Why Your AI Strategy Is Really Just Your Data Strategy

Christian Ward:

The problem with AI is you could put garbage in and get a masterpiece out.

 

Saket Saurabh:

Yeah, we call it slop.

 

Christian Ward:

If your data truly is a moat, and it has either competitive intelligence or other tools, that’s one thing, but you’re going to have to let it out. You can build whatever dashboard you want in twenty-two minutes, only if you’ve gotten your data strategy right. As I like to say, your AI strategy is really just your data strategy. I’d like everyone to take out your credit card receipts for the last two months, then look at the sun and see what days were really sunny. Everyone’s out there trying to build and train their own models, I think that’s adorable, that is not where I’d be spending my time.

 

Saket Saurabh:

Hello, everyone, and thank you for listening to another episode of Data Innovators and Builders. This is your host, Saket, and today I’m speaking with Christian Ward, Chief Data Officer and EVP at Yext. Christian, thanks for chatting with me today.

 

Christian Ward:

Thank you, Saket. Looking forward to it.

 

Saket Saurabh:

Very good, exciting. So why don’t we get started right away with a little bit about your background, and what Yext does? Let’s set the stage for your work.

 

Christian Ward:

Sure. I’ve been in structured data since I was twenty-three. I started my first data company on Wall Street when I was twenty-three, so I’ve been doing this a long time. First it was in finance, then it moved into more of what Yext is today. So I’ve split my time across my career, but honestly, the most consistent focus I’ve had has been helping people tell the right stories with their data and get the most out of it. I’d say right now is probably the most exciting time in both data and innovation that I’ve seen since maybe ’99, since the internet, which is, we’re seeing this opportunity where the data you have, through the power of AI, is unlocking all these new areas.

 

At Yext, think of us as the digital visibility solution for any brand, any business. When you go into Google Search, Gemini, OpenAI, Claude, any of these platforms, we make sure that the data, particularly the citations, everything they’re referencing, from your website to your Google Business Profile, any fact or objective piece of knowledge about a brand, a business, or their locations, we store that in knowledge graphs and synchronize it out to all those endpoints, hundreds of them, around the world, in real time.

 

The way to think about it is, if you’re a major brand, and we work with most of them, let’s say they want to close all their restaurants in a country for a holiday, they enter that once in our system and it synchronizes everywhere, so you don’t get bad information about the brand. It’s really a knowledge-graph-based system, but it’s linked to things like websites, social reviews, all those sorts of areas.

 

Saket Saurabh:

That’s a great example, actually, and it immediately highlights the challenge of data being fragmented, and the truth being hard to figure out. That’s the data variety and data siloing problem we see quite a bit out there. You mentioned knowledge graphs, and that’s a very hot topic right now, leveraging knowledge graphs in AI. Tell us a little bit about what’s happening in your world around the application of AI in your work.

 

Christian Ward:

Yeah, absolutely. It’s kind of a strange world when people say knowledge graphs are a hot topic, it’s a little disturbing when you think about it. Basically, a knowledge graph is really a series of databases in three dimensions, where relationships are just as important as the entity data itself. I might be a doctor who works in a hospital and does this type of surgery, but only takes this type of insurance, and has availability next Tuesday. That sort of linkage is really easy for AI to get a lot of value out of.

 

Whereas if I tried to explain that same thing in paragraph form on a website, it’s much harder for AI to ingest and compute. What we’ve found is that Google, in particular, and a lot of other sites, including AI systems, like knowledge graphs. I think the reason it’s become a hot topic is that people have found that through, say, markdown files with linkages, you can mimic these graph designs and add much more context without a ton of compute or memory usage. All of that ties together, and we’ve been in this business, Yext actually officially started almost twenty years ago, so we’ve been distributing data like this for a long time.

 

The real difference is that, over the last twenty years, it was principally around Google and their own website information, their own knowledge. What’s happening now with AI is you’re seeing the synthesis of all this knowledge in real time. We’re moving from a world where everything had to be crawled, indexed, and ranked, which was great and served us well, into a world where AI wants recency and frequency, something we’d never live without on Wall Street with our stock quotes, and now they want it for appointments too. We’re moving into an era where AI is demanding more information. There’s almost no world I see where AI or search wants less information over the next five years, it’s going to skyrocket.

