Moody's Corporation ($MCO)

Good afternoon, everyone, and welcome to today's call, and we're excited here at Moody's to have Andrew Steinman, Managing Director and Equity Research Analyst at JPMorgan moderating this session with Christina Poretti, who is a General Manager and Head of Generative AI Solutions at Moody's Analytics. The questions have been presubmitted, and Andrew will be moderating. And so Andrew, thank you so much for doing this, and over to you.

My pleasure, Shivani. Thank you, Christina. We enjoy this research dialogue with you. Christina, what you just start out with how should people think about the AI strategy at Moody's?

Of course. Thank you, Andrew, and thank you, Shivani. A pleasure for anyone that's listening to be here speaking about this topic that it's highly relevant and which we're extremely passionate about. So when I think about GenAI strategy and Moody's GenAI strategy, the first thing I think we have to keep in mind is everything seats inside Moody's agentic solutions, right? And I would encourage that everyone that's looking at this thinks about 2 layers and a third pillar that is about how those layers reach customers, right? So if I think and if we can show in the slides, we're going to show what those 3 the pillars are, right? So the first pillar is connected intelligence. And this is highly relevant. This is a foundation and the sequencing actually matters here, right? Because you cannot deal decision grade agents on poor data. And what makes Moody's unique, it's just not the volume of data, which, of course, is 600 million entities, 2 billion ownership links, our research, our ratings when it's the depth of the domain expertise embedded in this data over decades, right? In credit risk, as an example, we have Moody's Ratings, which are originally proprietary, for KYC and compliance, we have our Orbis database, which provides beneficial ownership mapping and NDP resolutions across 170 data resources, right? So that's -- and what we do is we collect the data, we connect that, we curate it, so that's the foundation of everything. Then we go to Pillar 2. Pillar 2 is about agentic workflows, right? And it's how we package that connector intelligence into purpose-built end-to-end workflows. And there are a couple of things that are very important. First, we focus on workflows where making bad decision is going to cost you a lot of money. And what we mean by that is you don't want to make a credit risk decision because you're going to lose a lot of money. You don't want to underwrite the right insurance policy and not look at the risk. Again, there's big financial consequences. You don't want to lend to and engage into a relationship with -- that is the result of making a wrong KYC check because, again, it's going to cross do fines, a lot of money, a lot of reputational risk, right? So it's about developing workflow solutions in those high-stake areas that are leveraging all the connected intelligence of Pillar 1. And then in the case of Pillar 3 is how do we then reach and distribute both the connect intelligence and the agentic solutions, right? And the idea behind this pillar 3 is, we want to meet customers wherever they are building and working with AI. So when we think about Anthropic, about AWS, about Microsoft, OpenAI, Databricks, Salesforce, we want to make sure that we are meeting our customers where they're doing their work. We don't really see them as competitors, we've seen them as partners that amplify our reach, and I think there's a couple of very important things in terms of that. First, it allows us to reach new buyer personas; second, in each of these cases that I named, we are maintaining the customer relationship and we retain the IP, right? So if you think again, to recap what I've said about the strategy, we are addressing 3 things. We are addressing the customer need for trusted defensible intelligence in high state workflows. We are addressing also customers' GenAI maturity from those building their own models, their own with our data to those that prefer to consume decision-ready workflow outputs, and then we're also reaching them where they need us to meet them. And that's basically the strategy.

Okay. Christina, Moody's has announced partnerships with 4 of the big AI players, AWS, and thoracic, Microsoft and OpenAI. Could you just give more color about those -- the nature of those partnerships?

