AI Limitations and Opportunities in Treasury (Valorean Technologies)

Speaker A: Foreign. Welcome to the Treasury Update Podcast, presented by Strategic Treasurer, your source for interesting treasury news, analysis and insights in your

Speaker B: car, at the gym, or wherever you

Speaker A: decide to tune in.

Speaker B: Welcome to the Treasury Update Podcast. This is Craig Jeffrey again, your host. I am joined by Arjun Krishnan. Arjun, welcome to the Treasury Update podcast.

Speaker A: Thank you, Craig. Um, pleasure to be here again.

Speaker B: This is our second podcast together and it's on the same topic. It's Arjun's book Enterprise, AI and Treasury. And you can find information on that in the show notes. You can get that at Amazon. I'll put the link down in the show notes or whatever you're listening to. You can find it there in this episode. We call it AI Limitations and Opportunities. Thinking about limitations and opportunities. Normally, discussions about AI are always super positive, but we're going to talk about some of the limitations and opportunities that Arjun outlines in the book. Now, mostly it's from a chapter or section on limitations, but on the flip side of limitations that always present some opportunities. So we'll try to tease some of those out there. So it's not just a, uh, just the sober reflection or the serious reflection on things. Maybe we can start with this. Arjun is. AI certainly has both hype and promise. Um, there's limitations in strength, and in your book you mention like, five major limitations that you talk about. I would love to talk about all of those through the course of the discussion, but as you talk about them, um, maybe you could talk about, here's the limitation, and here is the sober or wise response to those. Do we accommodate it? Do we avoid it? Do we have a workaround? And the first one is data quality shapes AI quality. This is probably talked about the most, right? You have to have quality data or it's, you know, it's learning on bad data and therefore the output's going to be very much compromised. Maybe you can tell us about data quality.

Speaker A: Um, just a general comment on, you know, on limitations, though. You know, every technology has limitations. Uh, and it's really important to understand clearly because particularly, um, with AI, they can be less visible than traditional software and the impact can be a lot more adverse. And it's important to understand the limitations because really we can mitigate or remove the limitations, right, to ensure that trust and confidence in the system is built in. And that's really important to understand when you're designing a system with respect to data. I mean, you know, data is not a supporting character in the AI treasury story. It's a, it's the foundation on which everything is built. This is not unique to AI. Um, you know, we've all been through treasury transformations, um, enterprise ERP implementations or a, uh, treasury management system deployment or a reporting overhaul. They all depend on the same underlying quality of uh, data. Right. It's absolutely critical. From an AI's perspective. It's even more crucial because any incomplete, biased or inconsistent data is going to start getting reflected in the inference and in the answers on outputs that it provides. And the thing is that the answers may be very, uh, you know, with a traditional system, you know, it's very easy to tell that it's a wrong answer. But AI can produce nuanced and subtly wrong answers because it's reasoning as well, and it's reasoning based on patterns that it's learned. Um, and what basically it has inferred from those algorithms.

Speaker B: It's maybe it's doing an interpolation, selecting something between the ranges or an extrapolation. Now on some of the AI machine learning or pattern detection or the ability to detect something that's anomalous is one of the ways that there's a promise is we can also use that to detect those outliers, those things that shouldn't be considered or should need to be fixed before we're relying on those. Is that a core promise?

Speaker A: Um, yes. In fact, that's one of the nice things with AI. I mean it depends on good data quality for good outputs. But you can actually use AI to improve the quality of your, of your source data. You know, harmonization and um, normalization, correction, synchronization of data. And treasury deals with master data, transactional data market, um, data. There's such a myriad of data sources and you can actually use AI, AI to clean those up and then once the AI is using that data, and I think that's also been the promise of uh, an erp systems like SAP and Oracle, where the data is all fairly standardized and quite a challenge to get it moved from legacy to them. But once it's done, the data, uh, is in fairly good shape. So yes, um, the AI can not only enhance it, um, and it uses that same data to provide better quality, um, output and answers.