 

So what we focus on with customers is the customer journey: what are they asking, and when they ask it, where are you showing up? And when you show up there, how can you control that, or at least make sure the content is correct? A lot of academics have looked at the problem of data quality over the years, and there have been several theories that really capture most of it, and it always comes down to a lot of the same signals: what’s the source of the data, what’s its authority, how consistent is it with other authoritative sources, when was it last touched or updated, is there a decay factor, and how deep is it, how many sources, how much data does it hold.

 

When you add all of that up, that’s a lot of what we work with clients on, making sure you’re doing that part really well, because AI is synthesizing all of it, but because AI is probabilistic, if I said to the audience, “kill two birds with one stone,” everyone knows that phrase well, because, probability-wise, it’s shown up so many times. Think of that in terms of your data. If I look at your business, open on this day, in this town, and you’re this doctor with this insurance, and I see that over and over again, and when it changes, it all changes at once, that gives confidence to the accuracy of the signal. That’s the whole game to us.

 

Saket Saurabh:

That’s so well explained, thank you, Christian. As I look at the evolution of AI, especially for enterprise use, two years back it was, “it’s great for chat, productivity, marketing, writing emails,” but it can’t really be used in the enterprise, enterprise is complicated and data is messy. Then, toward the end of last year, models improved, AI started to work for software engineering, and it became clear it actually can work in the enterprise. Now, when I talk to data and AI leaders, I consistently hear that they feel their moat is the data they have, how well it’s organized, and how much context and knowledge they’ve built around it.

 

So the question for you: you’re talking about twenty years of data built at Yext. Tell us how you see data being the moat, or not, in your case, and how you perceive AI’s impact on that.

 

Christian Ward:

Yeah, it’s interesting, we talk about this a lot, both internally and with clients. When you think about data, a lot of people are realizing this “garbage in, garbage out” thing, and, humbly, most humans can spot garbage, they have a pretty good sense for it. The problem with AI is you could put garbage in and get a masterpiece out, it looks.

 

Saket Saurabh:

Yeah, we call it slop.

 

Christian Ward:

Yeah, and it’s very hard, some slop, I think most of us can spot now.

 

Saket Saurabh:

But.

 

Christian Ward:

But it’s getting to the point where, if you’ve had a chance to try Fable, once you get it writing, you can’t tell. It’s actually quite incredible. To your point, I think the AI we were trying to work with two years ago, the adoption curve looked huge in terms of scaffolding, trying to get it to not break, or use the right data sets. Now the adoption curve is shifting, because if you try AI today and your company hasn’t really used it, the AI you’re trying today is significantly different from the AI Saket and I were working with two years ago. What’s starting to happen is it’s opening people’s eyes: wait a second, if I really do have great data, what are the moats in terms of what that data can do?

 

And if I combine that data, say, with my customer data, or weather data, or population demographic data, you start to see a whole new way of thinking. What used to be the purview of your engineering team and data science team, now we have CMOs and CEOs on the weekend pulling their data assets into Claude or ChatGPT and starting to see they can pull insights out of it. So I think probably the best unlock most companies are realizing is, if your data truly is a moat, with competitive intelligence or other value, that’s one thing, but you’re going to have to let it out, in some format, to give value to the market.

 

And then the other side is, once the data tells you to do something, do you also provide a service that gets it done? That’s why you’re seeing all these acquisitions, people buying data companies and merging them with a CMS, or with Slack, because what they’re recognizing is that intelligence alone isn’t the only answer. It’s great to have the intelligence, but what can you also do, via API or structure, that actually enacts it and affects something in the real world? We sit at a pretty nice intersection of that.

 

The way we look at it is, the more data we can gather, the better. We acquired a company ourselves last year that looks at all the data about a business. Say I’m the company powering a brand, it could be any brand in the world, let’s say Tom’s Pizza, and Tom’s Pizza has five hundred locations around the world, from Italy to California. When we look at Tom’s location in Newark, New Jersey, we also pull in the fifty other pizza places Tom’s is up against and analyze everything they have. We build a huge data set so we can correlate what’s working in what market, because what you have to do to show up in Google or AI in Roswell, Georgia is very different from Palo Alto.

 

So we advise most of our clients: you have a lot of data, everyone’s been telling your chief data officer, who you never listen to, that you’ve got to get it out of the silos into a play area where it can be utilized, a Snowflake, whatever you want to use. But then you start bringing in third parties, like my company or others, and now you’re playing with data in a way you could have only imagined. It almost seems like sci-fi, but it’s really amazing how fast you can get insights you never had before.