Absolutely. So -- and I'm going to start by entropic and probably because Anthropics every day on the news with a new announcement, right? So when I think about tropic, it's probably our most architecturally distinctive partnership. We have basically deal 2 things with them. Back in November, in Q4 last year, we announced the launch of MCPs, right? MCPs that allow our common clients to access our data through cloud. The other thing that we most recently announced, and we believe it's the first of its kind as far as we are aware, is the launch of an MCP app which is an interactive agent interface that lets users access moist agents, generate outputs and trade sources without leaving the cloud environment, right? So it's not a data feeder on API, it is Moody's Intelligence that is rendered, in the first case, as data that is -- can access through chat, and in the second case, through actually a workflow right? And an example of that workflow would be running an ownership check or running a portfolio monitoring workflow or running a credit memo write-off, right? So that's Anthropic. In the case of AWS, I think more about 2 things. One is access of our agents and our data through the cloud marketplace through the AWS marketplace. And then most recently, we've also integrated into Amazon Q, which is AWS native generative AI chat interface, which means that customers can [indiscernible] voice intelligence conversational, we've seen the AWS environment. So you don't have to leave again, if you're using AWS, you can buy our agents, and that gives us, again, increased customer reach to the marketplace, and then you can also converse with our data and our agents through [indiscernible], right? I'm going to move now to the third partnership, which is Microsoft. And we think Microsoft as our productivity layer play. We all know the reach that Microsoft has. We are all users of Microsoft. We are basically embedding decision grade intelligence directly into Microsoft 365 Copilot, Researcher and Excel through a dedicated Moody's agent and MCP integration. And I want to be clear, when we talk about Moody's -- a dedicated agents, this is the way that the Microsoft environment works. It packages things to an agent. But what it means is, in your, in using Microsoft Street Microsoft 365 Copilot, if you're using Excel, if you're using Teams, you can interact with the Moody's data directly through any of those environments, of course, provided that you're already a Moody's customer, right? And then the fourth partnership, which I'm going to describe today is OpenAI. There are MCPs live in ChatGPT Enterprise. And basically, again, similar model to the other ones. You have to be a customer of Moody's, and you can then -- you are using ChatGPT Enterprise and you can access the Moody's data. So if you think about the 4 partnerships I've described, there's a common thread here in every case. Moody's retains the customer relationship, Moody's controls the pricing and fulfillment, and our data is not used to train third-party models. The partners are the distribution surface, the intelligence, the IP, the customer relationship remains ours.

Right. And Cristina, you would imagine those premises that you just said will continue going forward as well. I mean, obviously, we're at the early days with these partnerships.

Absolutely. And you actually see how we have now dedicated teams inside Moody's. When I talk about dedicated teams, I would describe them as squads that are working with these partners. We've been very deliberate in the partnerships we form, but we are -- this is not a one-off. You're going to see continuous announcements from Moody's as this partnership launch more features, more skills, more tools, et cetera.

Okay. We'll talk a little bit more about the partnerships. But what I wonder is about LLM token economics. Like when I hear the word agentics and MCP applications for Moody's customers, I wonder who bears the cost of the tokens and obviously, tokens could be inflationary is the client bearing the token cost? Or are there also times when Moody's in these agentic workflows or MCP applications are bearing token costs?

Yes, yes. And this is a question we get -- we get a lot of questions around this topic. And I think it's worth answering it very careful because the model of token economics is going to depend on how the customer is accessing Moody's intelligence. And they are basically 2 paths, right? In the first path, the customer is consuming Moody's workflows directly through Moody's own environment. So let's leave those partnerships that I described aside for a second. We are basically contracting with the customer directly because the customer is buying from us an MCP or it's buying from us an angetic workflow. In that case, Moody's carries the underlying token cost and builds them into our pricing. So if you think about the risk of the token, it's on us. To understand what is the cost per token, of course, we do have -- we do negotiate the volume with the customers. But at the end, that token cost is bear by Moody's and it's including in the price. So the customer gets kind of a clean, predictable relationship, they pay for Moody's Intelligence and outputs without managing the variable token consumption separately. What they do have to mention, they have to manage kind of the volume, right? If they contracted for a certain number of outputs then, of course, if they go above that output, then they would have to -- they would pay more. So that's when they contract directly with Moody's. In the second path is when they're contracting their using our MCPs on our -- or agents through the partners I described in your previous question, Andrew. So basically, the customer is interacting with Moody's MCPs through a third-party AI environment, like Claude Enterprise, like ChatGPT, like Microsoft CoPilot. And in this case, the token cost sits with the customer. because they are already operating with them and paying for the platform environment, right? They already have a relationship with Claude. They already have a relationship with Microsoft CoPilot. So Moody's is not in the billing path for those tokens. The customer has a relationship with that provider and then they have a separate relationship with Moody's for the intelligence layer on top of it, right? So you can see yes -- I'm sorry, go ahead.