Speaker B: Yeah, ah, you hear people talk about get it all into Oracle or get it all into SAP and their companies are always buying other companies. There's always this continual process. So you can't just say once we're all on one system, it'll be fine. It's never that way. So having these tools to help clean the data is going to be, I guess that's the life that we live in. So that's. Number one, data quality shapes AI quality. Number two is quite interesting. AI correlates. It doesn't understand. Can you explain the correlation? But the lack of, um, understanding from a machine.

Speaker A: So we have to understand how AI thinks and how it reasons. Basically, it's based on patterns. It's based on patterns that it's learned. It's based on, uh, statistical patterns, basically makes decisions based on those patterns. It's very easy for it to miss context. It can make confident errors because, yeah, it's very confident, but it could be wrong.

Speaker B: And then you question and it's like, oh, sorry, I forgot about that. And then it corrects it, and sometimes the correction is still wrong, you know,

Speaker A: and it requires judgment. And that's why I keep coming back to, you can't put wisdom into an algorithm. Right. The wisdom comes from treasury practitioners, like, you know, you and your listeners. They're the ones who know the context. They have the judgment, they have the enterprise knowledge to be able to help the AI. So AI enhances your judgment, it doesn't replace it. So we have to understand that limitation. But in understanding that limitation, we can also ensure that we get better responses with the understanding that we always have the ability to, uh, override. Override the, um. What the AI is saying. So it's no longer, you know, should. Should AI make Treasury decisions? It's no longer that question. It's a question of how much to trust it and at what point to override it. That always stays within the domain of the treasury practitioner. That will never change.

Speaker B: Yeah, and you say override. I'm wondering, is it overriding or is it choosing not to accept what it's provided? I guess that's the same thing.

Speaker A: That's what, uh. Yes, that's what I mean. I mean, you don't go in. You don't have to go in and reprogram. Well, you may want to. There's ways to. To, um, um, to. To improve the, the learning. And it could be just improving the data. And, and, and again, it also, um, uh, goes back to what's the nature of the error? I mean, it's something that's very blatant, something that can be fixed. Right. That should be fixed. But if it's a judgment call, then you make the judgment call. You know, you don't. You don't have to explain to the AI, Ah. Why you made that. And typically, it might even say, you do this or that, and then you can say, why would you make that recommendation? And it's going to tell you there'll always be a trail and audit trail and there should be. That's why you have a documentation agent, uh, which will document everything that that's taking place.

Speaker B: So um, yeah, I think that's really, really helpful. You know, sometimes we like to think of activities that can occur as, you know, the expected value. You roll two dice, you're going to get seven. Well you can get two to 12. And so there's this bell shaped curve. Well, it's nice with dice, right? It'll eventually get there but sometimes we're like hey, it's seven. And if you get a two, three or four or something in life, whatever that distribution is, it may be very deleterious or harmful to the company. And so you need to do something else. And so this ability to manage things outside the norm or probabilistic, uh, situations like it may not understand that if you haven't programmed it, that these type of events are very, very harmful. We have to avoid those even if it costs more than would be recommended, you know, over time on average. But it's so harmful we're going to pay for that. We're going to do that. So that's number two. So one was data quality, shapes AI quality. Number two is AI correlates it doesn't understand. Before we get to number three, could you just give our audience, if they haven't listened to the other one, a quick background on uh, your career?

Speaker A: Yes, certainly Craig, currently, um, the CEO and founder of uh, Valorian Technologies, a company that I just started about six months ago. Um, focused on um, building AI powered agentic solutions primarily for enterprise, treasury as well as finance. I'm basically an accountant by profession. Uh, I've spent the past three decades privileged to work with and advise some of the world's most forward looking and dynamic corporations with their enterprise wide treasury systems, um, implementation as well as their operational treasury operational transformation. Prior to founding Valorian, I was a managing director at Ernst and Young, building and leading their SAP treasury practice in the United States for the past four years.

