Meet Moody’s AI Agents

Speaker A: By downloading or listening to this podcast, you are agreeing to Moody's legal terms and conditions found@moody's.com disclaimer, including that the information provided is not investment or financial advice and that Moody's will not be liable for losses arising, uh, from your use of the information.

Speaker B: This is Risk Reframed, the show, where we bring you the leading risk experts from Moody's, our partners and the wider ecosystem to discuss the risks organizations face, how they're interconnected, and ultimately what you, the risk professional, can do about them. Every conversation we have on Risk Reframe these days at, uh, some point comes onto AI agents and what comes next. Like most businesses in this space, Moody's has been busy working with customers on what to build and launch in terms of agents. And we're going to unpack some of that today and meet some of Moody's agents themselves. We'll do that with my colleagues, Olga and Keith Berry. Uh, Olga, I'm going to try to pronounce your full name because I'm notoriously bad at pronunciation. Uh, we have Olga Luasau Aslanidi.

Speaker A: Almost there.

Speaker B: Almost. Olga Luazo Aslanidi, you're the managing director for AI Solutions here at Moody's in the corporates and governments unit. And Keith, you're our general manager, so the one running the show, as it were. Although I know you don't like to take credit for too much.

Speaker C: That's right. Yep.

Speaker B: Very good. First things first, we've got an office move going on, which happens over the next week and a half. Um, and you guys are trying to find each other in our current sort of partial floor earlier. Did you manage to find each other before the show?

Speaker C: We did. We did, yeah.

Speaker B: Very good. I was, uh, sending you in opposite directions, and I think, um. And, Keith, the other thing I thought that was quite interesting was the last time you were on, we were talking about the launch of the, uh, sort of KYC franchise that you led at Moody's. As that grew, um, and since then, the world's changed quite a lot, technology wise. Um, we were trying to figure out whether it was before or after the initial OpenAI public launch. What did we come to? What were our dates?

Speaker C: I think it was after it. So I guess November 2022 was ChatGPT.

Speaker B: Right.

Speaker C: And the world changed. I, um, think, um, we spoke sometime

Speaker B: early in 23, something like that. When we release versus when we record is always. Yeah, I never quite remember, but exactly.

Speaker C: So. So I think. I think the world had started to change, but probably we weren't fully aware of it yet.

Speaker B: Yeah, we knew we could ask this funny website some questions and make it say funny things back exo get it

Speaker C: to come up with poems and rhymes

Speaker B: and stuff like that. I do actually remember one of, uh, another former guest, Rupert used to do, um, or maybe he still does, but he'd feed a few facts about people and do like a birthday poem or whatever if there was a team meeting and the occasion called for it. So simpler uh, times, some might say. I mentioned in the intro there we're talking agents these days. And Olga, as I said, you're leading this for us. Um, everyone says they have them or are using them, but what are they if, uh, we want to sort of sort the week from the chaff.

Speaker A: Yeah. So we started with a chat, right? Um, what you were saying, asking question and getting answer. What agent is, is really delegating work that you want to get done. So really giving it an objective, giving it a set of tools, giving it really a way how to do things, setting m guardrails and then get things done. So that's really what what agent is. So it still uses the large language model like ChatGPT as Brain, but then it can actually do stuff, not just tell us, okay, go and do this.

Speaker B: Okay, so if the LLM, um, is the brain.

Speaker A: Yeah.

Speaker B: You're saying there are other parts of the body presumably to do this.

Speaker A: Tools would be hands, then there would be some connectors and then agent would uh, would uh, do work to achieve objective that we give it.

Speaker B: Okay, so these are the component parts that are the how it works. Right, Exactly. And a lot of people say harness and I don't know, sometimes I understand it and I forget it and I hear it again. How does the harness work?

Speaker A: Yeah, so why people started uh, talking about harness is because when you just look at large language model alone, it's almost never enough for the thing that you want to do. So you want to bring the whole ecosystem or if you want harness. So if you think about large language model as your cpu, harness is kind of operating system around it. So when you give an agent, uh, tools, when you give connection through for example what is called model context protocols, when you give it access to uh, Data to your APIs, you do evaluations. So all these components of the ecosystem represent hardness.

Speaker B: Okay. So it's a nice sort of collective term for all of it that's making it work.

Speaker A: Exactly. So it's really agent is a large language model plus harness to get things done properly that you can trust Understood.

