#240 Generative AI in the Enterprise with Steve Holden, Senior Vice President and Head of Single-Family Analytics at Fannie Mae

Hi, I'm Maggie, VP of Curriculum at datacamp, and I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

Math at scale right, is what Excel brought us in some sense. Or from telephone to email to instant messaging, these are tech that have changed how we work, how we engage, but also it's accelerated our work. And I think that that's what this tech is going to do for us, but it's going to change our work in ways that we still don't really understand. And so our approach is to try to get this technology into the hands of our employees so that they can tell us what it can do in the context of how they're operating.

Hello everyone, this is Adele, Data evangelist and educator at datacamp. And if you're new here, Data Framed is a podcast in which we explore how individuals and organizations can succeed with data and AI. There's a lot of hype around generative AI today, and with this hype comes tons of expectations. Executives from every industry want to see the value driven from generative AI investments, but operationalizing genai and driving value from it is much harder. In reality, this is especially the case if you're in a highly regulated industry like financial services. So how do you go about driving value using generative AI within an enterprise setting? Enter Steve Holden Steve Holden is the Senior Vice President and Head of Single Family analytics at Fannie Mae, leading a team of data science professionals supporting loan underwriting, pricing and acquisition, securitization, loss mitigation and loan liquidation for the company's multi trillion dollar single family mortgage portfolio. He is also responsible for all generative AI initiatives across the enterprise. Throughout the episode, we talked about the use cases of genai in financial services, the challenges and opportunities of leading a genai program in an organization like Fannie Mae, how to get people excited about the generative AI agenda, how to build use cases responsibly and effectively, and a lot more. Before we move on to today's episode though, I wanted to say we're also hiring. It's been awesome seeing the growth of DataFrame the DataCamp Radar Conference, our live sessions and a lot of the things that we do on the media team and we want to take that to the next level. So we're looking for a couple of roles to join the team. First, we're hiring a data Evangelist to help me and Richie co host Data Framed, organize live sessions with awesome speakers and a lot more. So if you enjoyed this podcast and want to help make it better, this is a great opportunity for you. We're also hiring for a video creator. So if you love creating video essays about data and AI short form content, this role is for you as well, as both roles are based in our London offices, but we're super open to remote roles across the world as long as time zones relatively make sense. Now onto today's episode. Steve Holen, it's great to have you on the show.

Great to be here. I'm a big fan, very excited to join you today.

Adil, we're a big fan of you as well. So you are a Senior Vice President and Head of Single Family analytics at Fannie Mae. You are also in charge of Fannie Mae's Generative AI program and initiatives. So maybe setting the stage before we talk about Fannie Mae in specific, what are the opportunities for Generative AI if you are in a large financial services organization?

Genai brings it's actually sort of an incremental technological advancement that I think has unlocked a lot of really neat and interesting business possibilities that we're exploring. Things that we're all familiar with, like translation and summarization and content generation. But the things I'm most excited about, I put it into two buckets. The first is just a much more powerful search capability. And the simple example I use to talk about this internally is if you think about when you go to a website and you're looking at a frequently asked questions, and if you've ever had that experience of looking at the 10 FAQs and thinking to yourself, oh, those don't really apply to me, I actually have a very different question. What Genai does is it enables the context of your situation to be understood and then to be sort of melded with the content that's out there on the Internet or wherever you're searching, and the ability to go search that information in a much more smart and informed way to give you back results that matter to your situation. And to me that's going to unlock a lot of capabilities from a knowledge management perspective. Information doesn't always travel smoothly through enterprises and this is a place where, especially where you have good curated content. This is a way to really get that content into the hands of employees as they need it in a very efficient way. The second thing I'll point out is this power to interrogate data. It's really going to upend, I think, the way that our business teams interact with information. And so a way I talk about this internally a lot is we've made available a lot of reporting information, information about our business that gets updated automatically on a monthly or weekly or in fact daily basis. And our teams can self serve, they can go grab that information themselves. They never have to call up the analytics teams. But as soon as they have a question that hasn't really been contemplated before, that becomes a very bespoke and involved analysis. And I think what Genai is going to do is enable business teams to actually start to interact with the data directly and ask questions that we haven't really asked before without necessarily having to engage a technology team or an analytics team to prepare the work. And what that means is that business teams are going to actually have to become much more data literate. They're going to have to understand how data is, is a reflection of the business that they're trying to understand and its strengths and some of its shortcomings. But I think it's going to be a new world and I'm excited for it.

