Artificial intelligence (AI) is set to play an increasingly significant role in commodity trading over the next decade. However, while AI can enhance trading strategies and improve efficiency, human oversight and expertise remain crucial. In this episode of The Trading Desk, experts from Oliver Wyman and Veritas explore the role of AI in trading, how organizations can prepare for the AI journey, and what senior leadership should consider in order to maintain a competitive edge.

Interested in a particular topic? Skip to the following sections:

01:05 — How AI is different to analytics and can be leveraged in trading

02:52 — Where trading executives should think about deploying AI in their business

05:58 — How AI is being applied and whether it will replace traders

10:20 — Leveraging AI to build a competitive edge

12:10 — The business opportunities that can be achieved through leveraging AI and what organizations must do to help it flourish

16:39 — The intelligence aspect of AI and whether it can go rogue

19:16 — Who will succeed in the AI race and reap the benefits first

25:10 — Advice to C-suite and whether AI is worth the ROI

28:45 — The organizations that are adopting AI better than others and how they're winning in this space

Joanne Salih  

Hi, I'm Joanne Salih, a partner in our energy and natural resources practice, and the head of our Americas risk and trading business. I'm joined today by Christian Lins, a partner in our energy and natural resources practice, and also our risk and trading business, as well as Brad Kyer, a director in our Veritas business, which is a specialist technical arm of our risk and trading organization. Thank you so much for joining me today, guys, and today we'll be discussing the role of AI in the future of trading. So I'm always suspicious of these buzzwords and phrases that get bandied around, especially with systems. I'm just very, very old fashioned and I'm always thinking to myself, I mean honestly, when it comes to AI, is this really going to be something that we can leverage in trading? Is it something which is measurably different to analytics?  

Brad Kyer

Yes, AI is definitely something we can leverage in trading. It's a high-order view of analytics in the sense that it brings together an extreme amount of data, processes it and can look for signals in that data that may not be evident when looking at pieces of data individually. Yes, there are ways with statistics to do this, but AI has a more natural way of dealing with data that comes from the natural world, to process, unlike statistics, which you have a much more rigorous framework of data in the way you have to present it for it to be processed.  

Christian Lins

Clearly in the trading organizations we are serving, AI - as in advanced analytics, machine learning, neural networks and the likes - have been around for many, many years, decades. I think what's new this time around is that the generative capabilities that the latest generation really brings to the table opens up a host of new questions, and that's really new that the cost of training has come down and that all of a sudden you can answer and explore really explorative open-ended questions and you don't need to respond to rule-based, hard-coded, essentially requirements. So that's the novelty this time around.  

Joanne

So we always think about things in trading as a question of productivity, like what's my cost income ratio? How am I going to effectively beat the rest of the market? Where are there opportunities that I can pursue that potentially give me an edge over others? And with AI, I get what you guys are telling me. The question obviously that a lot of our clients ask us, and in turn we ask ourselves, is where are we going to deploy this first? What does this mean? Is it a front office capability? Are we actually going to replace traders? Or are we thinking about this more as a efficiency item - something that we're going to look at in risk or in the back-office processes that we run?  

Christian

Yeah, that's a very good question and I think many folks are still trying to figure it out, quite frankly. I think what we see is that AI that really uses numbers and produces numbers as outputs is a very structured, relatively "stable" problem statements that works well. And then in many asset classes or commodities, that's the norm. If you're required to respond within seconds, then no human is going to play in that space. Now if you're getting loads of different kinds of inputs, different languages, different formats, different meanings, different units, it's getting much, much more domain specific and that's where we see quite some interest in the generative capabilities to really shape and redesign some of these processes. 

Joanne 

Yeah, so is this a process question for you, Brad, or is it something else?  

Brad 

 Well, it's a combination. It's a process question. It's understanding where the best leverage can be applied. The front office for many decades has been using various forms of AI. Middle office has been using it in some forms. Back office, however, this is where interactions with counterparties, interactions with clients comes in multiple formats - domain-specific data sets - historically has been very manual or very programmatic in the sense you needed a large team of people to interact with that data so the back office could send out confirmations or invoicing, and the AGI and these new forms can process that, convert it into something that's much more friendly for the back office to operate. The leverage factor is much higher. There is a knock-on requirement where you have to train and manage and care and feed for this new feature. You're getting the ability to deal with the data faster, but somebody has to keep the quality control in check. And that's where I think your leverage factor is higher, but now you're kind of changing the domain in which your employees are operating.  

Joanne 

Yeah, so Christian, does that mean we can finally get rid of those pesky traders?  

Christian 

That's a challenging question or a difficult one. I think that the truth is in parts, we are already there. We have different types of trading, for instance in power in certain markets where the actual execution of the trade, the analysis -numbers in, numbers out - is already done by machines today. And it's simply because the velocity of the decision making is beyond what humans can do. Now, the type of AI that everybody's talking about these days is not necessarily that kind of AI. It's the more generative AI that can cope with more complex, more explorative types of problem statements where you might need to ask a question or try to find the true meaning of something. And that's something where people are scratching their heads at the moment.  

Joanne 

Yeah, so I mean honestly, we would've thought about this for systems like maybe 20 years ago, analytics like 10 to five years ago. But it is fundamentally now a competitive advantage and I think a definition of whether you're going to win or not in the next five to 10 years.  