Yes. So just maybe -- I totally understand the second point. Why don't we -- just to make sure we get it on the first part, we you're like, when we're contracting directly, we build token prices into the contract. My question to you is, as token course go up or the volume of consumption goes up, you're saying the price into the client adjust, and so this isn't a possible mismatch for Moody's, right?

Yes. Yes, it could be, but we are needless to say, very careful about it, right? So first, we are monitoring every single thing that the customer consumes. And we're also monitoring very closely our token cost, right? And it's not only about the token cost, but what is the model. We have the ability to select the model we're using for everything that we're providing to the customer, right? So -- and we are very careful to use the model that makes more sense, not only from an economic standpoint, of course, but also from a performance standpoint, from a reasoning standpoint, from a follow direction standpoint. But that us -- that gives us freedom to say, well, we're not maybe used for the task the most pricey model because it's not worth, it's not really going to make a difference, right? So we have several levers here to control first. We are looking at token costs very closely. Second, we're looking at the volume and the consumption, what is the cost for us of clients consuming, and then we have the lever of controlling the model. So as of now, we feel pretty comfortable, and we have the necessary buffers built in. So that's how we're approaching it right now.

Okay. That sounds good. I think you sort of just led towards this term that we hear a lot about consumption pricing. And maybe you're going to say, I just defined it for you. But just because there's so much discussion about consumption pricing in an AI context, how does that work for Moody's?

Yes. So I think this is something we've been extremely careful about. And I think we want to be -- we want to continue to be careful about, right? So, yes, I think consumption pricing can be something very powerful because as customers consume more data, run more agentic workflows this is something that can be beneficial for us. So I see it as a potential uplift, right? Now we've talked before with you and many other of our analysts and investors about there's a downside to it as well, right, which is the volatility here. So when we're thinking about consumption price, we're basically thinking about a base price, and we always price our arrangements so far as a base price that guarantees a minimum consumption. And then if you go above that consumption, then there's kind of consumption driven pricing, right? So yes, we -- as a customer -- as this gets more ingrained in the customer and the customer consumes more, we kind of benefit from the uplift on that. Of course, with have a pre-agreed pricing arrangement with our customers, but we also want to make sure that we minimize the variability or the volatility of our revenues. Does...

It sounds like the volatility could really only be to the upside, right, because you have your base amount of pricing and then you're paying for overage if you go above that.

Yes. Yes. Yes. That's basically the the idea, right? And just to -- because I also want to be mindful with our clients there, right? So when you think about the data, in the data, there is a potential to be more overage because of the data gets more democratized, and we're seeing when we talk to our clients and we engage our clients, there's more appetite for enterprise licenses, right? One of the things that has happened with GenAI, which we actually see as a tailwind is it has democratized the access, right? It makes simpler to use the data, it allows for more data to be used in more parts of the organizations. So you could see more increase there. I would say when you're talking about agentic workflows, you kind of know what's your business volume, right? So it's more difficult to go above that. But again, what you stated, Andrew, is, yes, the upside is also -- it's going to be generally upside. It's difficult to be downside because we're protecting that through a minimum through that -- I'm sorry, I'm missing my word, but yes, so that guaranteed subscription basically.

Okay. Great. I'd love to get into specific use cases or solutions. So like when you look at agentic AI at Moody's Analytics, could you go through maybe 2 or 3 solutions that you're prioritizing with Moody's clients today? And why did you choose these use cases to kind of be the priority first?