Speaker B: Bit of vocab there for everyone. So everyone make the notes. Keep that in mind as we go through today. Keith, you've been a technologist your whole career. I don't know if you like that term or not, but you understand technology at a level that I certainly don't. Is this just the current thing, right? Like three years? Ish. Three and a half years. Ish. And, uh, maybe a little bit more. But it's the Gartner hype cycle people talk about. Is this just the next one, or is it bigger than that in your view?

Speaker C: I mean, I think this is a fundamental kind of technology shift, like we saw with the Internet or with maybe cell phones and smartphones coming along. M. Um, and there's been kind of multiple steps to this, but I think if you go to really November, December of 2025, so we're in midway through 2026 now. That's when we really started to see these agents, particularly in coding, getting incredibly powerful. So a lot of people talk about going home for Christmas or the holidays, uh, December 2025, and starting to realize that these agents are, uh, writing code as good as a good engineer. Right. And that's been a combination of the harnesses getting better, as Olga's talked about, but also the fundamental models getting better and their ability to do things and then applying that to other domains, such as the work we do in KYC and M. That same approach. But this really does feel like the start of something fundamentally different, uh, in terms of the ability to automate things that historically might have taken people days, weeks and hours and get them done in, uh, those things done in a fraction of the time.

Speaker B: Okay. So we now have better tools, more mature models, and we're going faster. Everyone said the world was already moving too fast, but we're going faster. More gas is the solution. Olga, when you think about having this toolkit now, when you're building. Because it's not about building a product, right. It's about building a solution for customer so they can buy it and benefit from it. How do you think about bringing agentic capabilities into the suite of solutions you're putting out there? Uh, what's your sort of mental model for how, um, and when and why to use it?

Speaker A: Yeah, so it's really about building solutions because it's not about applying the cool technology, which is also fun. Um, but it's really solving real problems. So it's really either saving people time to do things or doing things more at scale. For example, producing various reports that would be you know, taking days or months sometimes to produce when you need to get data from different places. So it's really solving real problems for real use cases. So when we build agents, first thing we're asking, does this problem even need an agent? Uh, perhaps it's an uh, automated workflow. Or if it does need an agent, what exactly this agent is supposed to do? Then we're thinking about the harness around that agent. What harness do we need to build in order to solve that particular problem or a use case? And to give you example, for example in kyc, when we need to do screening for a company, we need to clear many different tasks that uh, analysts would uh, need to do. So for example check uh, uh, company names or owner name and see are there any false positive matches. So really to clear this faster and also to have human analysts focusing only on those important things where really you need to spend time and dig deeper while agent can do the rest of them and say this is likely much, um, or this is likely true positive. But then still as a human analyst we decide do we want to accept what agent suggested, um, or we don't. So it's really solving problem like that, writing report or going through those um, uh, for example screening um, uh, tasks.

Speaker B: Okay, that's interesting because I think a lot of us will have as consumers even right like kind of experience everything. They're not now having a chatbot or seemingly and sort of oh, it's the same service. But now I have this extra bit of friction to deal with because they want me to engage with something with AI. But you're saying it's not an add on, it's uh, you start at the fundamental question of what is the problem and then does an agent actually need to be used here or could it do a better job than our old terms? Is that right?

Speaker A: Exactly. So it's really not about sprinkling AI on top of the solution, but it's really embedding it where it makes sense. Automate what needs to be automated, bring some uh, judgment level when it needs to be escalated to human. And really uh, having this ability to pass, um, think about it as either you delegate your work to agents, uh, or uh, you do it yourself and then engage agent only in um, in the places where it makes sense. But in the end of the day it's us human decide uh, what can be delegated, where agent starts, where human jumps in and where agent can actually take over. So it's really embedding it into the work that needs to get done.

Speaker B: Interesting. I know it's not just about kyc, but because we've done a lot of episodes around that topic. I remember we did, uh, KYC Officer of the Future several years ago, and we talked about this idea of the risk architect M. And your job won't be to work cases. Your job should be to set up your system and then design around that. Well, maybe it wasn't after ChatGPT was released, but before I understood any of it at all. So feels like that. But now we've actually got a toolkit to get there. Whereas I was just thinking, look, just get your right sources and get your right APIs and put them in the right order and you're good. Maybe that was naive and now it's a little bit more, uh, possible.

Speaker A: Now it's definitely more possible because you could actually start the work simply by talking to what seems like a chatbot. But what's behind is actually the whole coordination of different workflows, agents grabbing the tools, the APIs, as you mentioned, and then bringing them in the right place at the right time. Um, and importantly, bring also all the trusted data underneath. Uh, because this is what's, uh, obviously very important for, um, trusted results of what the agent does.