Brad 

We are seeing AI being applied, again, in power trading, in power systems analytics - it's looking at huge amounts of data and we're not just talking about what's the price anymore, we're talking about the price, how much is flowing, what's the temperature, what availability do we have for capacity on the lines? It's balancing all of those things which humans can do. However, it's a very long-term skill that it takes to learn where AI can learn this quickly, get up to speed, can do the balancing. You can get your very fast high frequency trading out of this today without having to have a very expensive workforce to get there of humans, who are very talented, but very few humans do these today. We see this, again, in power. We see this in natural gas - we see managing pipelines across storage, looking at the holistic view of the portfolio and making sure that counterparties and collateral, all of that is being managed in a trade all at once from the top down. And today we have to build systems for humans to make sure that they can do the 'what if' trade - will I create a collateral event? Will I need to put more margin up?  These things I have to make sure this is okay. AI is doing that now. We're capable of doing that today within a single thought process. This is where we start getting into a situation where we're looking at a macro kind of trading environment with one process instead of we have five or six actors all acting together as humans. And that's where I think we're getting to. We're seeing that now. We're seeing that more and more. Systematic trading is getting further and further looking at huge amounts of data across interest rates. It's looking across the whole portfolio of let's say the larger corporation. This is going on. The level to which it's going on depends on how far down the road and how high quality the data is that the trading firm has gotten to. And that's where I think some have a very long, long runway to get there and some started there or that was their focus from day one.  

Joanne 

I think that's a really interesting point because I have to say back to the point around it being a buzzword, a lot of the questions we get asked is really how do I implement AI into my current environment? And for me, my starting point is do you have the data architecture to even be having this discussion and do you have the basic systems in place and organizationally, do you understand what this is really going to mean to you? So I think honestly, looking at the trading environment right now, there are those who may be far more able to pursue this than others, and there's a roadmap to getting there, but it's not going to be overnight.

Christian 

Yeah, I mean the leading players in the space have obviously top management buy-in, but they also have chief data scientists or chief digital transformation officers or whatever you want to call it, CXX, that benchmarked themselves against big tech that realize that physical commodity trading is essentially mostly about logistics. And they also realize that this is a moving target. So it's about how do you manage a workflow with a human and a synthetic workforce and which interfaces can be machine to machine? Which activities, and even decisions, could potentially be given to a machine? And you made a point around data that's very important, but also what we're seeing is that there's significant effort being put into training and even customizing, and it's almost humans teaching the synthetic workforce. It's almost like training, in parts, training a successor that might outdo you in the future.  

Joanne 

So in layman's terms though, Christian - might outdo you if you're very good, but also could be terrible if you don't have great organization.  

Christian 

Absolutely, absolutely. And that's why there will be, you can call it new overhead to the whole AI story and completely new roles for humans.  

Joanne 

I mean, what's really interesting for me is there's obviously significant challenges with all of this. There's also a lot of opportunity, and you mentioned power earlier, and I think a lot of people's minds go to power because of the pace that we're talking about. But I always think about, okay, well when we think about the more conventional side of things, you mentioned pipelines and fundamentally logistics infrastructure, the whole world of information basically whether it's weather or it's the activities of your competitors, and also just low carbon opportunities as well. So I'm beginning to think about, okay, well if you're not present in a market, does AI actually become the lever by which you begin to explore and look at opportunities rather than setting up a full-fledged trading organization to go and pursue a specific opportunity?

Brad 

Well, I think it's a bit dangerous to think AI is going to give you the magic answer to go and make money in the market without having properly identified the questions you're trying to ask and how you plan on interacting with the AI to get to an answer that you could execute. AI is definitely going to help, but AI, it is not a silver bullet. It's not magic. It's not going to say, okay, you do this, you're going to make a million dollars tomorrow. If you cannot, from the top down, build an environment around AI that it can flourish, you're going to get out of it what little you put into it is what little you get out or what garbage you put in is what garbage you're going to get out. So even as a general example in today's generative AI, if you were to ask it, the difference between a dressing and a sauce, it'll tell you a sauce goes on food and a dressing is a bandage that you use to cover wounds. And so again, if you give it very little context for it to operate, it's going to give you back what it implies as a context. And this is where properly trained people and processes behind it are necessary to get you the results you're looking for. I mean, it is not a silver bullet. We have used it before for monitoring prices along pipelines and looking at storage facilities and temperature deviations from the norm and then looking at physical and financial fundamental data altogether. And we've been able to track prices through the pipelines fairly accurately. But in that case, we had a lot of validated use cases. We were testing it, we were making sure that it was doing what we expected it to do and not running off on a wild tangent. And these are things you learn. It's like, okay, well if I didn't ask it not to do that, it may actually do that. And this is kind of the whole process behind it. And this is where training your people becomes paramount if you want to actually get results.  

Christian 

That is a really important point. So ultimately in the limit, if you solve technology, it's human adoption that will make it work or not. And if you are a manager of a support function, pick logistics again, and you're having parts of your team replaced by AI agents, that's a big thing because you're going to get KPIs, they're going to cost you something and they need to perform. And in many instances, if you're inheriting this, these are black boxes - about as black as they get. So how do you understand what do you actually get, how they perform, and how can you control that? So it is the human trust that needs to be ultimately also factored into this transformation. Otherwise you're just going to have island solutions and you're not going to see the big savings or productivity gains or the scalability of the platform that everybody is going after here.  