Speaker B: Thanks Arjun. Let's continue with limitations. Uh, for those who have acquired the book, you'll find it on starting on pages 23. Number three is AI struggles with genuinely not novel situations, new situations that it can't account for and has to adapt.

Speaker A: Yeah, and this is an interesting one and again, you know, I, I think it goes back to how well the AI has been trained or how uh, what kind of data the AI is working with. But if you think about the Inferencing capabilities are getting better and better. You know, if you take the example for example of uh, fraud detection, the AI may be able to, to identify most situations, but a truly novel fraud like an ah, email scheme perhaps might not be caught by the AI. So I think it has to be used in a way that it provides value to you. And again, when I think about enterprise AI, Craig, and I'm not sure whether this will really be a problem for uh, enterprise because what you're doing is, I think at least initially you should. No, it's to be able to get the information to tap into the incredibly rich data that the enterprise holds and be able to make meaning out of it. And I think that itself is like, you know, 80% of the battle won because it's going to make your decision making. You know, we talked about judgment and decision making. It's going to help your decision making. So you know, in terms of novel situations, they're going to come up because of the fact that perhaps that your Enterprise has not encountered it. But again if it does come up then you know, you can, it's identified and then you can try and do something about it. But uh, if it, if it's missed by AI, well you know, that's, that there's always that risk, right? There's, there's nothing is 100% foolproof. But again you have to see what's the value versus what you might miss. The uh, 80% that you do get, you wouldn't have got if you didn't have this kind of a system. So I think that's the way to kind of look at it.

Speaker B: Seems like then building in some of the safeguards would be putting in. If it falls outside these parameters, let me know, bring me in the loop and defining those to it to the extent that you can.

Speaker A: Right, Exactly.

Speaker B: All right, that's number three. Struggles with genuinely novel situations. Number four is, doesn't seem like this is possible, but it is. AI behavior can vary with inputs. AI behavior can vary with inputs. So how can we depend upon it?

Speaker A: Yeah, I mean again this is a big difference between AI and the traditional systems because of the fact that AI can reason uh, when you put data in, feed it to the AI, it's reasoning, um, across it and even a small change in how you phrase a question might actually produce a significantly different answer. And the system can, can give subtle as well as different answer versus any subtle variations in responses. This is because the behavior, the algorithm is stochastic, right? It is based on variability. It's not a deterministic kind of, uh, algorithm. So it will need a different approach to testing and validation as compared to traditional software. It does need to be tested thoroughly. And that's uh, some of the proof of concepts that I've been working on. You know, you build multiple, multiple checks and balances for each sub agent and um, make sure that the boss orchestrator agent is uh, keeping everybody in check.

Speaker B: I like it. You know, when we talk about probability and before we had AI, we would run, you know, models would run like, here's these variables, here's the range that interest rates can be or general economic growth can be. And you factor that into a model and then you say, well, what will that look like based upon the probabilities and the ranges that exist? And then the outcome, maybe it's some type of cash flow. And you run those and it plots them out on the chart. And is that, would you say that is how I mostly operates? Or is that a way that you can have AI help determine things like probability or confidence in forecasts or needs for cash flow or.

Speaker A: I think the real value, uh, would come from looking at implementing enterprise AI as a process, right? End to end process. I think that's really where the value comes in. And if you look at any process and I think we go back to the spectrum and see, you know, where exactly do you sit in that spectrum? Where will the value add come in? You know, where will AI be able to help you in a discrete process or in an end to end process? So it's not so much about the output or a specific report, it's what value you want to get out of it and in what format. And that could be in the form of a report, it could be a dashboard, it could be a plotted graph, or it could just be a language, a conversation. Right, Using a natural language. So I think it's important to understand your processes and then see where your pain points are and then assess, can AI help that process? Will it help to add value? Will it save time? Will it inform? Will it allow better judgment? Will it allow better decisions? And these are the things that treasury practitioners have to think through when, um, implementing AI. It's clearly, this is AI, enterprise AI, it's not automation, rebranded. Right. Um, but the question which, I mean I struggled, it was one of the reasons I wrote the book is if it's everything that it's cracked up to be, how can I use it in, in my functions, in my treasury functions? And if I'm going to do it, how do I get started and that's kind of one of the primary drivers for me writing the book because I struggled with it and you know, I certainly don't have all the answers. But I think if we all collectively ask the right questions, the tools are available and I think we'll get to the answers.