Yeah, I'm excited as well. And on the power to interrogate data, it's truly magical. As someone who's not a coder, for example, but has strong data literacy skills, has a coding background. When I work with a tool like ChatGPT or something along those lines with data, it really lowers down the barrier to entry, to working with data. Right. And I couldn't agree more. It puts a lot of pressure on data literacy and we're going to talk about that. But I think it's safe to assume, given what you just outlined, that there's a lot of nuances for using generative AI. Whether you're building use cases within your organization or you're rolling out a tool to do work. Right. And especially in spaces regulated in financial services, there's quite a few unique challenges to using generative AI. So maybe walk me through those challenges and how you see solutions to these challenges.

Look, this is a really important question. We deal a lot with models at Fannie Mae, as you might imagine, and these models are fundamentally different from what we're used to and in very specific ways. So let me outline a couple of things to think about. First of all, we're used to models that have measures of accuracy, think confidence intervals. We're used to models that are explainable. We're used to models that have, I'll say handfuls of parameters, not billions of parameters or trillions of parameters. And we're used to models that provide consistent output, meaning that if you provide the same inputs or close to the same inputs, you're going to get very close to the same outputs. And on every one of those dimensions Genai presents challenges. Lots and lots of parameters as I mentioned. You know, it speaks with the same level of confidence no matter whether it has results that are highly accurate or not so accurate. This is the concept of hallucination that's gotten a lot of press, the explainability. You know, there's a really great YouTube video video by Andre Karpathy on explainability and he talks about the fact that the engineers themselves that have built these models cannot explain exactly where these outputs come from. And that's quite profound. And so for all of these reasons there are challenges in how we introduce these models into the enterprise. By the way, I think all of these are going to be, they're all surmountable, they're going to get addressed. I have a lot of confidence based on just seeing how things are evolving that that will resolve these issues. But we're in the early days here. Some things that we think about a lot, making sure that we've got so called human in the loop. So there's an accountability there. Always back to a person. The computer is never responsible. It's the person that's using the algorithms. We do a lot of testing to make sure that specific to the use that this tool is being deployed, that it's going to do a good job in that specific narrow use. Obviously prompt engineering plays an important role. There's a lot of discussion right now out there about using LLMs in these sort of challenging situations, situations to sort of challenge the output of other LLMs. I think that offers a lot of potential. And so we've started exploring that and also chain of reasoning, making sure that when you get a result, you've got some data upon which those results are based that you can go and reference some sort of direct sources of information. And so understanding where that reasoning is coming from is important. So those are all mitigants that one might consider.

Okay, great. And there's a lot of challenges here to unpack. And you know, you've been heading up Fannie Mae's generative AI program. A lot of organizations today are thinking about how can we accelerate our generative AI capabilities. They have a generative AI lead. They have a generative AI program. And I think you have a very unique point of view. You know, we've talked a lot behind the scenes as well about kind of the weekly blogs that you write internally at Fannie Mae, kind of spreading the message, evangelizing generative AI within Fannie Mae. Look, we're going to definitely discuss those, but to first anchor our discussion, what do you think are the key components of building a successful generative AI program within an enterprise like Fannie Mae today?

I was asked to take this role on back in November of last year. So it actually hasn't been quite a year yet. And I spent a couple weeks really thinking carefully about the approach I wanted to take. And I largely concluded that we needed a set of guiding principles. And so I established three that I'll take you through. But what I'll say is, throughout this whole program that we've stood up, those three guiding principles will keep showing up again and again. And they've proven to be really important in our approach. And so they are as follows. The first is balance. And by balance, what I mean is I think there's a risk here of going too fast and of going too slow, especially for a highly regulated financial institution such as Fannie Mae. We've got to get that balance right. We want to make sure we're making progress, but we want to make sure that we're doing so in a responsible way. I think there are examples out there in the press that you can read about companies that went really fast out of the gate and ended up finding out that there are some real legitimate challenges with this tech. And so striking the right balance is important. And again, this is on both sides of the equation. It's not going too fast, but it's not going too slow. The second is around transparency. And you mentioned that I write a blog every week. That's one example. We also have a bi weekly knowledge share that we have our engineers come in and share their work so that as people are figuring things out, that information is finding its way across the organization. We're sharing any decisions that we're making in sort of very public ways. And then we set up advisory groups and councils so that knowledge flows through divisions in a very organized fashion. So transparency is important for us. And then finally I'll say, and this is probably the most important thing is humility. This is such a dynamic area. Literally every week something's changing. You have to understand that what you thought was true Yesterday may not be true tomorrow. And so we have a very much a test and learn approach. We're trying to figure things out. I was at the AWS conference back in November and there was a talk given by Jonathan Allen, one of their executives, who talked about one way versus two way doors. And I found that a really compelling kind of way to think about this, which is make sure that at these early stages that you've got this ability to back out of decisions. Don't make these sort of choices that are irreversible too early, because things are going to change and you're going to need that nimbleness, that flexibility as you figure things out as you move along. And so humility, I think, is a way to remind everybody to stay in that learning mode, to keep testing and figuring things out and not declaring really long term, final decisions about how things are going to turn out, because you will inevitably be wrong.