Joanne 

In a way, it is not quite the same thing obviously, but it's kind of like when you take on a cargo and it's the traceability of where that cargo came from. Where does the buck stop? And as we know, there's a lot of litigation in that space. So it feels like in a way a new problem. But one we've kind of known in the past, we know the situation. So can AI go rogue? Are we going to see the rise of AI and they're going to take over and do a bunch of dodgy things?  

Brad 

Well, I think we have to be careful about what we call rogue. I think individual bots can obviously go rogue if they're not properly trained and validated. They can make decisions that are completely erroneous or outside the rules that humans are aware of. As a synthetic workforce, we have to think about as a collection of these agents operating and performing human-like tasks, can the collection as a whole go rogue? It's quite possible. If you're feeding garbage through the system, you may not get out what you're expecting. I don't think it's going to launch nuclear missiles at anyone anytime soon. But there is a good chance that your AI system could start trading or interacting with counterparties in ways that violate laws or basically lose you money left and right. No different than high frequency, just simple probabilistic trading can do today. We've seen that in the past where bad inputs and it just starts lifting and hitting bids all over the place and the trading firm goes under in a matter of an hour. Same thing can happen here. And again, this comes back to how do you build, manage, and maintain and whether or not you build systems that are monitoring the system. We have this for humans. We have compliance, we have audit that are monitoring what the traders are doing. We have to have a similar thing on the AI side that says this AI bot is doing something unexpected. We need to shut it down before the whole system gets corrupted somehow in a way that the business could get shut down.

Joanne 

Yeah, exactly. For me, it's like that intelligence aspect of AI, which is truly that you're building something which in itself is intelligent and has the potential to undertake certain activities, but it's also a product of the environment you've created for it. So there is this notion of going rogue or whatever other phrase we want it to use, but similarly, if it's in an environment which is actually highly control dominated and has very low limits, and it is basically telling the traders continuously, look, this is the structure you need to sit within, and it's quite conservative. We don't want to take lots of risks. The AI will also inherit that. It won't suddenly be this kind of highly productive engine that butts against the machine because the machine is what it's learning from, I suppose.  

Christian

So reflecting on who's going to succeed and reap the benefits first, it's clear that you need a minimum size to venture into this, because developing and governing an AI agent requires a minimum workflow and there are economies of scale in deploying that, right? Even if it's just to think about maritime and terrestrial logistics, you want to have a certain flow there or B2B type businesses. You want to have a certain number of customers there. So, scale requirements are going to be critical for the early adopters. You obviously want management that's reasonably forward thinking. You want an environment where - essentially human labor is not infinitely cheap - so you're looking at markets where actually labor costs are relatively high, which I think in general in trading is the case.  

Joanne 

But why is that? Why is the labor cost such a factor?  

Christian 

Because when we see the environment, typically in some trading houses it's always been easier to "throw", allocate people to the problem rather than systematizing it. So, it's actually an old symptom of a bad operating model. 

Joanne 

So you need a mature trading organization.  

Christian 

You need a mature trading organization essentially with something to lose. And that's where we see actually most efforts at the moment.  

Brad 

Places that succeed will definitely be places that have put it upfront. Top down management is thinking through how are we going to systematize? How are we going to get this data controlled? And how are we going to look at processes in place that we have to reduce, not necessarily reduce the number of headcount for humans, but how do we get them oriented in the way so that they can feed the system and we can leverage the system to grow further and to grow faster? This is the whole goal of AI. It's not necessarily a human replacement as much as a leverage factor. You have to go back and think about, well, I'm going to have to retrain or upskill my staff to do this, and there's absolutely zero wrong with that. That is the future. You should upskill. And if you don't, they'll go somewhere else where they'll get upskilled. I mean, that's the only way you can look at that. But that is what's going to happen - you're going to have to upskill the middle office, back office personnel to instead of adding another spreadsheet, how do we train the agents so that we get rid of the proliferation of spreadsheets that occur in the background that become a very manual process that requires extensive human capital and it's very low value, low return on that effort. But again, that's a whole different mindset and way to think about running the business than it is today.  

Joanne 

So it sounds like, I mean, much like many of the questions that we typically battle with in trading, it's rare to say this is a trading only discussion if you're not in an independent trading organization.  

Christian 

No, exactly. And even the independent trading companies are no longer independent, in the original sense of it. They're all starting to look more alike.  

Joanne 

By alike you mean they have larger asset positions?  

Christian 

Absolutely. And one of our hypothesis in this space is that the opportunity here, and that's what I meant earlier by something to lose, is the actual interfaces with the assets that are not necessarily a pure trading problem per se, where you have activities, decisions, very heterogeneous, where similar questions could result in different answers at times. So that understanding, that pyramid, is really going to be key. And that's where we see the fixed costs, that's where we see the leverage that AI can provide.  