Speaker B: Definitely come onto that. Um, Kiva, I was going to just ask you, sort of your section from the leadership side and you and the leadership team, how have you sort of thought through, when do we want to direct the builders in the Org to go about pursuing agents versus where we specifically don't want them to? Is that a decision that you've thought about, or is it more letting them decide that for themselves?

Speaker C: So I think it's a bit of a balance. Right. Um, I mean, we've been talking internally about kind of type one organizations and type two organizations, and I think there's a set of organizations that are moving very fast in this space, and then there's organizations and look, a lot of the work we do at Moody's is with regulated companies. And there may be different regulatory regimes around the world that are more cautious around the use of AI. Other areas are more positive and almost encouraging, uh, the use of AI. So you've got kind of different pressures on these different organizations, but we do have some very early adopters who want to move very fast and sort of have agents at the heart of everything they're doing. We're definitely seeing that at this point in time. Um, and then we've got firms who feel less comfortable with that because of regulatory or their risk posture itself, and, uh, are moving slower and those are kind of the type 2 who are doing the more traditional, uh, approaches. But I think the two things come together. So very much what we're trying to do is embed, you know, it's a bit of a, you know, meet customers where they are. Right. And people always say that. But, uh, in this context it could be where if you don't want to leverage the AI, you can still use our products without. But we do believe there's benefit and having features and capabilities we can turn on that'll allow you to do that and you'll see the benefit is key.

Speaker B: Understood. So it's not a, uh, build these ones, don't build these ones. It's more understand customer base, customer demand, external forces, and then build appropriately for different points of that journey.

Speaker C: Yeah, I mean, the job to be done I don't think is changing the way you might achieve that job might be different. So I mean, maybe to explain a bit of the story behind the screening agent, um, I mean that came out of a customer who were a, ah, crypto firm who were moving very fast. And we really started with that back in pretty early, back in 2023, I think, uh, towards the end of 2023, uh, with actually a discussion over dinner with uh, their chief compliance officer who was talking about the spikes they were seeing from actually all the sanctions that were coming out with the Russia, Ukraine war, um, and how that was really impacting their analysts. And their analysts were having to work all sorts of crazy shifts because they couldn't forecast the peaks and troughs in demand. And after a discussion with him at dinner, I asked him, well, when we give you an alert because of a new sanction, what do you do with it? And he actually shared his playbook for what his human analysts do. Um, and I went back to our very, uh, early kind of machine learning AI team and said, could we automate this? And we kind of partnered with that organization to actually build a pretty early, to be honest, agent that took that playbook and automated it. I think the other interesting thing that brought to the table was, you know, when we talk about what's different here is the description of the rules that you're giving the agent is often in natural language.

Speaker B: Right.

Speaker C: So, for example, a common thing in the world of KYC and screening is to look at the distance, uh, you know, the geographical distance between the person who's maybe applying for an account and where you're seeing an alert from somebody in a risk database. And traditionally, you know, the chief compliance officer would have written that in a ah, policy document would have handed the policy document to his engineers. His engineers would maybe have then ah, worked with a vendor to implement that potentially in code somewhere or maybe as a configuration setting. In this world of kind of AI and agents, the chief compliance officer can go in there and change 50 miles to 75 miles.

Speaker B: Mhm.

Speaker C: And see how that impacts his portfolio of customers. Right. And so I think that's one of the. Back to your concept of risk architect. Actually being able to architect your processes in natural language rather than having to go through these translation steps is very powerful.

Speaker B: Yeah, well, uh, again staying with that theme, uh, or that concept, maybe we could all have been architects or I could have if I didn't have to understand physics and maths and all those things. Just talk. I want a house that looks like this. Um, Olga, we're talking about the screening agent. Is that the one you're most proud of or do you have like a top three that you've worked on so far that you're like these are my children as it were, that I'm sending out into the world and doing work with customers?

Speaker A: Yeah, uh, there are many agents that uh, we're working on and if I

Speaker B: have to pick, uh, I'm making you. So no offense to anyone that's worked on the agent that doesn't get mentioned, they're all good.