Yes,yes, yes. So I would say, I think there's 2 things that I would highlight here, right? And when we think about prioritization and most importantly, we think about our right to win, we're going to think about 2 things. One is where do we have data that is proprietary, that it's connected, that is that connected intelligence that we refer to because we all know that, again, you cannot build decision grade agentic workflows on poor data. So it doesn't matter who's building those agents. If it's Moody's agents, if it's third-party agents, we want to make sure that we have the right data, the right context data, that it's AI-ready first to make sure that we can focus on those agentic workflows. And then the other thing that we've been very deliberate about is that concept of prioritizing those places where the stakes are highest, where a wrong answer has legal, regulatory or financial consequences, and where Moody's has domain expertise, right? So I'm just going to repeat that, and then I'm going to give you a couple of examples. Places we have proprietary data, that it's connected, that has a proper context layer, so it's basically AI-ready. Second, those cases where stakes are higher because the wrong answer has a lot of consequences, and third place is where Moody's has domain expertise. So if you put these 3 things together, we basically, as of now, have come into 3 areas. One is credit risk, and that's where we started at the beginning, right? So it's what are the type of data and/or workflows that you need to leverage GenAI for credit risk assessment and for lending. So examples of that is how -- with automated credit memo, how we're doing automated early warning. It's all the MCPs that we have rolled out even either independently or through the partnerships in terms of ratings, research, probability of the fault models, thermographic, financials, et cetera. Second use case it's know your customer. That's our second priority. And those are things such as entity profiling, ownership mapping, adverse media, sanction screaming. So basically, a lot of it is coming from the Moody's Orbis database, right? And then the third one, which is it's -- we're just starting on, it's basically the insurance underwriting path. So Moody's risk models, climate analytics, ESG data that can create a differentiated foundation for underwriting work.

Okay. I wanted to get a sense of if a client, and I mean a current client that's already accessing Moody's Analytics data probably through an API. If they choose a smart API or more likely an MCP server, are they paying more to access the data in an additional way, or is that part of the existing contract? In other words, when they go from API to MCP, even if it's the same data set, same customer, is that like an upgrade where they're paid more? And if they are paid more, why would they switch?

Yes, absolutely. And this is a great question. So yes, even if they're an API customer, we are charging a premium for that, right? And the reason for that and how we justify to our clients is the following, right? When you're thinking about an API, an API is going to deliver raw data, which means that on the customer side, a group of data scientists, developers want analysts have to take the data and build something on top of it, a model, a workflow, a dash word. And of course, the data is valuable, but it requires a lot of investment, expertise, ongoing maintenance, right? So basically, you can think about when it's an API, they're buying kind of an ingredient, right? When you think more about what they're buying with an MCP and it's -- we are already packaging the data in a way that makes the agent -- and again, we're not talking necessarily about our agents, we're talking about large language models, we're talking about customer agents or any third-party agents. It makes those agents -- it makes the job for that agent, I'm sorry, much easier, right? Because the agent has an easier time understanding that it has to use this data and how it has to use the data because it basically has instructions for the agent on how to use the data. So you might say, well, Christina, that's great, but isn't that a nice to have, right? And why would a client pay more for that? And the answer has several reasons behind it. First, it's speed, right? You can basically -- by giving this clear instructions and that clear context layer, it means that the agent can go leverage -- connect with the MCP and get you an answer extremely fast. Second, there is a cost element, right? Because you are -- if you don't find the answer, if you are working with an agent or an LLM, it doesn't find the answer, it's going to keep looking everywhere it can to not only find the answer, but also, for example, at this -- here connected intelligence comes into play. If it needs an answer that requires several things, it might -- that looking for an answer might take more and more time as it constructs the answer, right? While if we are packaging everything in one MCP and we're giving clear instructions, that means that your token use is going to go down, right? And then the third is kind of the auditability, the knowing that the answer you're going in GenAI can -- it's backed by Moody's, right? But I would really, really emphasize the first 2, one is speed and the second time it's cost on the client side.

Okay. That makes sense. So you using a lot of phrases, and I just want to make sure the audience catches what you mean by each of these phrases. I'm just going to mention 3 phrases. Contacts layer, you say that a lot, decision grade data, and I forgot if you said this 1 today, but I definitely hear Moody's talk about Knowledge Graph. And just if you can go through in the context of AI and Moody's, what each of these mean for the Moody's universe?