 

Saket Saurabh:

That’s well said, and actually very relatable for me, because I started Nexla with the idea that I want to help companies bring in data from their customers and partners. That’s why Instacart became our very first customer, getting data from merchants, because it’s hard to get that data in. If I break it down in my head, I see three unique pieces about data in the AI world. One is being able to tap into unstructured data and documents, which was much harder. Second, with the knowledge graph, instead of a similarity search on vectors, you have a very deterministic relationship between things, which is super helpful for getting more deterministic answers. And third, in that knowledge base, there are these gaps, and if you can find the right data elements to fill them, the probabilistic model can start to perform better than it had before.

 

Christian Ward:

Right.

 

Saket Saurabh:

Okay.

 

Christian Ward:

Yeah, I think when you tie those three things together, what you end up with, you’d normally be beholden to an engineering team or a software platform to pull that together, but in some ways that’s already limiting. For example, and this isn’t to pick on Tableau, we love it, we love BI, we love Power BI, it’s more that, where they’re all going, and where you can go on your own now, is when you realize your data set is now available to other members of your organization, and you’ve done the crosswalk linking. So I can link through time, geography, store, retail, product, any of it, the graph gives you the easy cross-linking to infer a lot more about what’s going on.

 

That opens this world to everyone: you can build whatever dashboard you want in twenty-two minutes, only if you’ve gotten your data strategy right. As I like to say, your AI strategy is really just your data strategy. I appreciate everyone trying to build and train their own models, I think that’s adorable, that’s not where I’d be spending my time. Get your data right, and test it on all these frontier opportunities, they’re going to get better, and at some point they’ll asymptote, for certain reasons, and then you can use whatever’s best for you at that time. So many people jumped in thinking, “look at our data, let’s train on it,” and that’s not going to work, because you don’t have enough.

 

But if your data is in a great state, and you bring in one of these other tools, you can really get a lot out of it. And I don’t mean eighteen months from now, it’s almost immediate. When we show our clients how they can do this, I should mention, Yext opened up an MCP to all our data for our clients, so they can literally skip the software, go right into Claude or whatever their foundational model of choice is, and start asking questions. The neat thing is that’s now inside their firewall. All their payer data, insurance rates, we do a lot of healthcare finance, really private stuff, they’re pulling in our data to build an understanding of the competition, where they should build a new branch, what services they’re missing in a market. That’s the power of this, if you get the data right.

 

Saket Saurabh:

Okay, and I think that’s super powerful as you bring more data together. I do have a question, though, because one of the challenges now, in this world of AI-powered use of data, is that we can be a lot more exploratory. Whereas in the world of analytics, it’s more like, “I want this dashboard,” you build that, and you’re kind of boxed within it. But now it’s exploratory, I can ask a follow-up question, keep digging, and end up somewhere I hadn’t thought I’d start out. How do you make that possible? MCP servers are great, but when you get to an end user who has a lot of internal data, and is also connected to Yext data, exploring and finding the real nuggets of opportunity isn’t easy. Is it a conversational approach? Is there a discovery process that happens automatically? How does that work?

 

Christian Ward:

Yeah, it’s a great question, because I think it’s the difference in how many people, when we meet with them, are still treating AI like a chatbot. They’re thinking of it as a finite game, like chess, they move, then the AI moves, then they move. It’s not chess, it’s a continuous, infinite game, the whole point is to keep playing. Discovery is about continuing to play. What you actually want to teach people is inverse prompting: at the end of everything you ask an MCP through an AI, you say, “what am I missing? What else do you need to know? What data set did I not include that might add a third dimension of understanding?” Once you start realizing that.

 

And again, I’m not talking about the free models, they don’t do any of this well, you’ve got to be paying for some of these premium models. It’s worth it, even if you just want to try it. But once you do that, the dialogue becomes much more interesting, in terms of discovery. To your point, with some of my own systems, I might have twenty different MCPs hooked up, and what it’s able to pull in is things like, when was the last time Saket and I spoke to each other, on the calendar, when was the last email, what topics were discussed, how does that relate to my own blog writing, or to Saket’s last five interviews.