That's really great. I want to unpack a lot of these. So you mentioned balance, transparency and humility. And I love humility, and I'm going to discuss that in a bit more depth. But first, maybe let's chat about balance because I think moving too fast or too slow indeed is a major consideration for many organizations today. And it really depends on where you sit in the risk spectrum, the type of industry that you're in. So maybe Steve, outline for us kind of the key considerations or factors that determine speed to market, right. With these use cases, like maybe what are the different considerations that you have when thinking about the speed of execution for these use cases.

So what I would say is this tech is new and we manage a lot of different types of risk at Fannie Mae. And so probably one of the things that really is the determinant of the pace that we can move at is getting the governance. There are a lot of individuals within the company responsible for managing different aspects of our risk. And everybody's trying to figure out what generative AI means to our risk surface, how our risk surface is being impacted and changed. And so we're trying to engage our risk teams in an efficient way. And one of the ways that we accomplish that is through really establishing a prioritization for what we send through governance. And the idea here is that if every idea went through governance all at the same time, all of our governance teams would just grind to a halt because we couldn't possibly move, especially at this very early stage where there's just a lot of learning going on. And so we've been, I would say, fairly ruthlessly. Prioritizing the places that we want to engage our governance teams. And we have four criteria that we think about here. One is, well, we first start with is this a low versus high risk use case? Because we really want to be in the realm of low risk to get started. I believe that a lot of what we're going to do out of the gate here is very much internally focused and likely productivity focused, although there are some things that go beyond that. Second thing is uniqueness. What we don't want to do is have the same ideas being pursued in different parts of the company without that awareness of one another. And so what we're really looking for is, is this idea that we're going to engage our governance teams around something that we are already trying so that we can be really efficient on that. And the third thing is I call it tech stack aligned, which is if you're going to investigate some new sort of technology, let's make sure that it's nicely aligned with what we're already doing. Or if it's going outside of our existing sort of tech footprint or tech stack, that we're very purposeful, that sort of the juice is worth the squeeze that the opportunity there is really is really pretty high because we're going to have to engage a lot more of our technology partners if we're going to be able to pursue those ideas. So making sure that it's low risk, it's a unique idea and it's tech stack aligned. And then finally we ask that the teams are really clear on the business use cases so that we understand what the business outcome is that they're expecting from pursuing these concepts.

Okay, that's awesome. And a couple of things that we can kind of follow on this trail in the conversation as well is you mentioned that when building out use cases, it's really important that a lot of the choices that you make and it comes a lot on what you're saying, right? Like the use cases are low risk, it's tech stack aligned. A lot of the choices that you make make sure that they're also reversible because the space is dynamic and it's changing so quickly. How do you make sure that you're making choices that are reversible? Like what's an anti pattern that you need to avoid here?

Well, I think one of the places where we're going to see this play out is in which models we end up using. You know, we are looking at this sort of concept of a model garden. We certainly aren't sort of all in on one single model. We're testing lots of different models with a variety of capabilities capabilities and I think that ultimately our architecture needs to be able to sort of swap in and swap out those models depending on how the tech is evolving and what the needs are. You know, there's a lot of discussion out there now about small language models versus large language models find super interesting. There are times when we're going to want to go pretty deep into the tech stack, but things like doing rag implementations or fine tuning other places where maybe there's some capability that's just embedded in the software already that we checked out and seem to work pretty well. So you know, I think it's just this idea of not being too committed to any one approach and so just having an awareness of those decisions that you're making along the way.