Speaker A: Uh, uh, another good example is uh, a due diligence agent. Uh, so this agent, think about this agent as a digital research colleague that basically puts together different parts of information that may be scattered in different places. So for example looking at company, uh, financials, looking at company uh, ownership graph, uh, looking at for example financial performance of a company but also uh, from more integrated risk point of view, looking at let's say sustainability profile, uh, of that company or looking at um, cyber risk for that company. So really having that comprehensive analysis that yes, potentially, uh, analysts could do it for a uh, um, handful of companies in reasonable time. But when the number of companies increase and you think about thousands or even millions of companies that this due diligence report need to be put together, that is a very, very time consuming task. So here the idea behind due diligence agent is that it's your digital colleague that never sleeps that pulls all this different information together and produce a comprehensive report that can be used for example for enhanced due diligence report or in general for your integrated risk assessment of a company. Would you like to do business with that company? Maybe it's your potential supplier or maybe you're Onboarding that company. So really bringing this information uh together. Uh so that's another example uh of what I like to call uh domain specific uh agent so screening agent, due diligent agent. Another I think very interesting example is what I like to call Intelligent Automation agents. Um so I like to refer to it as um uh entity Enrichment Agent. So think about um getting all the data for a company, let's say when you want to verify if this company, who they say they are you could connect to Life Registry uh data. See you know uh, if this company is really there, get some information from there. Also get some curated uh data. When for example um something is not available in Life Registry but also using uh Entity Enrichment Agent you can fill in remaining gaps. So you can also go to trusted and uh curated uh sources um websites, databases, uh, and so on and bring information that may be missing. So for example some contact information or some uh, you know names of uh people on the board or other information. So it's really doing it at scale. That is very cool because if you think about doing it um, kind of one by one as an analyst doing search that is extremely time consuming. So entity enrichment is really enhances that uh, that entity profile data. So that's another I think interesting uh example of agent that uh, I'm quite proud of that we're working on that

Speaker B: starts to solve some of the. We uh had a conversation last year with Ian Godfrey one of our MDM experts and master data management experts for those that aren't familiar with the acronym and the data Interoperability Challenge like potentially that is a part of the solution then right? Because you can go across these things that might not be fully joined up or you might work across multiple vendors plus your own data and then join it up there and then you have your own view. Um whether that's your golden record as people like to say or whether it's just something that goes into whatever pre state before you make your golden record. It would seem to have a lot of uh potential uses across all sorts of use cases. Once you get that bit right.

Speaker C: Yeah, I mean maybe just to pick up on that, an example of that might be getting information from a company's own website. And where the AI and the agent can be very good is if you go to a hundred different companies websites they look very different but almost all companies will have an About Me page somewhere and it can be buried many layers deep. It can be on the front page. On that page it will have usually the management team, sometimes the directors, usually any Legal disclosures they have to have about legal entities. So being able to access that once you've understood who the company are, uh, and compare that and combine it with the registry data that we have, for example, is very powerful.

Speaker A: Yeah. And have a confidence in that as well. So really kind of assigning uh, a confidence score if you want, when you see this information in, for example multiple sources, including company website. So this entity enrichment agent not only brings the data but also brings more value by seeing how confident you would be in reaching that data with that uh, information.

Speaker B: That's a great point. I mean I was about to ask because I've been working myself on sort of some of the fraud use case and one of the things I look for is like lots of fake websites now because you can prompt a website into existence in a few seconds. Um, so like uh, presumably everything leaves a footprint. So there is markers of that. But if the agent knows to flag that if it happens, then you can protect yourself from the proliferation of too much information. Yes, indeed. Yeah.

Speaker C: And you've got to look at how, when did that website coming to being, uh, for example.

Speaker B: Right.

Speaker C: Because yes, it's very quick and easy now to create a new legal entity, register it, create the website. Is the website a real website? Is it selling something? Does it, Is it got a real phone number? Has it got real. Uh, you know, there's a lot of tells once you start digging into it and thinking about now the beauty with an agent is you can train the agent on those kind of tells.

Speaker B: Yeah.

Speaker C: And then apply that approach at a

Speaker B: larger scale than you only bring the data back if it passes, you know, at least five of these seven checks or whatever it is, uh, makes a lot of sense. You mentioned the story of the crypto provider Kepler. What was the impact on that in the end once it went into deployment?

Speaker C: So that was really significant. I mean they offshored their analyst team and it's now 10% of the size that it was. Right. And they're very much in that type one trying to automate every piece of their operation, uh, using AI at this point. But they've really gone a long way. And yeah, it's essentially a 90% reduction in uh, the false positives they were seeing, but also in the team that they were needing to staff to keep up with that.

Speaker B: It's interesting, the percentages, because I remember there's a guy called Aaron Stroud who was the og, uh salesperson of what we now call Grid, um, four movies. But he always talk about no matter what system, uh, you're using, you will have a 90% plus false positive rate. Cause it's the way the systems are designed. What you actually want to look at is the workload because 90% of 100,000 hits is a lot more than 90% of 10,000 hits and whatever numbers you want to use. But if you're saying 90% reduction in workload on top of whatever all the filtering stuff we did before then that is changing. The TCO or total cost of operations came up earlier today. Yeah, um, that's changing that hugely. Right. That's, that's what people generally buy screening systems on.

