AI can now do most of the boring, routine, repetitive work like data entry, transaction coding, and AP/AR, which raises the question of what exactly the human accountants will be doing. According to several major software vendors featured during the Institute of Management Accounting's Technology Showcase, the answer apparently is reviewing this work to make sure the AI did not make a mistake.
This was a point that was made over and over throughout event, which demoed a set of solutions for corporate accounting professionals. Much of this is because many of the routine processes in this field are very rule- and formula-driven, which enables consistent automation from task to task.
Dana Alhasawi, senior manager for mid-market and commercial solutions with payment solutions provider Ramp, raised this point during her own presentation, saying that the finance function is the domain that stands to benefit the most from AI than any other part of the business. Finance processes are built on learnable, enforceable patterns based on the rich data finance workflows produced during repetitive operational tasks that happen every day, such as coding transactions, matching invoices to peers, chasing receipts, closing the books month after month after month. This, she said, is especially important due to the high stakes of the finance function.
"We know finance is a zero error culture. You can't afford to have an AI agent making decisions in the dark. So everything we built, every AI powered workflow, has full audit trails, complete visibility into what the system did and why, and human control at every meaningful decision point. AI is applied selectively. You decide where it acts autonomously and where it escalates the compliance requirements that govern your function don't go away," she said.
Within finance specifically, she said, accounts payable processes stand as one of the most concrete examples of how finance works with AI. Laura Van Lenten, Ramp's lead solutions consultant for the mid-market, demonstrated the solution's ability to read invoices with optical character recognition to pull things like invoice numbers and dates, and automatically code approvals and optimize payments along the user's specifications. It can do this not only for invoices the customer currently has but for new ones they have never seen before. Overall, she said, the goal is to create as touchless an experience as possible.
"Our whole goal with everything in Ramp, whether that be on the AP side or the card side, is that every experience should be essentially touchless for you. What is typically an entry process for your teams today should really be more of a review process, taking away that manual work and obviously saving the team hours and hours of punching things in," she said.
With most of the routine work done, users will mostly be reviewing outputs and approving payments; and even here, the AI already flags invoices that demonstrate fraud risks, and even offers recommendations as to whether or not something should be approved based on user activity and data. While people generally review automations now, Ramp can automate even parts of the review itself to the point where even this process is done with little need for intense scrutiny.
"And then as individuals become more and more comfortable with these recommendations. We're seeing teams just coming in here and saying, Yep, these are good to go. No more diving in, scouring through line by line. And again, because it's providing that level of detail, teams are gaining confidence very quickly with the solution and taking hours out of approved time now really quickly before we jump to payments," she said.
A similar sort of message eventually emerged from Asaf Gover, CEO and cofounder of finance workflow automation solution Apprentice. He began by noting that while there are similar needs among all businesses, like document processes, the particulars vary too much for a truly universal solution. He said that it can actually be very complicated to automate finance functions as a result. Even if someone there were an AI that could perfectly stand in for a human, people would still need to explain the specifics of the individual business, including its processes, its data structures, its controls and more.
Yet, difficult to automate does not mean impossible. Gover said it just needs a different approach. Rather than following a chain of rules and procedures, the program instead learns by doing. Users, when they start out, record themselves doing their jobs and provide a voiceover explanation of what they're doing step by step. The system then synthesizes that explanation and, from there, builds unique automations that match it. After that, people can fine tune the automations and provide feedback to the system if it got anything wrong. He noted that, once the system has a process down, it is not run via probabilistic AI but a deterministic system that will produce consistent results every time.
And so while Apprentice does not rely on specific defined rules from the start, it is still intended to automate much of the mundane work. And once again, once this has been established, the human's role will mainly be to review the outputs and supervise the system.
"It could be a very long or very short process, but the user gets to review it step by step. They get the peace of mind that the system actually captured it in its full complexity, and they can trust it," he said. "And the user reviewing the automation is the same user that runs the process, the same user that knows the ins and outs and knows whether the automation got it right."
Noel English, senior solutions consultant with payments and finance platform Bill, similarly noted that with so many processes automated, the human largely becomes a reviewer. Documents, whether scanned physically or digitally accessed, are processed instantly, and then reviewed automatically for things like duplicates or incomplete information. The system will also automatically pick up header-level details via a combination of AI and OCR to input information like vendor name, invoice number, due date, total amount and payment terms. It does similar things for purchase order workflows, the system automatically grabbing the purchase order and filling out the details itself based on user behavior. Once that's all finished, the human mainly needs to look at the information to see if everything is okay.
"So as you can see, pretty much about 90%, maybe 95% of the work is already done here. For me, it's really becoming more of a review process at this point," he said.
Mark Fisher, senior vice president of marketing with Vic.ai, an AP automation solutions provider, made a similar point as Alhasawi in that AP is uniquely positioned for AI automation because it is a very low risk, high reward area where there is already a lot of repetitive manual tasks that AI excels at, which is why so many firms are focused there right now.
"And so, we see AI against AP clearly as one of the best entry points in accounting," he said.
Steph Hartnett, a solutions engineer with Vic.ai, noted that the program predicts the header data that's typically written on the invoice, like a vendor or a total amount, along with less typical information like a department or a location for any purchase order backed invoices, with with 97% accuracy out of the box. She noted that the program even goes through the user's email to find invoices or vendor communications that require attention. Like the other solutions, much of this is automated and the human mostly reviews things.
But Vic.AI, too, does much to automate even the review process. Invoices marked in green are high confidence and so can be set to bypass AP review entirely and automatically route the invoice to the approval stage (which, ultimately, must be decided by a human.) In the end, the human does not need to review everything the AI produces, just whatever requires human attention in the first place, which is not as much as people might think.
"So imagine your AP team is really working on a subset of invoices that you receive. 50% are autonomously processed by the AI, and of the remaining 50% that end up in their queue. Only 20% would require them to make an edit to an AI predicted field. The team can easily navigate to that review queue directly from the dashboard," she said.
And even if a system is not AI-native, they can still possess powerful automation capabilities that, once more, require human review. Jonathan Grimes, technical accounting consultant with finance operations platform FinQuery, noted that his platform automatically scans all the information from invoices that someone would need from an organization perspective.
"So the system is purpose built again to make sure that this process is easy and streamlined for your organization, not to mention, we have all of the additional details for your organization to get more context around your portfolios, if we're talking about your prepayments and your invoices," he said.
The webcast took place in March 12. It was the first technology showcase the IMA hosted.
