 28 players in the space, with the leaders being ABBYY, AntWorks, Automation Anywhere, IBM, Kofax and WorkFusion. A “major contender,” according to Everest, is IDP vendor Hyperscience

The market has exploded in recent years, said Hyperscience Chief Operating Officer Charlie Newark-French. “With the rise of machine learning and artificial intelligence, banks and other financial services firms can find it challenging to navigate the evolving marketplace offerings and select a solution that adds the most value,” he said.

Citizens didn’t specify which solution it chose, but it opted for a “fully integrated solution that worked with the bank’s existing industry-leading financial statement software,” Jha said. It used a machine learning (ML) component that would learn the preferred approaches of the bank for financial statement spreading over time, Jha said.

Expect a learning curve for IDP

Citizens was an early adopter and found the ML component of the tool “was not yet as plug-and-play as initially highlighted,” Jha said. The tool had to be trained with business rules on how to read all the variations of financial statements sourced through the systems. As the ML component became better trained, the adjustment errors decreased and it became more reliable, he added.

What has the bank learned from deploying IDP?

“No system will be as plug-and-play as desired and each will require ample ramp-up to reach its full potential,” Jha said. “Further, it is important to think about long-term sustainability and consider tools that can learn over time and not be restricted by merely hard-coded mapping fields.”

He added that tools will need to be tweaked to accommodate changing accounting rules, as well as bank statement spreading conventions.

The solution also leveraged robotic process automation (RPA) to increase efficiency and reduce errors in the data transfer process, Jha said. In the end, IDP accelerated the financial statement spreading process, allowing Citizens Bank to underwrite faster, he added.

Market trends

Vendors are trying to address the learning curve Citizens experienced by offering more pre-trained models made specifically for certain document types, like mortgages, Vijayan said. These package solutions reduce the training time and the volume of training data an organization needs to get up and running, he explained. On the other hand, it has led to more niche players specializing in certain document types.

Some IDP providers are also reducing time to deployment by prebuilding integrations into the systems of record, such as enterprise resource planning (ERP) systems. Typically, these pre-built integrations are API-based, Vijayan said.

Finally, the technology is maturing with the machine learning tools becoming better at processing unstructured data, Vijayan added.

Financial institutions will also see low-code/no-code options, cloud-based deployment, and more support for non-Latin script languages, such as Japanese and Mandarin. Many providers have also partnered with RPA vendors — or, like Automation Anywhere, they are RPA vendor themselves — in order to incorporate bots to the process as well, Vijayan said.

Bank Automation News would like to know what our readers think about bots. We invite readers to take part in a short BAN reader survey.