Speaker B: Right. That's number four. AI behavior can vary with inputs. And number five of five on the limitations is explainability remains challenging. So that's partly back to the black box. And so maybe you could take us through that one.

Speaker A: Yes, this is really important when it comes to enterprise AI, and I think we talked about this earlier, um, in the previous podcast, is that we are the stewards, the stewardship function, we are the custodial function. Right? Controls, governance and good data is critical for Treasury. So when it comes to systems and automating, especially AI systems, you can have uh, complex AI models running in a black box or a monolithic system. The problem with that is of course that you know, there will be no audit trails. Explainability becomes a real issue and you, you really cannot have that because you know, from a governance, regulatory or um, internal audit perspective and treasury covers all three. It's one function that is, is very um, you know, highly, um, highly controlled because the stakes are so high in Treasury. So in order for the explainability it's absolutely critical that we are able to explain to an auditor or to a regulator why the AI made a certain decision because why did it recommend this hedge portfolio? On what basis did it do it? Did it just come up with it or was it based on compliance? Hopefully it's based on the compliance policy that the treasurer has laid down. It'll faithfully should reflect that. So we should be able to explain that. We should be able to document everything that the system is doing and if something goes wrong, we need to be able to pinpoint exactly where it went wrong. That's what auditors look for, that's what um, um, regulators look for. Um, and this is important and I think from an audit perspective, I remember when ERP systems were coming in, auditors would still think about, oh, these were the manual controls. They were not able to shift the locus uh, of their thinking into an enterprise systems based controls. Same issue is going to come with AI, but the more you have more information you have to provide to them to give them that assurance that um, you know, then that's the key. Explainability is absolutely critical. I would say the two major limitations which really are relevant for enterprise AI for treasurers is the data quality and the explainability, because both of them cut to the heart of what, uh, enterprise AI will do for Treasury.

Speaker B: Arjun, thank you. So that was explainability remains challenging. That sounds very reasonable and, uh, rational, based on everything that we've talked about so far. So, as we wrap up, what are your final thoughts about limitations? What's most important? How do we think about that?

Speaker A: Thank you, Craig. Many of the limitations mentioned, uh, apply, to be frank, more in the general application of AI which have been trained on large data sets in the public domain, think, chatgpt, etc. In the context of enterprise AI built around your organization's processes and data, data quality and explainability are the critical limitations to overcome. Organizations will build closed systems that access and act primarily on their own organizational data and providing solutions to their specific needs. Um, I talked about the sub agent architecture. It's particularly powerful and suited to building safe, reliable enterprise AI solutions. Um, I think the question, Craig, is no longer can AI handle Treasury decisions? The question is, which decisions, under what conditions and with what safeguards? The override, and I've mentioned this before, the override is the most important human AI interaction in any treasury system.

Speaker B: Excellent, Arjun. Thank you so much. We really appreciate it. And for those looking to, uh, examine or get his book on enterprise AI and Treasury, we'll drop a link in the show notes. Thank you so much, Arjun.

Speaker A: Thank you. Thank you for having me. Really appreciate it. You've reached the end of another episode

Speaker B: of the Treasury Update podcast.

Speaker A: Be sure to follow Strategic Treasurer on LinkedIn. Just search for Strategic Treasurer.

Speaker B: This podcast is provided for informational purposes only, and statements made by Strategic Treasurer LLC on this podcast are not intended as legal, business consulting or tax advice.

Speaker A: For more information, Visit and bookmark StrategicTreasurer.com.