JAIC Partners with DIU on AI/ML Models to Resolve Complex Financial Errors
- By: JAIC Public Affairs
The Joint Artificial Intelligence Center’s Business Process Transformation Mission Initiative in partnership with the Defense Innovation Unit is launching a project that will build “intelligent” or “cognitive” automation prototypes to resolve complicated financial and accounting errors without any human intervention.
“This is, as far as I know, the first example of the Department of Defense actually interfacing a machine learning model with automated data systems that can start making judgments about financial transactions,” explained Rachael Martin, Mission Director for the BPT MI. “We’re hoping to prove that auto-ML software capability will be effective in interfacing with Robotic Process Automation (RPA) software, providing an additional level of complexity for resolving challenges with DoD financial data.”
DoD financial communities are already incorporating various types of automation solutions for routine financial management tasks to streamline operations, reduce errors, and save time.
The Deputy Assistant Secretary of the Army for Financial Operations and Information (DASA-FOI), for example, leveraged automation to eliminate nearly 30,000 labor hours in FY19 and expects to free up more than 100,000 labor hours in FY20. Meanwhile, the Office of the Undersecretary of Defense (Comptroller) has stood up an Automation Shared Service within the Advana platform that allows DoD personnel to share automations and discover cross-agency efficiencies in scaling automations.
Automations are currently used to resolve a range of financial errors and process financial data for a range of audit requirements. Unfortunately, standard RPA automation is limited in its capabilities, according to Martin, as it cannot handle the audit category known as an unmatched transaction. RPA and other types of automation are too logical to resolve this type of financial error because it’s either too complex or it doesn’t fit into a well-defined workflow. “When that happens, a human has to step in and resolve the error,” Martin said.
Unmatched transactions can take anywhere from a day to weeks to months to resolve and they are a major headache for financial managers. Currently, the Army estimates that though these type of errors only account for less than 2 percent of the DoD’s financial transactions, it translates into billions of dollars in unresolved financial activity.
“Right now, there is a backlog of a year or so of these unmatched transactions that DoD personnel can’t get through,” Martin said.
This is where the new AI/ML models come in. To fix as many of these unmatched transactions as possible without any need for manual aid, the JAIC’s BPT team is working with the OUSD (Comptroller) and the Army’s ODASA-FOI to develop an AI/ML model prototypes that deliver intelligent or cognitive automation.
“While regular automation follows logical, well-defined workflows, intelligent or cognitive automation is able to take irregular, complex financial information and data, make decisions, and apply judgments and solutions to a high level of accuracy or confidence without having to have a human in the loop,” Martin said.
To actually build these models, the JAIC solicited the support of the DIU to help curate and identify commercial solutions for these problems. More than 50 vendors provided solution briefs through the Commercial Solution Opening and DIU’s program managers and technical experts worked with DoD partners to downselect to the finalists. Within four months, two contracts were awarded to two vendors Vertosoft and Summit2Sea—as part of its Humanless Unmatched Transactions (HUnT) product line.
“As AI companies move from Robotic Process Automation to machine learning, the DoD can leverage these solutions to make their financial systems more efficient providing program manager support on this effort,” said Jeff Klugman, Director of the DIU AI portfolio. “We expect these vendors to supply working prototypes for field trials with the goal of saving the DoD labor hours per year in finding and correcting unmatched transactions.”
Each company will integrate auto-ML and AI software with existing RPA infrastructure, but one will work with the ODASA-FOI’s General Funds Enterprise Business System, while the other will work with the Defense Agencies Initiative, a general ledger and primary accounting enterprise resource planning system used by 24 Defense agencies and managed by OUSD (Comptroller).
“These vendors are building the same capability on two different systems and two different sets of data, but what we’re going to do is let the two vendors compete to see which one comes up with a better solution,” Klugman said. “And if they both provide great solutions, that’s still fantastic because that would give us an expanded realm of choice for other agencies interested in pursuing an AI/ML solution for similar use cases.”
The HUnT models will be trained to first categorize the unmatched transactions as either simple or complex. A second phase will focus on training them to triage the transactions, such as which transactions are the easiest to tackle, which are the most important ones, which ones can be sent back to the RPA automation with direction on how to fix, and which ones are so complex that they still require human interaction to resolve.
If successful, the HUnT models will enable a number of long-term benefits for the DoD financial community and the larger mission. These include:
- Time savings.
- Cost savings.
- Greater regulatory and fiscal compliance.
- Increased efficiency.
- The ability to maintain more accurate accounting of where DoD dollars are located and where and how they are being spent.
In fact, the impact on the DoD and its ability to field new AI capabilities will be wide-ranging as the overarching software capability that’s developed will be built out into Advana, an OSD community platform for DoD users/developers. This will allow for agile development and experimentation with new AI capabilities from a variety of partners suited for the requirements.
“This is definitely one of our most important proofs of concept,” Martin said. “It’s a capability that right now people are trying to figure out how to incorporate and leverage, so this is a good opportunity to be able to help provide that.”