Joanne 

You have to fundamentally believe this is going to add something to your system. But even in the way that we work today with the systems that we have today, as simple as the use of an LP and what an LP tells you, to a certain extent, you have to believe it and then say, okay, well this data set seems right and the answers seem as jumping off point and we as a human organization are going to figure out now what to do with this. Or agricultural commodities that are much, much more heterogeneous than your typical hydrocarbon play. So many more factors, so many more stakeholders involved, very complex, very fragmented logistics, a lot of different formats of receipts and notices flowing around. So that's typically where we see the opportunity. And obviously as said earlier, it only makes real sense if you have a large enough system to deploy this. So let me ask you, if you are a CXX of an organization and the powers that be are asking you, do we do this or do we not? Is AI something we need to be investing in? I mean, what are some of your reflections on it?  

Brad 

I would say if you're not at least earnestly looking into this, you're going to probably be left behind. We see today in the systems that there are high leverage capabilities from them in different areas of your business, even in simple things as document processing. I mean, there are significant lever factors, 2-3X, that you can get just by properly putting in place some of these AI agents and these synthetic workforce agents to do this for you. And then you can actually use your existing workforce, move them into the higher ROI parts of your business. And again, there is a cost. You have to rethink your business process, rethink your data models? How does all of this flow through the system? And so if I'm the CXX, I have to think about these things - like today, I may have very fragmented parts of the business. How do I bring them together cohesively so this whole process can flow? And how does the data access rights work in such a way that I can really lever up how much business I can do with the same workforce I have today? It's not about reducing workforce, it's about with the same workforce, how do I leverage that up into much more business, more income, more revenue, and that's the way I need to be thinking about it. Yes, certain jobs will probably vaporize, but they are going to transform into a completely different role. And that's the way we need to think about it, because everyone else is thinking about this.  

Joanne 

So kind of invest today to save tomorrow.  

Brad 

Yeah. Yeah. I mean it truly is. You have to invest today. You're going to have to rethink your business in a way that you can facilitate this. But as you go through this one, you learn where the weak spots in your business are. There's a lot of businesses today that have siloed environments, and it's very hard to get information out. That has to go away. And then on top of it, you can still, with AI agents, maintain your controls over data who can see what, where, how, because the agent doesn't have to tell you where the data came from and what the actual data is. It can just tell you that here's a result, right? Today you may have complex systems in place to manage who can see what and how they all interact with different humans because of compliance and audit controls. Well, you can still implement those with your AI agents without having to expose that to an outside world. And so again, it simplifies that whole flow process, but it is a lot of work. I'm not going to lie.  

Christian 

I think there are different stakeholders here as well beyond the usual suspects. So one that I like to highlight typically is also human capital and not just because of the human adoption that is needed, but also that in order to really drive impact in a more corporate setting where it's difficult to move people around or make them redundant, there's also an element of how do I actually transform my workforce to be able to cope with this new environment? And that will be the pace setting, especially with the more corporate trading analysis.  

Joanne 

So it sounds like your CXX is going to spend a lot of time talking to a lot of other people to get this right. So maybe then it's like a question of, well, who do we truly believe is going to do this better than others? Are there organizations that today, for example, are conducting algorithmic trading who are pretty lean, I think from a systems angle, either have completely built out their own systems architecture or have what we believe to be leading class capabilities. Are those the guys that are going to be winning - guys and girls -that are going to be winning?  

Brad 

So I think in today's world, the corporations that are winning, especially in this algorithmic space, they are winning with more probabilistic trading. They do have AI-based trading. We have to be somewhat cognizant that the high frequency trading is nanosecond, sub millisecond space of decision-making processes. And that really comes down to simple math in most cases of a decision - do I put hit or lift? AI is not there yet at that level of performance. It's making great strides every month to get closer and closer and closer, but it is not there yet. So in those sort of systematic trading where they are trading on the sub milliseconds, no, but everyone is working towards it. That is coming. Nobody's going to tell you when it's there, it's just going to be there. And then everyone who's not there will be left behind and playing catch up. But today's AI probably is still in the subsecond range on response times for a fairly decent response using a fairly large amount of complex data to get to those sort of decision processes, which is not - in some cases, that's fine - there are certain commodity trading that really trades in minutes, and this would be a fine example where that could fit in. But in the systematic trading where it is sub second or sub millisecond, it's still probably a couple years out.  

Christian 

And ultimately, I think what you're describing are really proprietary agents. So it's a tailor-made suit, but also what we're going to be seeing here is an ecosystem of off-the-rack suits, where they fit well for certain types of occasions. And these occasions are not in the sub milliseconds, and many of them are more middle, back-office type or front office support type activities. And that's where we see a myriad of efforts also with certain players in actually structuring this and understanding how they can cope with agents that might be 50% third party sourced, but then 50% internally trained or entirely externally sourced, or for certain applications really based on internal proprietary data and a minimalistic external shell. So that's going to be a challenge. And again, the governance of this is something that requires human oversight.  

Joanne 

Just reflecting back guys, AI isn't really a question of technology. It's obvious technology is a fundamental enabler, but it becomes a question of people, of regulation, and compliance. It's a question of organizational structures and adoption. And so it's new, but as we've always said, the challenges are challenges we've always faced in the commodity trading space. But if there's an organization or a capability that can potentially get it right, given the strive for value, I would argue this is probably, like you said, Brad, I think this is the place to do it if it's going to be anywhere.  

Christian 

Absolutely.