Yes, yes, yes. So I'm going to go through the 3 of them and then actually go through the 3 of them in the way we construct them, right? So the first one, of course, is we get -- we have our raw data. And we -- I'd like to talk about it as decision-grade data because we never expose to our customers or to our internal application just the raw data, right? Where we end up exposing is what we call decision-grade data, which is basically the step that we hold our data to. So what does it mean? It's sourced, it's curated, it's explainable, it's auditable, which if you think about where we are focusing our efforts is extremely importable, right? Because it's then feed for decisions that carry legal, regulatory or financial consequences, right? So if you think, for example, about data and you scrap from the web, that's not going to be the decision grade. But if you think about data that has been collected, sourced, QA connected that then is what we call the decision grade, right? So the decision matters a lot because in regulated financial services, the provenance and the auditability of the data is as important as the data itself, right? So that's what we call decision-grade data, data that we can stand behind, that our clients can say, I can trust the data, it's coming from Moody's and then I can say that to the regulator. The second term, you asked me about, and I don't think I had mentioned it in the call yet, but we're talking about it a lot, and we believe it delivers a lot of value to our clients, and it's constructed on all the years of data and different acquisitions that we've made is in knowledge graph. And this is basically the architecture that makes that decision-grade data interconnected rather than siloed, right? So it connects those 600 million entities that we have in the org database, with 2 billion ownership links. Those 2 billion ownership links are across jurisdictions, right? It connects those ownership links with rings, with -- I'm sorry, with ratings, with other credit scores, with catastrophe models, with tenants, if we think about commercial real estate in one single intelligent fabric. And of course, the other thing we're doing with the knowledge graph is we're doing knowledge graft that are specific than 2 use cases, right? So you have a knowledge graph for a sales and marketing use case. You have a knowledge graph for a compliance use case. You have a knowledge graph for a credit risk use case because the time of data that is relevant for you and that you want to be connected, it's going to defer by the use case, right? So that's the knowledge graph piece. And then the third piece is, once we've connected all that data, so think about the process, right? First, I described decision-grade data, you're cleaning, standardizing collecting making sure you can stand by that data. Then we're connecting that decision-grade data. So you have -- you can get all the relevant insights when you're analyzing something. And you're not -- again, I'm going to go back to the previous question. You're not relying on an agent or you're token cost to build all those links. And then once we have that connected data then we're going to build a context layer. And the contact layer is what sits between the knowledge graph and the AI reasoning engine, right? So think of it as the instruction layer for AI, a structure governed representation of what the data means, how it relates, when it should be applied, what caveats apply. And that basically is what it's -- translates is into increased accuracy and increased efficiency, right? So it's fair to say that without a context layer, an LLM can access data, but cannot reason about it in any way that it's defensible in a regulated environment. So I'll stop here to see if you have further questions on this.

No, not on those 3 terms. Maybe we'll move on to the data moat. Obviously, MA breaks up its business into 3 subsegments: data and information, research and insights, decision solutions. My question is, what is the strength of your data moat in each of those 3 subsectors? And then also, a lot of terminology goes around this word proprietary. Maybe you should just -- as you talk about the strength in your data moat, just to find what you guys mean when you say proprietary data?