And you know, a big part of you mentioned this creating use cases, especially in the early run on internal use cases focused on productivity. I assume it's really like linked to the idea of making sure that it's a low risk use case. But maybe walk us through a bit in more depth your thought process behind use case prioritization. How do you determine what is the use case that you want to pursue? So I'd love to kind of see how you think about use case prioritization.

So say a couple things. First of all, on the engineering side we've created, I call it centralized governance with federated research. So the idea is because again, engaging governance teams is going to be a big resource draw. We've created centralized areas for our engineering teams to operate. So the engineering teams themselves could be in lots of different parts of the company. But there's one place where they can come together and do the work and all that work is inventory. So that we're very clear on what's going on. There's user agreements and manager approval and things like that. But we've got a place for people to engage. We've separately gone out to every division in the company and we've engaged with them around this technology and what it is and what it isn't. And really sort of laid out in fairly straightforward terms what the business capabilities are that we believe that generative AI is unlocking. And then what we've said is you are the domain experts, you understand your area really well. Now that we've explained this technology, what are the things this tech could help you accomplish as you think about your business objectives? And so we sat down sort of in a follow up session and had them do that Ideation and really start to list out these opportunities. And it's not that we're going to go pursue all of them, but it's that we've got them thinking sort of differently in terms of the tech and what it can do. And we are organizing all of that information and we're in the process right now reflecting that back to management to say, here's what the teams are thinking about. Some of these things, by the way, are big ticket items. They're things that sort of have more enterprise type feel to them. Other things are much smaller, but in their totality they could have a pretty big impact. And so that's the process we're in the middle of right now. But what I'll tell you is that these sessions have been extremely engaging. There's been a lot of excitement across the whole enterprise and we're really excited to see what comes of it.

Yeah, I was about to follow up actually on the excitement element of it because what were your expectations maybe coming in before these meetings? Because I assume that there would be in any organization, some folks that are excited, some folks that are not excited about AI. Right. So I'd love to see how you manage these conversations and how, what were your expectations going in and how did they end up panning out in terms of like the level of excitement around these use cases?

You're absolutely right. We've got a group of pioneers who are out there really pushing on the tech. We've got a lot of people that are really interested in the tech but don't quite know how to engage in it. And so we tried to create mechanisms to enable people to kind of lean in and see what's going on. One of the things that we did really early on, actually in February, I announced that we were going to set up a old company internal conference. And I announced the date, which was in the middle of May. And by doing that, I signaled to our engineering teams, look, if you've got some concepts that you're thinking about or you're working on, here's a date that you can aim at and see if you can bring this to fruition and we can start to share it across the company and show, you know, in a sort of spirit of transparency, as we talked about earlier, we can help our employees see what it is that we're really thinking about here. And so we did that. We had about 50 concepts, I think, that got registered on our sandboxes. People were actively working on about 17 of them. By May 15th was sort of ready for prime Time ready for demoing. And so we ran that conference. We had about 850 people show up in person in our Reston office in Virginia. We had about 1300 people join virtually. We had some other sessions. We brought in some technology companies, some vendors who have LLM model that they could put on display in a developer lounge so people could just sort of look, touch and feel the tech and get a sense of it. And it was, I think, by all measures, a huge success. There was a lot of energy and enthusiasm, and we brought a lot of transparency to what we were up to. There wasn't sort of this thing going on behind the curtain. People could come and see what was going on. I would say, you know, the other thing I'll just mention on this front is you're right. I think there is, you know, on the one hand, excitement. On the other hand, there is concern. And my view on this is just like all other disruptive technologies that have come through over the last many decades, the one thing that I'm pretty sure about is it's going to accelerate our work. Things are going to move a lot faster. The examples I often give are from pen and paper to hand calculator to Excel spreadsheets, right? Math, you know, math at scale, right, Is what Excel brought us in some sense. Or from telephone to email to instant messaging. These are tech that have changed how we work, how we engage, but also it's accelerated our work. And I think that that's what this tech is going to do for us, but it's going to change our work in ways that we still don't really understand. And so our approach is to try to get this technology into the hands of our employees so that they can tell us what it can do in the context of how they're operating. And, you know, I always tell people I'm an optimist, personally, I believe that the parts of your day that are probably more mundane and less interesting are the parts of your day that. Jenny, I can come in and actually make some of those things that you're doing a lot easier. That's my hope. That's my expectation. And some of the engineering work that we're going to be doing internally focused is going to be looking at things like that. We talked about knowledge management. I think that that's a huge opportunity for us. If you spend a lot of your time trying to understand a certain policy that you're trying to deploy in the context of where you're working, that information is going to be easy to get, easy to Understand easy to make use out of those sorts of things I see in the days ahead.