Speaker C: Yeah. And then look at those analysts that are there are looking at the, the complex cases. Right. The ones that really need human judgment. Um, and it's really a reduction in I guess the false positives.

Speaker B: Yeah. And Olga, is there any other customer stories with these other agents? Obviously due diligence would use screening, but it's only screen is only one small part of it. The NC enrichment I just mentioned, there's loads of applications there. Is there any customer sort of stories that you're able to share, um, around

Speaker C: some of those other favorites I could maybe pick up. I was just at a banking conference and actually showing a number of people these solutions. Right. And one of the compliance officers I showed, uh, the due diligence agent to, his remark was, you know, the way it works today in the world of KYC and compliance is you do enhanced due diligence on a small set of your portfolio, the high risk customers, partly because it takes three to five days on average, partly because there's a cost involved in doing that deeper due diligence. M and his remark to me was that with that he would actually do due diligence on a lot more of the customers and he would do it more frequently than he can today. So I think that's the sort of thing that you see coming out.

Speaker B: So because the total cost of operations comes down to do the thing, he can actually raise the bar in terms of risk detection and protect the firm better. So that one isn't the efficiency as such, it's efficiency pursuing better risk control.

Speaker C: Correct?

Speaker A: Uh, exactly. And because those reports can be produced on a regular basis, so customers can schedule for example production of these reports every certain interval, uh, that they would not be able to do before because it was just too much work. So now it's really, you know, raising the bar as, as you said and also bringing even more information about those, those uh, companies than, than it was possible before. And the entity enrichment part uh, that I mentioned before. It's also part of those due diligence reports, for example. So if an agent for example is, is missing some, some data from a preliminary analysis, it can also go on those websites, trustworthy websites, and then bring that uh, that information as well. So you get three sources there, you get um, Life Registry that we have connection to, then curated data and also enriched uh, data that uh, that we built into, into the solution. So it's really a very nice combination.

Speaker B: Yeah, I'm just thinking obviously as I mentioned in the intro, like almost all companies seem in, at least in our space are working on something agentic or they're talking about AI, etc. Like what's then the difference with the agents we're talking about that Olga, you and the team have built and released versus other agents that might have a similar narrative out there. Um, how does somebody sort of thinking through oh, screen agent sounds good or due diligence agent sound good to me. However, there's these 10 options, five options, whatever it is, what would you be saying to them would be different about the ones you're working on?

Speaker A: Um, to me the most interesting part is what Keith was mentioning before is we meet our customers where they are and where they work. So not all our customers and the same place of their AI journey. So some of them uh, they prefer to use the uh, regular kind of UI of the platform, familiar interfaces. Some others they may be already thinking about using those workflows in large language models, for example uh, Claude or ChatGPT, Enterprise or Copilot, and then they want to do their work from there. Uh, so the agents that we built uh, enable customers to do so. So if they would like to do the screening, they use our um, model context protocol connectors and model context protocol applications to do the screening, to do um, the due diligence analysis, uh, where they, they actually work. So it's really what I really like is the way how they could do the work from different, from different places and then it activates the agents that need to be activated, it surfaces the information that needs to be surfaced. So it's really frictionless, um, exercise. So it's not agent per agent, it's really kind of building blocks.

Speaker B: Where does it go, what does it interact with, et cetera.

Speaker A: Exactly.

Speaker B: Um, and I know we talk a lot about the data as well in terms of obviously the cliche of garbage in, garbage out. Um, I won't accuse others of having garbage, um, but what goes in matters. Keith, could you talk about that a

Speaker C: little bit Yeah, I think at, ah, the heart of everything that's happening with AI is the need for really good quality data, or decision grade data as we've started to call it at uh, boonies. And um, especially in these kind of mission critical, uh, use cases that we operate in. Right. If somebody makes a mistake in the world of kyc, then there's potential financial repercussions, reputational repercussions. So thinking about the trusted data that goes into that's really important. Um, and I think that's one of the things that makes us unique is that we have this large data estate of, of millions of records, but millions of records on companies and individuals associated with companies and the facilities those companies work out of and everything from that to cyber events and uh, other sorts of risk events, extreme weather events for example, and how we join all those up into a knowledge graph. Now the power of that is you can access that information very quickly and easily because we've structured that data and joined it up. And that's really at the heart of what differentiates us and what we do. Um, I think we often hear, well, could you just use an LLM, an AI to go and find some of that data? And there are elements of that data where that is true. But I think as organizations are, uh, starting to understand if you have an AI go and scrape, uh, the data across the web, there's a cost to that. And if you're doing that every time you're looking up the name of a company, there's a cost to it, then do you trust it? There's been a couple of stories in the last few weeks about people planting fake research really for the purpose of demonstrating that you can't always believe what an AI is going to, uh, bring back. Um, and so, you know, having data that's trusted, having data that's joined, joined up, and the joining up of those data sets is one of the unique challenges and kind of the validation that we put into our data becomes very powerful. It also makes the AI that you're using a lot more efficient from a token perspective, as the cost of, you know, of accessing tokens is becoming a really important topic.