 

You start to pull this together, but you have to understand, I call it the tyranny of the blank box: you don’t walk into a blank chat box and expect it to work, you need to invite the AI into the discourse early, and it will help guide you toward discovery. Now, that doesn’t solve the problem of, if it’s so big, even the AI may not get you there. Generally, that’s a business acumen thing, you should have a perspective, a theory, a thesis, like we’d have in academia, and then you should go pursue that hypothesis, you don’t just run all the tests and let it hand you your hypothesis. So it’s about focusing on what areas you want to improve.

 

For many businesses, that might just be automation, or better billing, it could be a very basic process. But ask the AI to help you with the data you’ve already given it, and then what data you should maybe append to it, and you’ll start to see some real magic happen.

 

Saket Saurabh:

Yeah, I’ve seen different paradigms people use with AI. There’s “AI is like your intern,” you give it a task, it does it, then you check on it. But then there’s what you’re describing, AI as a thought partner, you work with it, dialogue back and forth, maybe ask it for ideas, “what else could I be doing?” That’s a process where you’re learning along with it, not just assigning a task to an intern and waiting for output.

 

Christian Ward:

Yeah, I think the intern framing was true two years ago, and now I think we’ve reached parity, certainly with every intern I’ve ever had, frankly better. So what you’re saying is absolutely correct. For example, with my children, they’re in their late teens, in college, and they’ll come home and we’ll be talking about something, and I’ll say, “let’s debate it,” and I’ll hit record, record the debate, and I won’t have the AI in the debate itself, I want to talk to my human child for a while. We’ll have the debate, I’ll send a forty-minute transcript to the AI and say, “what do either of these people not understand about this topic? What should they understand better?” You’re using it as a thought partner, in a really fun way.

 

For my kids, I want them using AI, not to write the papers they need to write, but to debate what tariffs are, as a high schooler who barely knows or understands economics, it’s a phenomenal thought partner. And while I’m using high school as the example, most of us can benefit from that in our professional lives, well beyond high school.

 

Saket Saurabh:

Yeah, yeah. Well, AI, from an intern, to a thought partner, to what, an overlord and master?

 

Christian Ward:

Yes, look, there’s a possibility, for which I’ll continue to be optimistic. For everyone out there worried about that, do you really think the elected humans have done such an amazing job that we shouldn’t at least try this?

 

Saket Saurabh:

I showed my kids The Matrix this summer, some sentimental movies, so, okay, you should see what future was expected according to sci-fi twenty, thirty years ago. One really cool thing you’d worked on before was this NLP tool at Ocean Research, doing real-time sentiment analysis from TV closed captions, which is super interesting. I’m curious, tell us a little bit about that, and how some of those experiences helped shape the CDO you are today.

 

Christian Ward:

Well, a lot of times, when you’re looking around at the world, one of the real benefits I’ve had is I was always taught to try to see everything from other angles, to change your perspective as often as you can. What you start to realize is almost everything in our lives is a signal, in some way, shape, or form. Your body’s constantly sending you signals, the weather is a signal, how you feel, what you’re actually doing. When I worked on Wall Street, I was very fortunate to work in what’s called independent research at the time. Independent research is where I got my start, things like satellite imagery of malls, you’ve heard of this, checking the parking lot to see how many cars are there before Black Friday, to see whether they’ll hit their numbers.

 

There are all these alternative signals, not the financials of the company, but how often is it talked about in the news, how often is it mentioned. The idea was to take a very simple, almost Occam’s razor approach, and look at how many times someone mentions a company on a given day. This is way before large language models, we’re talking twenty years ago, when closed captioning came out. We’d build a complete data set of what’s being talked about.

 

As we did that, we saw real correlations, not just that someone like Jim Cramer is jumping up and down, yelling a stock name seven times, but we could model acceleration, seeing that a company is being discussed in a window of time three standard deviations above its normal noise. Again, it’s just signal processing. We ended up selling that company, the whole goal was helping financial advisors manage their relationships, because we could tell them, at seven in the morning, which of their clients were going to call them, based on the portfolio of investments they held and what was discussed the day before.

 

We literally knew, “you’re going to get a call from Mrs. Johnson, and she’s going to be very upset because Jim Cramer talked about her top holding for two minutes.” You could build that out. It’s similar to what I’m saying today: if you’re not thinking about your visibility, how you’re showing up in AI, through the lens of the consumer memory that AI now has, which search never really did, memory, search, your competitors, and your location, wherever the business is, you’ve got to see it from all those angles, because that’s what’s going to let you understand how to pivot your business, or how it’s being talked about.