Okay, I couldn't agree more. Especially on. I'm very excited about Generative AI taking the parts of my day that are the most mundane away. And Steve, you shared a lot of the blogs that you write like the culture of transparency, creating a conference internally, and I think that's really wonderful. And a big part of driving success with Generative AI is harnessing and building that excitement within the organization, getting people to become believers, to be creating a decentralized network of evangelists within your organizations. Maybe walk us through kind of your methodology for driving excitement a bit more. How do you drive transparency? How big of a focus it is? I'd love to see your philosophy on driving excitement. As a Genai leader.

I can promise you in any organization, there are people throughout your company who really want to get engaged. And what they're looking for is just a little bit of structure they're looking for. How do I take my energy and point in a direction that's going to be aligned with where the enterprise is trying to go? And so I view my job is to send those signals and help those teams take that energy and use it for productive good, move it in the direction that we're trying to go. Again, I see with Genai, there's a lot of opportunity and there's a lot of risk and the trick is going to be to strike that balance. And so you've got to find those people, that community. And it's not a sort of an organizational, direct reporting, centralized structure. It's there in lots of places. And so blogging is a way to communicate. And you know, I very purposely don't email my blog out to a lot of people. I put it out on a website, on an internal website, and I have a link and people who want to go read it can go read it. And I have regularly about 300 readers every week. But those are the people that I want to talk to. Those are the people that really want to engage around this tech and want to understand it. And so they sort of opt in as opposed to of me sort of pushing this out and not really knowing who I'm talking to. And so now as we start to engage the community, I know who I need to talk to because I know that I've got a subgroup in the company who are going to help drive this tech forward.

You know, you mentioned here the folks that are necessarily that there's fear within any organization today by AI taking jobs about, you know, The AI risks, et cetera. You talked a bit about this, but I'd love if you can expand a bit more. Right. Like how do you make sure that you're communicating a message that inspires, given the fear around AI?

When we did our conference, one of the things that we did was we didn't just do concepts, we brought in specific teams in the organization where we felt this tech had a pretty important place to play. And so, for example, let's say that we did nothing. Let's say that we sort of sat back and this tech sort of evolved out in the sort of ecosystem, but we didn't really lean into it. From Fannie Mae's perspective, that creates its own risk. And the risk that it creates is that fraud or cyber attacks are going to get better with this tech. Others outside of our company are going to be using this technology also. And so I think expressing the importance of all of us sort of leaning in and trying to understand it, because it's literally affecting all of us. I think it's just the point that's really important to communicate. And so the way we did that was we invited business leaders across the company, both, for example, our chief economist came in and gave a talk about the disruptive nature of the technology, the economy, broadly. Fannie, may we sit on 17 million single family mortgages. Right. And these are everyday people across the country who are employed in a variety of jobs that are going to be impacted in one way or another by this technology. We brought in the cyber teams to talk about cyber risk. We brought in the fraud teams to talk about fraud risk. We brought in our HR team to talk about Persona based curricula in an era of advancing AI tech. And so the idea here is how do we create the mechanisms for our teams to lean in and really think about their own skill sets as AI is going to change potentially the way they engage in their day to day. And so really being open about what we're up to and sharing things as we know it, and then again, putting the tech in the hands of our employees and having them lean in and having them tell us what they're seeing, I think all the ways that we navigate this, we don't have all the answers. I believe we can't ignore this technology. I think there's a lot of risk in ignoring it, as I've mentioned. And so I think we just have to embrace it in a responsible way.

Yeah, and this is a great segue to my next question. You know, embracing generative AI in a responsible way and Kind of making sure that you're aware of the risks and you're able to navigate the risks. One thing that you wrote in one of your articles internally that I was privileged enough to read, that you shared with me, how governance and financial services is a competitive advantage when it comes to AI. So maybe walk us through your thinking a bit more depth here. I love this notion of how governance is a competitive advantage, so I'd love if you can expand that notion a bit more.