This transcript has been edited for clarity

Artificial intelligence (AI) is set to play an increasingly significant role in commodity trading over the next decade. However, while AI can enhance trading strategies and improve efficiency, human oversight and expertise remain crucial. In this episode of The Trading Desk, experts from Oliver Wyman and Veritas explore the role of AI in trading, how organizations can prepare for the AI journey, and what senior leadership should consider in order to maintain a competitive edge.

Interested in a particular topic? Skip to the following sections:

01:05 — How AI is different to analytics and can be leveraged in trading

02:52 — Where trading executives should think about deploying AI in their business

05:58 — How AI is being applied and whether it will replace traders

10:20 — Leveraging AI to build a competitive edge

12:10 — The business opportunities that can be achieved through leveraging AI and what organizations must do to help it flourish

16:39 — The intelligence aspect of AI and whether it can go rogue

19:16 — Who will succeed in the AI race and reap the benefits first

25:10 — Advice to C-suite and whether AI is worth the ROI

28:45 — The organizations that are adopting AI better than others and how they're winning in this space

Joanne Salih  

Hi, I'm Joanne Salih, a partner in our energy and natural resources practice, and the head of our Americas risk and trading business. I'm joined today by Christian Lins, a partner in our energy and natural resources practice, and also our risk and trading business, as well as Brad Kyer, a director in our Veritas business, which is a specialist technical arm of our risk and trading organization. Thank you so much for joining me today, guys, and today we'll be discussing the role of AI in the future of trading. So I'm always suspicious of these buzzwords and phrases that get bandied around, especially with systems. I'm just very, very old fashioned and I'm always thinking to myself, I mean honestly, when it comes to AI, is this really going to be something that we can leverage in trading? Is it something which is measurably different to analytics?  

Brad Kyer

Yes, AI is definitely something we can leverage in trading. It's a high-order view of analytics in the sense that it brings together an extreme amount of data, processes it and can look for signals in that data that may not be evident when looking at pieces of data individually. Yes, there are ways with statistics to do this, but AI has a more natural way of dealing with data that comes from the natural world, to process, unlike statistics, which you have a much more rigorous framework of data in the way you have to present it for it to be processed.  

Christian Lins

Clearly in the trading organizations we are serving, AI - as in advanced analytics, machine learning, neural networks and the likes - have been around for many, many years, decades. I think what's new this time around is that the generative capabilities that the latest generation really brings to the table opens up a host of new questions, and that's really new that the cost of training has come down and that all of a sudden you can answer and explore really explorative open-ended questions and you don't need to respond to rule-based, hard-coded, essentially requirements. So that's the novelty this time around.  

Joanne

So we always think about things in trading as a question of productivity, like what's my cost income ratio? How am I going to effectively beat the rest of the market? Where are there opportunities that I can pursue that potentially give me an edge over others? And with AI, I get what you guys are telling me. The question obviously that a lot of our clients ask us, and in turn we ask ourselves, is where are we going to deploy this first? What does this mean? Is it a front office capability? Are we actually going to replace traders? Or are we thinking about this more as a efficiency item - something that we're going to look at in risk or in the back-office processes that we run?  

Christian

Yeah, that's a very good question and I think many folks are still trying to figure it out, quite frankly. I think what we see is that AI that really uses numbers and produces numbers as outputs is a very structured, relatively "stable" problem statements that works well. And then in many asset classes or commodities, that's the norm. If you're required to respond within seconds, then no human is going to play in that space. Now if you're getting loads of different kinds of inputs, different languages, different formats, different meanings, different units, it's getting much, much more domain specific and that's where we see quite some interest in the generative capabilities to really shape and redesign some of these processes. 

Joanne 

Yeah, so is this a process question for you, Brad, or is it something else?  

Brad 

 Well, it's a combination. It's a process question. It's understanding where the best leverage can be applied. The front office for many decades has been using various forms of AI. Middle office has been using it in some forms. Back office, however, this is where interactions with counterparties, interactions with clients comes in multiple formats - domain-specific data sets - historically has been very manual or very programmatic in the sense you needed a large team of people to interact with that data so the back office could send out confirmations or invoicing, and the AGI and these new forms can process that, convert it into something that's much more friendly for the back office to operate. The leverage factor is much higher. There is a knock-on requirement where you have to train and manage and care and feed for this new feature. You're getting the ability to deal with the data faster, but somebody has to keep the quality control in check. And that's where I think your leverage factor is higher, but now you're kind of changing the domain in which your employees are operating.  

Joanne 

Yeah, so Christian, does that mean we can finally get rid of those pesky traders?  

Christian 

That's a challenging question or a difficult one. I think that the truth is in parts, we are already there. We have different types of trading, for instance in power in certain markets where the actual execution of the trade, the analysis -numbers in, numbers out - is already done by machines today. And it's simply because the velocity of the decision making is beyond what humans can do. Now, the type of AI that everybody's talking about these days is not necessarily that kind of AI. It's the more generative AI that can cope with more complex, more explorative types of problem statements where you might need to ask a question or try to find the true meaning of something. And that's something where people are scratching their heads at the moment.  

Joanne 

Yeah, so I mean honestly, we would've thought about this for systems like maybe 20 years ago, analytics like 10 to five years ago. But it is fundamentally now a competitive advantage and I think a definition of whether you're going to win or not in the next five to 10 years.  