 

I’ve always looked at the world that way, it’s a ton of signals, how do you process them, how do you prioritize them. The fun part, for me, is that AI is making that much easier than it’s ever been.

 

Saket Saurabh:

My first startup was a company called NearbyAd, back in 2009, and we were looking at location as a prominent signal for mobile advertising. 2009 was really early in the apps world, and a lot of the ideas we were working on then, like how, if you get the data right in terms of patterns, you can identify things, were hard machine learning problems, and we did solve some of those. But was it just not possible to solve them at the velocity that’s possible today? How do you see that, by the way? Do you see traditional machine learning approaches continuing to run side by side with LLMs for specific problems, or LLM-based approaches taking over pretty much everything?

 

Christian Ward:

I’m a firm believer, if you follow the debate between the large language model community and, say, Yann LeCun and others who are really focused on world models, here’s the reality, I want to advise everyone, don’t get married to any of this. Date for a long time, this isn’t marriage material yet. You really want to try different approaches. Classic machine learning is still somewhat of an undisputed champion in some correlation studies, it’s really good, it’s easy to understand, easy to convey, easier to build into products. The moment you enter probabilistic territory with LLMs, it’s really hard to explain exactly what it did, and therefore you can’t build software to the same level of deterministic accuracy. Each one has its own flavor and purpose, and I think that will continue.

 

Now, the really exciting stuff, if you’ve had a chance to try Fable, and to some extent other labs have been trying this for a while too, a lot of people have been trying to optimize token spend by model. It’s a similar thing: I really just want Sonnet 4.5 doing this for me a thousand times, and it’s cheap enough to do that, or I want a Gemini Flash. I think where we’re going next, as with most things, is several models working together. There was a book by Pedro Domingos, he’s pretty outspoken on AI, he wrote some brilliant work on the five different theoretical schools of AI and machine learning, it was called “The Master Algorithm,” a very good book, this is going back ten, fifteen years. He predicted it would be a combination of things, from the biological to the neurological to classic machine learning to Bayesian approaches, these different families of models, and his whole point was, it’s not going to be one, it’s going to be a master of models, an ensemble, where you really have the ability to use all of them. So again, I’m not getting married to any of these, I think they’ll continue to evolve.

 

For most businesses, probably the proper path is having at least a team, or a team member, or a consultancy, or a product, that helps you understand how to optimize that going forward, because cost is a very attractive, and often wrong, partner. So often I see people say, “well, we run the cheapest model because we just run it five times over,” and I’m like, yes, but that five times isn’t even close to the value the full model would provide. I see what you’re doing, but don’t let cost blind you to what your real efficiency, or proper outcome, actually is.

 

Saket Saurabh:

Yeah, we’ve been through different phases of that conversation. At some point it was token maxing, “who’s spending the most, that’s my most productive employee.” Then what’s also happened is token consumption in certain models has increased, and people aren’t sure why, maybe there are levers companies are pulling as they head toward IPO or whatever’s happening. There’s certainly anxiety about cost. But you’re right, it’s not just about cost as one parameter, it’s about ROI, and ultimately the outcomes, when they’re valuable enough, you can justify the cost. You can’t just look at that one dimension.

 

Christian Ward:

Yeah, I wrote about this recently. My post is on something I call the token exchange, because tokens really are a currency. Just like you have currency exchange rates for different countries, and outcomes, think of buying one currency in a country with severe inflation versus another, and what you might achieve in terms of interest rates there, but then you have to process it back into measurable outcomes. I think the biggest barrier right now is that it’s really hard to measure the outcomes, so tokens are hard to pin into a classic currency exchange model. I think we’ll get there. It’s a bit like people who rent out their Tesla, or mine Bitcoin at night, there’s going to be capacity at different times, with users, without users, and I think there’s going to be an entire exchange around tokens. But we first have to solve, exactly to your point, the outcome: what’s the value of that outcome, and was it actually achieved? For now it’s inefficient, because every human is valuing the outcome differently. I think we’ll get there, but there’s going to be a way to ultimately have this kind of token exchange, I think that’s very likely.