Yeah, I mean, as I mentioned earlier, the whole nature of the risks we're managing is changing. And so our governance teams are all learning this as we are learning this. We've got a set of policies and procedures that were written in a time that predates the arrival of this technology. We have new regulations that are coming from our regulatory environment. We have a set of AI ethical principles that we've embraced as a company, which will actually largely look like what you'll see from any company embracing AI responsibly. Things like transparency and accountability and privacy and reliability, et cetera. So these are all pretty standard. But in terms of a competitive advantage, what I would tell you is all of that stuff is really in service to making sure that the things that you end up doing are done safely and responsibly and that the things that are beyond the pale in terms of risk, you don't do. And so how do you navigate that efficiently? How do you make sure that those innovative ideas that have real business impact and can be embraced responsibly, make it through the system whilst you stop those things that really have some challenges? And so the way we thought about this is we have a group that we stood up that's cross divisional and cross functional. So a lot of different expertise has come to the table. And what we're really doing is saying, how do we engage our governance systems efficiently? And what I will tell you is the first tools that go through the system, it's difficult, right? It's challenging because we're all learning, but having that mindset of continuous improvement, doing those after action reviews and saying, how can we go through this process next time a little bit faster, whilst not compromising any of the safety that we're embracing. So that's how we're thinking about it. We've got a system that we've stood up and it just gets better week over week over week. And I believe that we run this through a few cycles and we're going to have a very efficient way to start to bring this tech in that's going to be responsible and safe.

I love that. And I think it's a great segue as well to switch gears and talk about the skills transformation agenda here because having folks within the governance program needing to know how important key AI concepts, how AI use cases are being built. But you also mentioned at the beginning that one of the main use cases of generative AI within an organization like Fannie Mae is the power to interrogate data at scale, which really puts pressure on data literacy. Maybe how big of a moment is this for skills transformation overall? Whether the need for AI literacy skills to be able to understand how these systems work or for data literacy to be able to understand how to work with data at scale as well. So I'd love to understand your perspective on the importance of skills transformation today and how do you think the skills agenda will evolve over the next couple years or so?

This is a super interesting question and I'll tell you some things that are on my mind. Definitely data literacy, I touched on this earlier, but this idea of we sit in the secondary mortgage market, we actually don't directly interact with borrowers, but we have a huge stake in understanding that borrowers who take out mortgages in the US market are being set up for success, that these are sustainable mortgages, that these borrowers are going to be able to be successful at home ownership. And so we understand that through the data. And so understanding the information content of that data, understanding how to navigate the data, that's becoming increasingly important in an era where again, business teams can self serve. But I would say that beyond that, things like curiosity are going to become really important. Asking good questions and then knowing how to get those answers, because those answers are now going to be at your fingertips. I think those are going to be important. This prompt engineering skill that's gotten a lot of discussion, I think that's going to be a legitimate skill, both in terms of asking questions the right way, but also because these tools are expensive. And you have to think about both architecturally, how you design the tools in your ecosystem, but then how you train your teams to engage around them in a way that manages costs responsibly, that get answers that are coherent and accurate and usable. So those are some things that I think about. I also say that aspects of what we do are becoming more automatable, but there are elements of what we do that in my view are not replaceable. And one of those things is sort of the human element of this, right? The trust, the rapport that you build through human interaction. And I think one of the huge opportunities with this technology is it can free us up so that the parts of our interactions that are automatable can happen in the background. And the parts of our roles that aren't automatable but are incredibly important can really be focused on. And so, for example, there's a lot of discussion in the Genai space about how this technology can transform call centers. And think about what a call center is. You're having an engagement with another human being and it might be a variety of settings. It might be someone who's in a mortgage who's having trouble making a payment. It might be originating a mortgage where the borrower is being evaluated for credit. And if you think about the ways that humans can build trust and that information can be captured in a more automated way in the background, I think those are opportunities that I think present sort of interesting ways that we're going to evolve in terms of how we operate.

Okay, great. And I want to unpack maybe a few of those notions that you just mentioned. So doubling down on data literacy, for example, as an essential skill when working with data.

Right.

Like you mentioned that Fannie Mae use case of working with borrower data and mortgage data, I think is so foundational. Do you think that technical knowledge will be important or will it be subject matter expertise about business data that will be important here, or is it both? For example, for me, when working with a tool like ChatGPT, knowing Python and knowing that, okay, I'm able to edit that particular line of code, it's quite important. But also knowing the data is important, like how, how important do you view these skills evolving in the next couple of years? I'd love to also understand.