Speaker B: So what you're starting to touch on there is this ecosystem, right, Talking through, obviously customers wanting to get the best bang for their buck again, to use a, uh, cliche data providers, Moody's being one of them. But there's the model providers. Olga, you've mentioned a number of them already. Um, then you have the agents, whoever they're built by, and then how does it all stitch together who's building your system or integrating your system? And you've got potentially, as you mentioned, regulators. So you cross that span. Does it come down to this token efficiency and how do you run it as cost effectively or energy effectively as possible? Or how do you sort of see that ecosystem developing and sort of playing out together?

Speaker C: I think it's very much a work in progress for everybody right now because, you know, there's a lot of talk, uh, I was listening to a podcast on the way here this morning about, you know, a big tech company who already burnt through their whole year's uh, budget, um, on AI in the first quarter and uh, how are they managing that? And there's different models coming out with different costs based on the model. And I think people still trying to understand when do they need to use the most complex advanced model versus when will a simpler M model work. But at the end of the day, if you're tapping into this intelligence, uh, for the, you know, trying to optimize the end to end cost of a process, right. Think about the end to end cost of onboarding a new customer into your business or a new supplier. Obviously the token cost is a key part of that, right? There's, it's potential if, if you're not careful, you could end up with it costing more than uh, it did before because you've spent so many tokens. So how are you going to optimize the process, uh, the people involved in that process? Where do you need a human in the loop and where do you need those decisions driven by humans? And what's that overall end to end cost? They're all key factors.

Speaker A: Yeah.

Speaker B: And a couple of one slightly funny four and then one I'm not sure if is a story that was real or not, but one is, I'd love to know on the pie chart of that budget that got blown through, how much of it was people like making funny images and sending them to each other on the internal chat, whatever tool they use. I just couldn't say. It's fun, but maybe we need to put a lid on that. And then the other one was I saw some story and I don't know

Speaker C: if it was true.

Speaker B: As I said, I didn't verify it, but it's um, saying there was some company there that like hollowed out their engineering because they were like, oh, we're gonna get, you know, it coded by these, uh, agents or these tools, um, and then token uh, costs and they were like, actually it's cheaper to hire Junior devs and train them up a little bit and then get a balance. And so it's sort of like this,

Speaker A: I think I've seen that one actually.

Speaker B: Is that kind of, you know, the Onion, if anyone's familiar with that, these sort of satirical stories or was it real? I don't know. But um, if that is real, that's bit pie on face for the uh, team there. Uh, uh, do you know Keith, have you come across that story or.

Speaker C: I mean there's multiple stories floating around at the moment about. Yeah, I mean certain large tech companies having a leaderboards of the amount of tokens people consume. Right. And that can cause the wrong behavior because yeah, I could write something that will consume millions of tokens and that'll cost a lot more than, than any of the engineers salary if I got to.

Speaker B: So it's, I could just prompt it and say please use the most tokens possible today and leave it.

Speaker A: And that's why harnesses exist, because they are designed to use tokens in the most effective and efficient way.

Speaker B: Right, okay.

Speaker A: Um, so building those solutions using right harnesses and then providing them is indeed a way to actually save tokens. But otherwise, yeah, you can, you can burn so many tokens just by trying things that m. Maybe is not the optimal way to do what you want to do.

Speaker B: And then that's back to Keith's point. If the data's already connected, if it can take the fastest path to the um, answer, you get the efficiency.

Speaker A: Exactly.

Speaker B: Staying on data. AI ready data is a phrase that's now starting to bubble up. Maybe I should have been more aware of it. But just the last few months it seems to be more and more like what does that mean specifically? Is it just what we've talked about? Is it just knowledge graph and joined up data or does it mean something else where data is ready to be used by agents and yeah, how does that play out? I don't know who wants to take it first but yeah, maybe I can start.