Yes, yes. So the first -- I probably will say that if I think about the proprietary data mode, I think it's not necessarily data in each of these places, it's kind of the foundation of each of this, right? So the 3 segments is how do we organize and deliver value, right? We deliver value by providing data and information. We deliver value by providing you research and analytics and then our decision solutions, which are KYC, lending, insurance, et cetera. But actually, what I can think and when we think about the data mode, it's what makes everything defensible. It's basically the foundation, right? And let me tell you why we think this is a moat. And there's 3 important components on it. The first one is access, right? So a large portion of the data we have, it's simply -- it's not publicly available, right? And we have -- over the years, have created and have developed a lot of commercial agreements, licensing arrangements, royalty relationships with over 170 sources that, again, have been either exclusive or semi-exclusive so that provides a barrier to entry. But it's -- you would have to basically reconstruct a global network of supplier relationships from scratch, right? So it's not data that you're going to go and access in one place, it's basically built over a network of relationships in different jurisdictions, different countries. The second one is kind of proprietary creation. And those are assets that Moody's originated and that exists nowhere else, right? Now of course, the prime example would be the Moody's Ratings, right? So no one can generate a Moody's Rating. No competitor can replicate the regulatory acceptance and the institutional credibility behind it, right? And then the third angle of this is the construction and curation, right? It's what I've been talking again about connecting intelligence about it's a work of linking, resolving, standardizing and continuously maintaining data across jurisdiction. And I think that we're -- the part of maintaining its incredibly relevant, right? Because you can do this once, but this day changes constantly, right? So all the time, we're continuously refining those links, refining that entity resolution, working on the standardization and making sure that everything is data, it's decision grade data, right? So if you put all these things together, you put the fact that you have all those relationships with providers, right, more than 170 sources, update on that is not publicly available, you have your -- the assets that you are complete -- you're creating, right, the ratings being the prime example. And you put then the construction and the curation and the linking of all of this then you have a pretty robust moat, right? So going back to your original question, then I would say data and information, it's basically more of kind of the the pure data, right? And again, it's -- then you have -- we try to describe, right? And it's all the linkages that I described, it's all the curation, it's all the standardization. That's number one. When you move to research and insight, it's transforming that decision-grade data into analytical output, right, into credit opinions, into sector research, into those probabilities of the fold that we've built out of our historical default database. And then when you think about the decision solution is when that intelligence becomes workflow-ready tools, right? An example of that being credit lens or some of our catalyst solutions, et cetera, right? So the point that I would like to leave is, it's a combination of all of this that makes you powerful, right? It's a combination of having that connected intelligence as a foundation. It's how we build that through as I described. First, the access that you [indiscernible]; second, the proprietary creation; third, the curation and then the analytics we've developed on top of those and then the subject matter expertise and the relationship we have with our clients to be able to automate those workflows. Now I don't know if I answered your question about the proprietary data or -- okay, good.

Here's a question. Within your MCP protocols, what data protections do you have to prevent the LLM -- a third-party LLM from memorizing your data sets, training on your data sets and particularly in your partnership agreements with companies like Antropic, is it specifically in your agreement that they're not allowed to train on your data?

Yes. So we are extremely deliberate and focused both with those partnerships and with our customers that there's no training allowed in our data, right? So Number one, is from a contractual position, we -- there's a firm contractual position across our partner agreements, right? We have a dedicated privacy program, information security program, all are publicly documented that govern how the data is handled across all products and integrations, right? And the same standards are going to apply with, again, as I said, customers or with the partners, right? That's number one. Number two, the MCP -- we're being very focused on MCP architecture as a way that we want to distribute our data for GenAI purposes because of what it means, what are the implications of an MCP, right? So it basically allows our data to be accessed through a controlled interface. So it's not transfer. When a customer runs a workload inside Claude or another primary environment, they're clearing Moody's data through the MCP, they're not receiving a copy of the underlying data set. The data remains within Moody's governed infrastructure and then the outputs are generated on demand, their stored, they're attributed and the underlying data is not really exposed in raw term, right? So that gives us a lot of -- and then I would say the third angle is we do monitor, right? We monitor the volume of calls that is done through an MCP or through smart API, et cetera. So I would say between the contractual agreements, the fact that you're not receiving a full copy of our database, and then the fact that we're monitoring all of this, there's a robust framework there to prevent the training of -- to protect our MCP protocols and prevent the training by LLMs. And maybe I'll add 1 more thing, Andrew, which is because of the nature of the MCPs, even let's say that you pull a lot of volume at one point, it's going to be a point in time kind of data dump, right? And when you think about the nature of our data, it's very important that you have real-time data. So even if you were -- if a snapshot was theoretically possible, it would not solve the customer's problem because our data is continuously updated, curated and enriched, so the ball is not in the static data set. It is in the leading govern current intelligence that reflects what are today's entity structures, what are today's ratings, what are today's news.

Yes, that makes a lot of sense. Christina, a term that you used just a moment ago that card is that we could monitor the volume, like if it of our clients are trying to download an unusual amount of data unusual relative to them. My question to you is it just because you monitor the volume. But do you have audit rights? Obviously, you have these contracts with partners and clients and do you retain the right to ensure to audit that the data is being used in the scope of the contract and not, let's say, go outside the contract?