I think this is going to be an area of active debate and here's why. The idea that some of the skills that have been typically thought of as technical skills and professional disciplines, some of those things can now be handled in certain situations through tech. And so there's this discussion around low code and no code, right? This idea that you don't need to write code anymore because you know you can interact with the system. And the question that I think this raises is just because you can, should you? And I'll give you two examples and I don't have the answers. So let me set it up front. I don't have the answers, but let me give you two examples that I've thought about. So one is we've introduced GitHub Copilot into our environment here at Fannie Mae. And so we have been gradually rolling out to Our developers, we have a lot of developers here. We have, I think, 4,000 employees who would be classified as a developer of some sort of software engineer, a data engineer, a data scientist. And so we've made this capability available. It's a co generation sort of support tool. And you know, when we were first contemplating rolling this out, one of the questions that I was challenging the team with is which job functions are we going to make this available to? Are we making this available to professional developers as a productivity capability? Are we providing this to all employees as a way for anybody to start doing development work? And we very squarely landed in category one. We felt that this was a place to enable productivity for employees. The second place I see this showing up is in the data science area where you've got a lot of pretty advanced data mining capabilities that you can really open up to almost anybody because the barriers to entry have come down to almost zero. You can start asking questions, hey, go run me a random forest model. And you might actually not know what even you're asking. And again, I think this presents a lot of risk because you have to understand how these models work in order to understand some of the risks that you might be introducing into your environment by using these models. And so again, you know, I think that sort of these professional designations around data science and statistician and econometrician, these aren't going away. These are going to be incredibly important. But this is a frontier that's going to have to be navigated because these barriers to entry are coming down. And so we're going to have to be really purposeful about how we handle that.

Yeah, and I love that example of GitHub Copilot within the organization. I think providing a tool like GitHub Copilot to folks who may not necessarily have the skills to challenge the output could be not the most ideal situation. In fact, trust in data down the line. Right. Like if people don't trust data because of hallucinations. So you've been doing this role for a while now, being the generative AI lead at Fannie Mae. In a lot of ways you are building the plane as you fly it. Right. So what have you learned about leading a generative AI program or about leading the generative AI program, Fannie Mae, that you wish you knew before you started?

I think I knew this intuitively, but it really played out, which is it is really easy to demo something that captures people's imagination. You can get people really excited about something with this technology. You can set up very quickly that's actually not that hard. What's really hard is to scale it in a responsible way across an organization. That's really hard. And I think that reminding people of that over and over again is important so that there's this awareness of what it really takes to deploy this technology at scale at a company like Fannie Mae in a way that's going to harness that productive opportunity, that do so responsibly and in a way that doesn't end up creating unnecessary risk.

A couple of final notes. What would be advice you would give other leaders who are in a similar boat or position?

I've always been a big fan of both taking a real test and learn incremental approach. And that's, I think, paid off for us thus far. So I would say that that's worked really well for us. And then also sometimes there's a tendency for control, trying to really sort of control sort of every aspect of a program like this. And I think that's folly. I think that if you want to really create an innovative environment where great ideas are going to have their moment, I think you have to again put the tech in the hands of employees. You have to really set up some structure and some basic programmatic capabilities. But then really get out of the way and let the people who know how to innovate go innovate and show you what they can do. And I think that if you can do that, I think that you'll be off to the races and you'll do some really interesting work.

Yeah, it's not only folly, it's also really good for your mental health. And maybe one final note as we wrap up, Steve, what are trends that you're excited about in this space?

Look, I go back to some of the modeling comments I made earlier. I do think that explainability is going to improve and I'm looking forward to that. I do think that some of the issues around sort of confidence in the output and being able to measure that, that's going to get better. I look forward to that. But I think honestly starting to put generative AI into our workflows is something that I think is on the roadmap ahead. And I think that's where we're going to really start to see the Valley show up. There's a lot of talk out there about agents and agentic framework and we started exploring that and we're seeing some early interesting opportunities there. So as this stuff starts to come to fruition, I think it's going to become real for people. And so that's what I'm excited about.

Okay, that's awesome. Steve Holden, thank you so much for coming on Dataframe.

Thanks for having me.