 

Saket Saurabh:

I think, in some ways, this isn’t new. Even if you remember the SETI project, distributed computing, any computer plugged into the network, we’ve gone through different evolutions of that, and sometimes what happens is compute grows fast enough that some of it isn’t relevant anymore. I worked at Nvidia for six, seven years, so I’ve been in that world myself, I understand that’s a concern. We’ll see how it evolves. Ultimately, I’m also seeing open-weight models becoming extremely effective and productive, and I’ve seen some of our own use cases where token consumption or token cost has gone down by a factor of ten, for exactly the same measurable outcomes.

 

Christian Ward:

It’s really amazing, and I think that’s one of the most exciting things about this. As I was saying earlier, the adoption curve isn’t the same every month, because the curve itself is changing shape and changing capabilities. Some of these open-weight models, for a while, felt like they were a good year and a half behind, now it feels more like they’re forty days behind. So we’re getting to a place where, I think that’s a really promising element of the open market around this. Obviously, hardware is something you’d know far better than I would, in terms of the gating nature of it, but I’m very optimistic about what open models can do.

 

I have friends who’ve started experimenting even with small, twenty-billion-parameter models on a Mac Mini, and they’re very impressive for simple things, organizing things for your home or family, really simple, and it does a great job. So there’s definitely opportunity there. Again, I think once you get into multimodal, real-time processing of vision, auditory inflection, all these things, we’ll be right back to where the frontier models are out way ahead. But it’s pretty exciting that there’s already a viable open-source community built around this.

 

Saket Saurabh:

Very true. Switching gears a bit on the data side, you mentioned you made the decision to expose Yext’s data as an MCP server to your customers. That’s a decision a lot of companies have been thinking about, and you actually went through it. What was the thought process, what did you figure out, and what followed?

 

Christian Ward:

Yeah, so, number one, we’re dealing with a very particular type of data set, one where the more it’s out there, real time, and connected, the better it is for a client. If an AI went looking for, let’s use Tony’s Pizza again, and found Tony’s Pizza’s data was the exact same everywhere it looked, two hundred instances of it, and every other pizza place after that had only ten instances, Tony’s is going to rank number one every time. So for us, exposing and extending our data isn’t dangerous, it’s part of what we do, it’s the distribution.

 

But really, where this comes up is, you’re an entrepreneur, so you’ll appreciate this: the number of times in my life someone has said, “I’ve got this idea for a great business,” and I ask what it is, and they say, “I can’t tell you.” The problem is, if you do that, you’re never going to realize the value of what you’re actually doing. And that pretty much holds true for almost all data assets. I’m not talking about HIPAA, let’s not go crazy there, but generally speaking, by merging or allowing your data to be leveraged by your clients or partners, you’re instantly going to see things, because they’re coming at it from a very different perspective than you are, and that’s sometimes the most influential.

 

It also helps you refine your own data strategy going forward. So for anyone concerned about this, obviously you have to choose your own path, but for us, it was so obvious, once these tools came out, that unleashing some of these capabilities would inform us, drive the strategy going forward, and, overnight, change the way our clients interact with what we can do for them.

 

Saket Saurabh:

Yeah, and I’d say data is certainly one of those places where one person’s exhaust can be super valuable to somebody else.

 

Christian Ward:

Yes.

 

Saket Saurabh:

And you’ve probably seen that in financial services, you’re an entrepreneur, you built a company in a sector where you’re leveraging data to create financial outcomes, investment decisions, and so on.

 

Christian Ward:

Yeah, and I think that’s what prepared a lot of what I do now. When you talk to a hedge fund, they’ll voraciously consume weather data, voting data, license plates in California with the letter R, it doesn’t matter, they’re correlating anything they can find. I think that mentality is actually quite healthy in the world we’re getting into, because it opens your mind to what’s possible. We just developed something for a client, we took our data, where they had stores, we took weather data, and we looked at sunlight versus their same-store sales and their competitive matrix.

 

If you don’t know this already, I’d encourage everyone to take out your credit card receipts for the last two months, then look at which days were really sunny, and I guarantee you spent more money on the sunny days. That’s just the way we’re wired, it’s been known for a very long time. Now take that and extrapolate it across the world, ten thousand stores and all the competition, to understand which markets, on which days, you could predict good sales just from weather. That’s the fun stuff you start getting into, and it’s not possible if you hide all your data, you’ve got to take this open approach.