Brad 

We are seeing AI being applied, again, in power trading, in power systems analytics - it's looking at huge amounts of data and we're not just talking about what's the price anymore, we're talking about the price, how much is flowing, what's the temperature, what availability do we have for capacity on the lines? It's balancing all of those things which humans can do. However, it's a very long-term skill that it takes to learn where AI can learn this quickly, get up to speed, can do the balancing. You can get your very fast high frequency trading out of this today without having to have a very expensive workforce to get there of humans, who are very talented, but very few humans do these today. We see this, again, in power. We see this in natural gas - we see managing pipelines across storage, looking at the holistic view of the portfolio and making sure that counterparties and collateral, all of that is being managed in a trade all at once from the top down. And today we have to build systems for humans to make sure that they can do the 'what if' trade - will I create a collateral event? Will I need to put more margin up?  These things I have to make sure this is okay. AI is doing that now. We're capable of doing that today within a single thought process. This is where we start getting into a situation where we're looking at a macro kind of trading environment with one process instead of we have five or six actors all acting together as humans. And that's where I think we're getting to. We're seeing that now. We're seeing that more and more. Systematic trading is getting further and further looking at huge amounts of data across interest rates. It's looking across the whole portfolio of let's say the larger corporation. This is going on. The level to which it's going on depends on how far down the road and how high quality the data is that the trading firm has gotten to. And that's where I think some have a very long, long runway to get there and some started there or that was their focus from day one.  

Joanne 

I think that's a really interesting point because I have to say back to the point around it being a buzzword, a lot of the questions we get asked is really how do I implement AI into my current environment? And for me, my starting point is do you have the data architecture to even be having this discussion and do you have the basic systems in place and organizationally, do you understand what this is really going to mean to you? So I think honestly, looking at the trading environment right now, there are those who may be far more able to pursue this than others, and there's a roadmap to getting there, but it's not going to be overnight.

Christian 

Yeah, I mean the leading players in the space have obviously top management buy-in, but they also have chief data scientists or chief digital transformation officers or whatever you want to call it, CXX, that benchmarked themselves against big tech that realize that physical commodity trading is essentially mostly about logistics. And they also realize that this is a moving target. So it's about how do you manage a workflow with a human and a synthetic workforce and which interfaces can be machine to machine? Which activities, and even decisions, could potentially be given to a machine? And you made a point around data that's very important, but also what we're seeing is that there's significant effort being put into training and even customizing, and it's almost humans teaching the synthetic workforce. It's almost like training, in parts, training a successor that might outdo you in the future.  

Joanne 

So in layman's terms though, Christian - might outdo you if you're very good, but also could be terrible if you don't have great organization.  

Christian 

Absolutely, absolutely. And that's why there will be, you can call it new overhead to the whole AI story and completely new roles for humans.  

Joanne 

I mean, what's really interesting for me is there's obviously significant challenges with all of this. There's also a lot of opportunity, and you mentioned power earlier, and I think a lot of people's minds go to power because of the pace that we're talking about. But I always think about, okay, well when we think about the more conventional side of things, you mentioned pipelines and fundamentally logistics infrastructure, the whole world of information basically whether it's weather or it's the activities of your competitors, and also just low carbon opportunities as well. So I'm beginning to think about, okay, well if you're not present in a market, does AI actually become the lever by which you begin to explore and look at opportunities rather than setting up a full-fledged trading organization to go and pursue a specific opportunity?

Brad 

Well, I think it's a bit dangerous to think AI is going to give you the magic answer to go and make money in the market without having properly identified the questions you're trying to ask and how you plan on interacting with the AI to get to an answer that you could execute. AI is definitely going to help, but AI, it is not a silver bullet. It's not magic. It's not going to say, okay, you do this, you're going to make a million dollars tomorrow. If you cannot, from the top down, build an environment around AI that it can flourish, you're going to get out of it what little you put into it is what little you get out or what garbage you put in is what garbage you're going to get out. So even as a general example in today's generative AI, if you were to ask it, the difference between a dressing and a sauce, it'll tell you a sauce goes on food and a dressing is a bandage that you use to cover wounds. And so again, if you give it very little context for it to operate, it's going to give you back what it implies as a context. And this is where properly trained people and processes behind it are necessary to get you the results you're looking for. I mean, it is not a silver bullet. We have used it before for monitoring prices along pipelines and looking at storage facilities and temperature deviations from the norm and then looking at physical and financial fundamental data altogether. And we've been able to track prices through the pipelines fairly accurately. But in that case, we had a lot of validated use cases. We were testing it, we were making sure that it was doing what we expected it to do and not running off on a wild tangent. And these are things you learn. It's like, okay, well if I didn't ask it not to do that, it may actually do that. And this is kind of the whole process behind it. And this is where training your people becomes paramount if you want to actually get results.  

Christian 

That is a really important point. So ultimately in the limit, if you solve technology, it's human adoption that will make it work or not. And if you are a manager of a support function, pick logistics again, and you're having parts of your team replaced by AI agents, that's a big thing because you're going to get KPIs, they're going to cost you something and they need to perform. And in many instances, if you're inheriting this, these are black boxes - about as black as they get. So how do you understand what do you actually get, how they perform, and how can you control that? So it is the human trust that needs to be ultimately also factored into this transformation. Otherwise you're just going to have island solutions and you're not going to see the big savings or productivity gains or the scalability of the platform that everybody is going after here.  