Speaker A: So, um, when we say AI ready data, usually what it means is when you would normally use your API, um, it would be used by a human deciding, you know, which endpoints to use, you know, what information you need, and so on. However, when we build AI solutions, it may be used by both human and also AI agents that you know, someone, someone built. So we need to be ready for this type of situation too. So instead of building only APIs, we would also look at building what we call smart APIs, which is basically a way how to enable agents, uh, to consume and understand the API, uh, and have that kind of translation layer, uh, explaining it basically how it should be used, what endpoints to take, what does that mean? And on top of that what I mentioned earlier, a model context protocol, which is basically in a very simple term, it's a USB C connector, it's how you connect one AI to another AI. So think about it uh, uh, as a buying washing machine and then you have an instruction how to use it. So that model context protocol, it basically gives an AI a protocol how to use that data. So it actually, when you go and uh, for example open your chat and say I would like to screen this and that company or I would like to build early warning system, it would already know what to take from that Genai data. It would already know how to use that data without us, you know, feeding in lots of documentation and reading through uh, you know, ways what this data is reading through data dictionaries and so on. So it's really a combination of smart APIs and model context protocols and MCPs.

Speaker C: Yeah. So maybe from the engineer's perspective, yeah, it's about that self describing nature of what the API can do and then how the AI agent can use it to do things. So an example, you know, when the MCP protocol came out, which must be what, a year ago now? Roughly, Yeah, I think we created some MCP wrappers around our APIs. Traditionally an engineer would code to that API and would always use it in the same way. And we have a uh, thing called the eva. MCP can go direct to corporate registries, look up a company, find its information or an EVA API and if you were a programmer writing that into a script, you would use it in one specific way. M, uh, uh, we put an EVA MCP together which describes to the agent how to use that API. And, and it worked. Right. So you could ask the agent to find information on Moody's UK Limited and it would bring back information on Moody's UK Limited. But what was really powerful and what's different is you then you know, I remember doing this and this was one of those light bulb moments with AI where it does something you're not expecting. I think I said tell me all the Companies based in 1 Canada Square, uh, London, which is where we're sat today. And it literally used an uh, endpoint I wasn't as aware of because it was documented in the MCP and searched all the companies that were based in 1 Canada Square and gave me a list back. And so that that documentation built into the Smart API or MCP allowed it to understand that it had those capabilities.

Speaker B: Okay.

Speaker C: And go and do that for me.

Speaker B: So when we say AI ready data, the database itself or the service in the case of either, uh, is the same. It's literally the documentation and the connectivity, I guess, or how you connect with it that is changed, correct?

Speaker A: Yeah, that's right. It basically comes with instructions. It explains to AI agents, uh, how to use that data and then the possibilities are endless. You can search for companies as Keith described, you can generate reports, you can ask uh, to build the whole application or generate dashboards. And it would already know what endpoints to pick, um, what to do. And without you being as much involved,

Speaker B: I'm going to say they should stop calling it AI ready data and call it AI ready manuals or AI ready wrappers, which I think rolls off the tongue better. Stop confusing me. Um, there's a couple of audience questions I just wanted to put to you guys and then we'll start to wrap up. Um, one was somebody sort of been learning about agent skills. They said and they were like, their understanding is, look, agent skill is a set of repeatable steps that an agent can follow and say, well, uh, repeatable steps. That's kind of what we've been doing for years. And they have been in their roles and previous roles. Right? Orchestration. Do this API, then do this one, take this data field, do this one, do this one, et cetera, get to an output. He's like. And what he's hearing is that it's like very energy expensive to run these agents for. And you don't always need to. It's like why wouldn't you use an agent's skill versus uh, just have an agent go and run some orchestration flow that doesn't need the ainis for lack of a better term. Or me and him both missing the point.

Speaker C: I think both approaches are possible. Right. So if you think about some of our solutions, we are wrapping the solution with the ability for an AI to call it. And the solution has a pre built workflow. Right. We have products in the market that follow an onboarding workflow and that's possible. But it's really back to that what I explained to you about um, ah, that chief compliance officer at that crypto firm, right. A skill is really written in the domain language of a business user.

Speaker B: Okay.

Speaker C: Right. So imagine the compliance officer, uh, just writes the skill for how he wants his onboarding process to be done in English, right.

Speaker B: Without or any other language.

Speaker C: Or French or any other language.

Speaker B: Correct.

Speaker C: In natural language. Sorry, um, and then rather than having to pass that off to a developer or to an implementation consultant, that becomes the playbook. And so that ability for the business person to evolve the solution with skills is very powerful.