Yes. So we do have -- we usually have audit rights within the contract and that's something that we have even before GenAI. So again, I'm going to answer -- I'm going to say yes to your specific question. But I would say, again, it's 2 things, right? In all our agreements, we're defining very clear what the data can be used for, in what context and by which users, right? And then yes, we have -- then we have the monitoring in place in terms of not only the volume, but what type of data they're using. And of course, it's not only because we want to monitor, it's because we want to make sure that we are investing in the right places. And then the third thing is, yes, we do have auditability clauses in our contracts.

Okay. And usually, when you find that you audit the data, and there's like, let's say, more users at a client, the client usually just paid for that, right?

I'm sorry? Yes, if there -- when there's increased -- yes, when there's increased usage by client, yes, the client will pay for it, yes.

Okay. How about let's talk a little bit about cross elegant upselling. What within AI capabilities across Moody's -- the Moody's platform will drive more cross-sell and upsell?

Yes. So I would say -- I'm going to point to 3 things, Andrew. One is the metrics that we see, right? And we, in general, when we look at our customers that are using GenAI solutions by Moody's, we observed 2 things. We observed higher retention in that cohort and then we observe that they tend to consume more content, right? So that's a clear indicator that when we have AI adoption, it deepens the commercial relationship rather than substituting for it. We actually see higher retention and we see higher consumption which, of course, is a leading indicator for us to be able to increase our revenue or our commercial relationship with that customer. The second thing, and I touched on it earlier is when I think about the possibilities with Jenny, we are seeing -- if I think about the data, we are seeing, especially from Tier 1 institutions, more of a drive to enterprise licenses, right, to we want to use our data kind of throughout the organization as opposed -- versus in silos, right? So of course, that drives more consumption of the data. And then when we think about the agentic solutions, then there's a possibility of automation, which also allows us to tap into a different kind of -- a different part of the wallet of our customers, right? And then the third part, when I think about cross-selling, and I think this is -- I'm extremely excited about this is the partnerships, right? Because in that, it's not only that we're meeting the customers where they're working, right? But it also allows us to tap into a new buyer personas, right? So -- and that means customers that were not necessarily previously direct Moody's customers, but that now have -- can access our data through this new platform. So I would say there's a deepening of the relationship we have for our existing customers -- our existing customers through more retention, I'm sorry. There's increased consumptions for those organizations because of that use of more data for genai solutions, the need for reputable data in GenAI solutions. There's -- when we work with workflow solutions, we're tapping them into the automation budget. And lastly, there is the ability to tap into new buyer personas through our partner ecosystem.

Okay. Obviously, that all sounds credible and good. You know research analysts is supposed to have some healthy skepticism as well. And so my question is what I'm going to ask you about. When your team looks at Moody's business, what are the credible risks from AI? In other words, when the Moody's leadership team realizes that there's benefits and risks, what's like one area of risk where you like we have to get this part right?

Can I give you 2.

Yes, I think so.

Okay. Good. Great. So the first one I would say, and it's -- that's one that kept me on my toes every day, it's speed, right? I think we have to make sure that we are really, really focused on the speed of embedding our data, right? I think the risk here is not that our data becomes less valuable, it's that that the customers establish agentic workflows with with other intelligence providers because we were not there, right? And that's why you've seen from -- from very early on, you saw us launching in 2023, the -- what this, I'm sorry, I kind of believe -- I launched this product, and I adjust planned on it's name, Moody's Research Assistant. And then we saw -- you saw us coming with agentic workflows. And then you saw us coming with MCPs very early on. We were the first one to launch an MCP app in the market. So -- and you'll continue seeing this from us. And yes, sometimes they ask us, are the clients there? And I would say, some of the clients are, the most sophisticated are there, some others are not. But we want to make sure that when the clients are there, we are ready with all our data, all our analytics, all our agentic workforce ready for them to implement. So I would say it's about the speed of embedding and making sure we keep that momentum, right? I think in this market, you cannot say -- and you asked me at the beginning, Andrew, you said, well, you're going to continue all this work with Claude, AWS, Microsoft, absolutely, right? You cannot skip a beat here because then you have the risk of not being in the play when a customer is going to finally embrace -- sorry, yes, it's going to start their GenAI journey. And I think the second we talked about it, right? The second is we need to make sure that we're protecting our IP. And that's why we're so laser-focused in the type of engagements that we sign with our customers and with the hyperscalers because we want to make sure that, yes, we are there, we are embedding again our connected intelligence, but we're also very mindful of retaining our IP and retaining the customer relationships, right? So we want to do it extremely fast, but we wanted to do it safe. I would say that is the approach, right? And that's what -- where we are very, very focused on making sure we make this a win.