Joanne 

In a way, it is not quite the same thing obviously, but it's kind of like when you take on a cargo and it's the traceability of where that cargo came from. Where does the buck stop? And as we know, there's a lot of litigation in that space. So it feels like in a way a new problem. But one we've kind of known in the past, we know the situation. So can AI go rogue? Are we going to see the rise of AI and they're going to take over and do a bunch of dodgy things?  

Brad 

Well, I think we have to be careful about what we call rogue. I think individual bots can obviously go rogue if they're not properly trained and validated. They can make decisions that are completely erroneous or outside the rules that humans are aware of. As a synthetic workforce, we have to think about as a collection of these agents operating and performing human-like tasks, can the collection as a whole go rogue? It's quite possible. If you're feeding garbage through the system, you may not get out what you're expecting. I don't think it's going to launch nuclear missiles at anyone anytime soon. But there is a good chance that your AI system could start trading or interacting with counterparties in ways that violate laws or basically lose you money left and right. No different than high frequency, just simple probabilistic trading can do today. We've seen that in the past where bad inputs and it just starts lifting and hitting bids all over the place and the trading firm goes under in a matter of an hour. Same thing can happen here. And again, this comes back to how do you build, manage, and maintain and whether or not you build systems that are monitoring the system. We have this for humans. We have compliance, we have audit that are monitoring what the traders are doing. We have to have a similar thing on the AI side that says this AI bot is doing something unexpected. We need to shut it down before the whole system gets corrupted somehow in a way that the business could get shut down.

Joanne 

Yeah, exactly. For me, it's like that intelligence aspect of AI, which is truly that you're building something which in itself is intelligent and has the potential to undertake certain activities, but it's also a product of the environment you've created for it. So there is this notion of going rogue or whatever other phrase we want it to use, but similarly, if it's in an environment which is actually highly control dominated and has very low limits, and it is basically telling the traders continuously, look, this is the structure you need to sit within, and it's quite conservative. We don't want to take lots of risks. The AI will also inherit that. It won't suddenly be this kind of highly productive engine that butts against the machine because the machine is what it's learning from, I suppose.  

Christian

So reflecting on who's going to succeed and reap the benefits first, it's clear that you need a minimum size to venture into this, because developing and governing an AI agent requires a minimum workflow and there are economies of scale in deploying that, right? Even if it's just to think about maritime and terrestrial logistics, you want to have a certain flow there or B2B type businesses. You want to have a certain number of customers there. So, scale requirements are going to be critical for the early adopters. You obviously want management that's reasonably forward thinking. You want an environment where - essentially human labor is not infinitely cheap - so you're looking at markets where actually labor costs are relatively high, which I think in general in trading is the case.  

Joanne 

But why is that? Why is the labor cost such a factor?  

Christian 

Because when we see the environment, typically in some trading houses it's always been easier to "throw", allocate people to the problem rather than systematizing it. So, it's actually an old symptom of a bad operating model. 

Joanne 

So you need a mature trading organization.  

Christian 

You need a mature trading organization essentially with something to lose. And that's where we see actually most efforts at the moment.  

Brad 

Places that succeed will definitely be places that have put it upfront. Top down management is thinking through how are we going to systematize? How are we going to get this data controlled? And how are we going to look at processes in place that we have to reduce, not necessarily reduce the number of headcount for humans, but how do we get them oriented in the way so that they can feed the system and we can leverage the system to grow further and to grow faster? This is the whole goal of AI. It's not necessarily a human replacement as much as a leverage factor. You have to go back and think about, well, I'm going to have to retrain or upskill my staff to do this, and there's absolutely zero wrong with that. That is the future. You should upskill. And if you don't, they'll go somewhere else where they'll get upskilled. I mean, that's the only way you can look at that. But that is what's going to happen - you're going to have to upskill the middle office, back office personnel to instead of adding another spreadsheet, how do we train the agents so that we get rid of the proliferation of spreadsheets that occur in the background that become a very manual process that requires extensive human capital and it's very low value, low return on that effort. But again, that's a whole different mindset and way to think about running the business than it is today.  

Joanne 

So it sounds like, I mean, much like many of the questions that we typically battle with in trading, it's rare to say this is a trading only discussion if you're not in an independent trading organization.  

Christian 

No, exactly. And even the independent trading companies are no longer independent, in the original sense of it. They're all starting to look more alike.  

Joanne 

By alike you mean they have larger asset positions?  

Christian 

Absolutely. And one of our hypothesis in this space is that the opportunity here, and that's what I meant earlier by something to lose, is the actual interfaces with the assets that are not necessarily a pure trading problem per se, where you have activities, decisions, very heterogeneous, where similar questions could result in different answers at times. So that understanding, that pyramid, is really going to be key. And that's where we see the fixed costs, that's where we see the leverage that AI can provide.  

