But these are search companies, and their needs aren’t the same as other enterprises. For example, the amount spent on the AI capabilities of natural speech processing is likely to be much more at Google and Baidu than at GE or Johnson & Johnson. So in order to gain some perspective of AI, Robotic Process Automation (RPA), and machine learning in commercial enterprises, we spent some time talking to Eric Wilson, vice president of purchase-to-pay at Basware (www.basware.com). A leading provider of purchase-to-pay (P2P) and e-invoicing solutions, the Finland-based company has its North American headquarters in Stamford, Conn.
MACHINE INTELLIGENCE
Wilson outlined the progress made over four generations in invoice processing for Basware’s P2P. The first generation was essentially computational, using manual coding and matching. At $15/invoice and $25/PO, that process had an efficiency rating of 80% on-time. That started in the mid-1980s. The second generation dealt better with exception handling as new rules and coding improved the computer’s performance to $5-$10/invoice and $18/PO, 95% on-time. By the third generation (starting approximately five to seven years ago), the AP automation has become a process that includes predictive analytics, machine learning, and RPA. The results now can reach ≤$4/invoice, ≤$10/PO, and ≥99% on-time. A planned fourth generation will use RPA to completely automate many aspects of P2P, leaving only judgment-based decisions for the people monitoring the process. Going forward, the company plans to incorporate more AI and machine-learning capabilities into its solutions to reach that goal.
Looking over the three decades of the evolution of AI at Basware, it’s obvious that this isn’t just a recent development, and the accelerating efficiencies in RPA and machine learning are bringing greater savings and accuracy.
Wilson pointed out that exception handling has reached a 90% touchless level today. He estimated that the last 10% will be addressed with a likely 98% accomplished in leading purchasing and payables organizations in the next three to five years. And he raised an important point about one essential element of AI—the databases used by the AI functions. The deeper the data, the smarter the AI system. It’s all about the database, he explained. A fully automated process is only attainable when Big Data is combined with machine learning. And that data includes more than digits on spreadsheets. For instance, invoice processing now includes machine learning using previous invoices combined with language and country identification in order to apply country-specific business rules.
A kind of microcosm of the automation process can be seen at the level of exception handling. At first generation, the exceptions, which might include suspect entries or first-time, unrecognized accounts, would be flagged for human attention by AP clerks. By the third generation, RPA is handling exceptions with more efficient rules integrating related tasks with data validation and synchronization across multiple systems. The goal of fourth generation will be to leverage RPA and AI to get as close to fully automated processing as the exceptions will allow.
There’s an expanding role for AI at Basware. The company provides a number of solutions within its holistic purchase-to-pay offering beyond procurement, AP, invoicing, and financing. These include the mobile-designed travel and expense management as well as CloudScan, which scans and imports PDF, email, and image invoices. Wilson says that all of the solutions, including expense management and CloudScan, are possible fields for incorporating new AI functionality.
Finally, we discussed a new species of vulnerability of AI due to its increased autonomy. When programs are writing their own code, as can happen in unsupervised and reinforcement learning modes, there could be a new kind of vulnerability beyond those already exploited by hackers. Wilson explained that those doing risk assessment for development are now incorporating these kinds of risks as part of their responsibilities.
The changes brought to financial software by AI and RPA already have been substantive and extensive, and they have been at work longer than most of us realize. In the future, they will likely become even more pervasive and powerful.
April 2017
