Q&A with Eric Wilson, Basware’s VP of P2P: Can We Teach Machines to Get It Right 99% of the Time?

NLP agsandrew/Adobe Stock

In Part 1 of our conversation with Eric Wilson, VP of P2P at Basware, we dug into the links between data and AI — is data the be-all, end-all behind AI advancements? What does one do with the data once they get it? And how best to manage it?

“Capturing all the data is only one part of it,” Wilson noted. “The second thing is to actually make use of it.”

In this second installment, we get a bit beyond the data collection and management, and more into how context and eventual wisdom play into it — and whether those are even truly attainable.

Spend Matters: Let’s say we’ve figured out how best to store and manage all these disparate pieces of buy-side and sell-side data. How are we then going to learn from this data fast enough to derive anything useful?

Eric Wilson: That's a good question. You and I, as human beings, are [not] going to do it, to be blunt. That is where we really take advantage of the power of artificial intelligence. You don't have the ability to crunch a million rows of data in a split second and then predict that you “should not issue this purchase order right now” — but the machine does. That's how we will use that data: by utilizing what is already available, honestly, for providers like Google and Amazon and a bunch of others plugging those technologies into a transactional process. Whether we're talking about within purchase-to-pay or CRM or ERP, plug it into the contextual process to actually make changes to how you're acting in real time. That's how we will make use of that data.

SM: You mentioned the contextual process. How does one teach a computer to understand context as well as human analysts can, to bolster the AI function? Is that something that’s being thought about?

EW: Good point. There's a whole world of technologies out there that take different flavors of what broadly is known as artificial intelligence. Some of that is cognitive where they're truly taking into account the context of the environment and of that particular customer or user and they're learning over time. They're not just saying, "Okay, I can crunch a bunch of data and therefore spit out a predefined set of solutions. I'm crunching a bunch of data but I'm also learning from what happened last time to be able to alter my prediction as a machine.”

SM: So context and knowledge of other factors that are not so tangibly documented or called upon is getting built in, to result in progress toward something like “wisdom,” then?

EW: Exactly. That's the difference between advanced process automation and the true value of machine learning. If the machine is not actually improving and learning — not just from the data, but from the particular variables of your organization — then we're just talking about process automation, right? We've been doing that for a long time. With actual AI, that machine is truly learning from the behavior of all the users in the system from the context of the organization.

SM: Are you reasonably confident that machine learning is improving to the point of achieving 99% accuracy? In other words, as accurate as human experts can be in “getting it right”? I’ve heard that it’s difficult to auto-classify spend data to even 95% much of the time.

EW: Machines will very soon — and very accurately — replace the tactical transactional activities that humans carry out in the world of purchase-to-pay today. For example, a standard transactional activity that has to happen is when an invoice comes in and doesn't match the purchase order or somebody has to go in and say, "Okay. I'm still going to approve that invoice or I'm not going to approve that invoice." To do that, they're looking at several things. They're calling their colleague down the hall and saying, "Hey, is this okay?" Machines are going to get very good at taking care of that activity very quickly. What that's going to do is free up humans to then add more value to the organization by doing higher-level tasks. It'll be an iterative process where machines get more and more accurate at more and more difficult tasks. But that initial level of taking the transactional elements out of the hands of humans, that's going to happen very, very soon. It actually already is.

SM: But about those higher-level tasks — do you foresee a specific timetable for machines to overtake even the high-value tasks? Or are we not quite there yet?

EW: It depends on what level of sophistication we are talking about. It will certainly rapidly get more sophisticated than the example I just gave a moment ago. Here’s another example of what an analyst might do or say today in the sourcing space. They're looking at all the spend across the organization in a particular category, let’s say MRO. They're looking at a number of MRO suppliers and they're looking at competitive information about the pricing that some of their competitors are getting on MRO and also just kind of scouring the Internet for other publicly available data on pricing and they decide, "Okay, you know what, it's time to go out and resource our MRO category, send it out to bid and see if we can get better pricing."

One hundred times out of 100, that’s a person that makes that decision today. It will not be a really long time before a machine can do that pretty well also. You're looking maybe three years down the road, that particular activity likely is going to be run by a machine on a pretty regular basis. Not 100 times out of 100, but a pretty regular basis.

SM: Fair enough. How do you quantify the risk of a machine making mistakes, versus a human making the same mistakes?

EW: Certainly there are risks that a machine is going to make a mistake. As you said, there's a risk that humans are going to make a mistake as well. It's incumbent upon us, the solution providers, to make sure that in those situations that might have more significant risks associated with them, machines are likely not going to actually execute the decision. The machines are going to serve up the probability that this is the correct choice and then a human ultimately would say, “Yes, go,” or “No, I'm going to look at it myself.” Over time, the probability of error on that particular activity gets lower and lower. Certainly that's a great question and something that providers need to be taking into account when they are applying machine learning to the day-to-day transactions in their systems.

Eric Wilson will be speaking at Procurious’ Big Ideas Summit this week.

Discuss this:

Your email address will not be published. Required fields are marked *