Joanne 

You have to fundamentally believe this is going to add something to your system. But even in the way that we work today with the systems that we have today, as simple as the use of an LP and what an LP tells you, to a certain extent, you have to believe it and then say, okay, well this data set seems right and the answers seem as jumping off point and we as a human organization are going to figure out now what to do with this. Or agricultural commodities that are much, much more heterogeneous than your typical hydrocarbon play. So many more factors, so many more stakeholders involved, very complex, very fragmented logistics, a lot of different formats of receipts and notices flowing around. So that's typically where we see the opportunity. And obviously as said earlier, it only makes real sense if you have a large enough system to deploy this. So let me ask you, if you are a CXX of an organization and the powers that be are asking you, do we do this or do we not? Is AI something we need to be investing in? I mean, what are some of your reflections on it?  

Brad 

I would say if you're not at least earnestly looking into this, you're going to probably be left behind. We see today in the systems that there are high leverage capabilities from them in different areas of your business, even in simple things as document processing. I mean, there are significant lever factors, 2-3X, that you can get just by properly putting in place some of these AI agents and these synthetic workforce agents to do this for you. And then you can actually use your existing workforce, move them into the higher ROI parts of your business. And again, there is a cost. You have to rethink your business process, rethink your data models? How does all of this flow through the system? And so if I'm the CXX, I have to think about these things - like today, I may have very fragmented parts of the business. How do I bring them together cohesively so this whole process can flow? And how does the data access rights work in such a way that I can really lever up how much business I can do with the same workforce I have today? It's not about reducing workforce, it's about with the same workforce, how do I leverage that up into much more business, more income, more revenue, and that's the way I need to be thinking about it. Yes, certain jobs will probably vaporize, but they are going to transform into a completely different role. And that's the way we need to think about it, because everyone else is thinking about this.  

Joanne 

So kind of invest today to save tomorrow.  

Brad 

Yeah. Yeah. I mean it truly is. You have to invest today. You're going to have to rethink your business in a way that you can facilitate this. But as you go through this one, you learn where the weak spots in your business are. There's a lot of businesses today that have siloed environments, and it's very hard to get information out. That has to go away. And then on top of it, you can still, with AI agents, maintain your controls over data who can see what, where, how, because the agent doesn't have to tell you where the data came from and what the actual data is. It can just tell you that here's a result, right? Today you may have complex systems in place to manage who can see what and how they all interact with different humans because of compliance and audit controls. Well, you can still implement those with your AI agents without having to expose that to an outside world. And so again, it simplifies that whole flow process, but it is a lot of work. I'm not going to lie.  

Christian 

I think there are different stakeholders here as well beyond the usual suspects. So one that I like to highlight typically is also human capital and not just because of the human adoption that is needed, but also that in order to really drive impact in a more corporate setting where it's difficult to move people around or make them redundant, there's also an element of how do I actually transform my workforce to be able to cope with this new environment? And that will be the pace setting, especially with the more corporate trading analysis.  

Joanne 

So it sounds like your CXX is going to spend a lot of time talking to a lot of other people to get this right. So maybe then it's like a question of, well, who do we truly believe is going to do this better than others? Are there organizations that today, for example, are conducting algorithmic trading who are pretty lean, I think from a systems angle, either have completely built out their own systems architecture or have what we believe to be leading class capabilities. Are those the guys that are going to be winning - guys and girls -that are going to be winning?  

Brad 

So I think in today's world, the corporations that are winning, especially in this algorithmic space, they are winning with more probabilistic trading. They do have AI-based trading. We have to be somewhat cognizant that the high frequency trading is nanosecond, sub millisecond space of decision-making processes. And that really comes down to simple math in most cases of a decision - do I put hit or lift? AI is not there yet at that level of performance. It's making great strides every month to get closer and closer and closer, but it is not there yet. So in those sort of systematic trading where they are trading on the sub milliseconds, no, but everyone is working towards it. That is coming. Nobody's going to tell you when it's there, it's just going to be there. And then everyone who's not there will be left behind and playing catch up. But today's AI probably is still in the subsecond range on response times for a fairly decent response using a fairly large amount of complex data to get to those sort of decision processes, which is not - in some cases, that's fine - there are certain commodity trading that really trades in minutes, and this would be a fine example where that could fit in. But in the systematic trading where it is sub second or sub millisecond, it's still probably a couple years out.  

Christian 

And ultimately, I think what you're describing are really proprietary agents. So it's a tailor-made suit, but also what we're going to be seeing here is an ecosystem of off-the-rack suits, where they fit well for certain types of occasions. And these occasions are not in the sub milliseconds, and many of them are more middle, back-office type or front office support type activities. And that's where we see a myriad of efforts also with certain players in actually structuring this and understanding how they can cope with agents that might be 50% third party sourced, but then 50% internally trained or entirely externally sourced, or for certain applications really based on internal proprietary data and a minimalistic external shell. So that's going to be a challenge. And again, the governance of this is something that requires human oversight.  

Joanne 

Just reflecting back guys, AI isn't really a question of technology. It's obvious technology is a fundamental enabler, but it becomes a question of people, of regulation, and compliance. It's a question of organizational structures and adoption. And so it's new, but as we've always said, the challenges are challenges we've always faced in the commodity trading space. But if there's an organization or a capability that can potentially get it right, given the strive for value, I would argue this is probably, like you said, Brad, I think this is the place to do it if it's going to be anywhere.  

Christian 

Absolutely.

This transcript has been edited for clarity