Speaker B: So the value isn't the fact that it's agentic or AI driven, it's that, uh, it's faster time to value, it's less middlemen or women in the way basically to get the thing done. It's just, I want to do this, I do this thing, the tool does

Speaker C: it and the agent can, you know, that can be an iterative process. There's some really interesting stuff into the latest skills about running evals and having, uh, processes that self tune those skills to make them better over time. So there is learning that can go back into the skill, there's ability to branch. But I think the core thing is really capturing the business knowledge in natural language.

Speaker B: Trust in. And the other question we got was somebody sort of observing that. So there's the AI race, right? Or the model race between the various providers who are building that, but then there's also a race for proprietary data. And we probably consider ourselves in that race as a data business primarily. But the question the person asked was like, will AI and agentic workflows, will that make it easier for new data providers to emerge? If somebody wants to build a proprietary data set, can they just do that much faster, much cheaper now? Because the toolkit is fundamentally different to the one that the big data companies started with, whether it was 20 years ago, 30 years ago, 50 years ago, et cetera.

Speaker C: I mean, I think there's elements of, uh, you could argue that there's elements of that that's true, but it's really the depth of the data and the complexity in joining it up. M And the most. I mean, we've tried to do some of this internally at Moody's, right? And we obviously have a huge database on companies. You can give an AI a company name and it will bring back to you something quite plausible about that company. However, number one, do you trust that? Where did it find that information from? And as we talked about earlier, there are examples now of people creating information that's obviously false and proving that that can come through an AI. So number one, there's the kind of provenance of that data. But number two, what we've actually demonstrated to ourselves is the complex data. So you might find information on that company, but will you find information on the other 124 legal entities in that corporate group? You probably won't just from a simple question to a lord or a chatgpt or um, you know and we've kind of red teamed some of this and tried to, to to create data in that way and really found that it's not practical. So it's, it's back to that decision grade data and trusted data. How you know, can you speed up some pieces of it potentially? Um, and that's. But we're doing the same thing right. That's where our uh, entity enrichment agent comes in to layer in uh, data that is, is more widely available.

Speaker A: Exactly. And to that previous question that, that you had from audience on skill. So basically building some of those skills and then using that to enrich the data and then building an agent around it because in the end of the day skill just sits on top of that hardness and being application. So yes in theory it can be you know, uh, faster and uh, you know it's more possible than before. However just uh, the pure scale of it and you know all the interconnections make it very, very complex task to do.

Speaker B: Good. Well, we will wrap up there but one last question. Um, Olga, I'll start with you. If somebody wants to learn more anywhere you'd send them anything either you've written or you've read that you think is worth others uh, checking out.

Speaker A: I think in general just staying kind of uh, um, up to date with all the amazing uh, AI race, uh, seeing what's new in AI world. Um, so I follow various researchers um, on Twitter on substack. So um, for example I read a lot of Andrei Karpath blogs. Uh, he always uh, has something interesting to say. So really kind of staying up to date and understand what you could be uh, using in your work and also being kind of open minded of trying things because our own workflows change. Uh so when we were building agents, uh, we were also using lots of agents to build those agents. Because really now when I do certain tasks, first thing I do is can I use AI to help me do that. Um, so staying up to date with what are the latest developments, what are skills, what are harnesses, uh, what can I do today that I couldn't do yesterday? I think uh, it's a very important skill to have for everyone. So that would be my very good.

Speaker B: We'll get the link to the uh, author or blogger that you mentioned and we'll put that in the show notes. Keith, anything you'd suggest people check out?

Speaker C: Yeah, I mean blogger's been very modest. She's published a number of things, uh, that are on the Moodys.com site that we can link to around how we're using agents and AI in our products. I mean, I think follow any of us on LinkedIn, uh, because we often talk about what we're doing. And I would agree. I think staying. There's kind of a LinkedIn community around AI and agents. There's definitely a community on X, uh, of people who are posting pretty frequently. Um, and YouTube actually is, uh, the other interesting venue because it's literally changing every week.

Speaker B: Yeah. World's biggest education platform.

Speaker C: Correct.

Speaker B: I think. Um, anyway, guys, we'll link to all of those things in the show notes. Thank you very much for coming on and updating me, updating the audience, and, yeah, look forward to having you back another time.

Speaker C: Thank you, Alex.

Speaker A: Thank you. Thank you, Alex. M. Thanks for listening to this Moody's Talks podcast. To find out more about the topics discussed, please follow the links in the show notes. You can check out other Moody's Talks podcasts by visiting Moody's.com podcasts.