That sounds right. Okay. So last question is really, Cristina, it's a summary question. So feel free to kind of bring together things that we've already spoken about. I'm sure you realize investors are sensitive to the AI risk to Moody's business. But why should investors see AI more in total as a tailwind than a risk to Moody's business going forward?

Yes, yes, yes. And maybe I said I think this is probably not only the summary, but it's probably 1 of the most or probably the most important question here. I think there's 2 scenarios, and we hear it every day, right? And I'm going to start by the not good scenario, what I would call the bare scenario, right? And the bare scenario goes a little bit like this. AI will commoditize the data, the LLMs will synthesize everything from public sources, the customers will no longer license proprietary data sets. We are -- all these hyperscalers are going to be able to automate all the workflows that we sell through the decision Solutions. So we basically are -- there's no data to sell because everything has been synthesized by LLMs, everything has been commoditized and then there's no workflows. And I think -- why I think this [indiscernible] does not stand is because basically this [indiscernible] case is misunderstanding what Moody's sell, right? We do not sell data. I'm going to go back to we sell decision-grade intelligence, data that is structured, that is governed, that is continuously updated, that is explainable, that is auditable, again, for decisions that carry legal, regulatory and financial consequences, right? Yes, you can go and scrap all the data of the world. But if you're going to have to present -- if you're JPMorgan and the mention in JPMorgan because it's your firm, Andrew, and you have to stand in front of a regulator, you're not -- and the regulator are asking, well, how did you make the KYC decisions? How do you make this creditors? How do you make all the decisions and all the reports that you have to do in front of the regulator, your answer is not going to want to be. Well, I scraped this data from here. I don't know if I have the necessary risk and yes, there was an issue in linking this data, right? You want to be able to say to stand and say that this was done by -- it came from a reputable source, right? So I would say that's basically -- that takes me then into kind of the the good scenario, right, the bullish scenario, which is, with GenAI, we not only have -- we have an amplifier for that data. The importance of good data, it's more important than ever, right? And then the data becomes more importable because you want to make sure you want to avoid the risk of hallucination. You want to have data that is sourced and auditable. But then once you start embedding that data in agents, switching that data becomes extremely painful, right? So the data is going to become more stickier, not only there's increasing -- there's an increased demand for data, but then as you embed those data in your agentic work flows and as you embed your data in those automation workflows, it becomes more embedded. So there's basically -- as more agentic workflows are adopted, Moody's becomes more deeply embedded in the decisions our customers make every day. And then there's finally the partner ecosystem through which we are reaching buyer personas we have never reached before, right? So I would say those are incremental relationships with incremental revenue, not substitutions, right? And that's, I think, the picture, right? First, we are not playing in places where you're going to be comfortable with great data. We play in high stakes where high stakes assistants are made. Our data assets more used and becomes more embedded, more stickier, more intelligent. And the third part, we don't see the hyperscalers as substitutions. We see them as amplifiers of our reach. And by that, we see that as mechanisms to deliver incremental revenue.

Well said, Christina. Go ahead, Shivani.

I think that's a great kind of note to end the call on. And I just wanted to thank you both for making the time to help us kind of educate our external stakeholders on Moody's GenAI strategy and the topics that have been top of mind for many investors and analysts out there.

Absolutely. Thank you very much.

Okay. Thank you very much.

Thank you.