 

I’ll also say, something like that, even a few years ago, would have been a data scientist’s whole project. Now you can go out and pull in free data sets, like we just did for a hospitality company, we found a free data set of every FIFA game, every Rose Bowl, football, baseball, every major sporting event, it’s all out there. You merge that into a hospitality company’s calendar of advertising and content, and they can start planning six months in advance to get ahead of everybody else. This stuff is compelling.

 

Saket Saurabh:

Very true. In your world, with that MCP data, your customers are exploring, coming up with cool ideas, and one question is always, how many of these cool ideas actually become production use cases? How do you think about that, especially being the data provider, where you’d probably want many of these to become production use cases?

 

Christian Ward:

Yeah, it’s interesting. When you say “production use case,” you’re usually coming at it from the perspective of software engineering, something they’re going to build, but that’s not necessarily how this works. For example, if I’ve got a nationwide chain of dentist offices, there are some big players in dental and a lot of little mom-and-pop practices, it’s a very particular market. Operationalizing it might just mean choosing to spend more in markets that are so competitive against mom-and-pop shops that they’ll never beat them organically in Google, and spending less in digital where they’re already doing well organically.

 

If we can save them five to ten percent on their national advertising budget, that’s not something they need to bring into production, that’s literally just a data feed that changes how they spend their own ad money. So it’s not always about putting something into production in an engineering sense, sometimes it’s just a behavior change or a shift in methodology. I’d say almost every one of our clients picks up on a few of those pretty quickly, and the experimental, fun stuff is more like R&D, for when they have the extra budget.

 

Generally, we show up to every meeting with pretty specific starting points, based on what we already know for their industry, region, and competition, a good mixture of ingredients, in a data sense, to build something. So, when you think about “production,” in the engineering sense, many of our clients aren’t even using it that way. They’re basically using it to inform everything else they’re doing, and taking that forward.

 

Saket Saurabh:

One thing I was thinking about, in your background, you’ve worked quite a bit on data partnerships, and I’m curious how that world has changed with AI and the growing relevance of data. On one hand, I see companies that have become huge just being data suppliers or data labelers for the AI labs. On the other hand, I also see enterprises becoming more guarded about their data, not knowing who might learn from it. How do you see that space today, from a partnership standpoint?

 

Christian Ward:

Yeah, to your point, in many ways it’s never been easier to build out a really robust data partnership strategy. When it comes to data partnerships, you’re really trying to bucket what’s valuable, why it’s valuable, and at what scale, and then, when identifying someone you might want to partner with, you often have to convince them to let you use it. The thing about data partnerships nowadays, though, if you’re talking about the partnership between you and a paying customer, that’s already a valid partnership, but you usually think of them as using your software, that’s not really it, you’re both a data source to each other.

 

In other words, every business is a data business. I once worked with a company that did retail sales for things like vacuum cleaners and window washing, they’d call a hundred thousand homes in the US a day, and log who they spoke to, whether it went to voicemail, all of it. That exhaust, who has someone home to answer the phone at millions of residential addresses, is worth a fortune to other businesses trying to sell other products. That ended up being more valuable than what they were actually doing in sales. You get to this understanding that a data partnership may not be initially obvious.

 

With MCP, and with data architectures and tools like Snowflake for synchronizing everything, it’s easier to do a data partnership. But, to your point, you also have to be careful, because you don’t want to enable someone to knock you out of a market. I’ll never promote lawyers, but you really should have one, and you should think through this. I’d recommend using AI for this too, but you really want to think through the potential downsides, and one of the most important things is analyzing what derivative data sets you’re going to permit your partner to build. Derivative data sets are a way for them to create something new with yours combined with theirs, or another source, and if they’re doing that, and you’re replaceable, it’s very easy for them to swap you out for another source without you knowing.

 

You want to think all of that through. My brother and I actually wrote a book on this, and the value you can build at a company with a few really smart data partnerships is mind-blowing, and it’s gotten much easier now, but it’s something you’ll always have to be careful with.

 

Saket Saurabh:

Very true. And in your case specifically, you’re getting data from hundreds of partners and synchronizing it, that itself is an architecture and engineering challenge. Any architectural advice or lessons you can share with our audience?

 

Christian Ward:

Yeah, that’s a huge portion of what we do at Yext. Think of us, on one end, as intelligence, organizing everything going on in the market, and on the back end, as the system that actually synchronizes and pushes all the data out. That’s a huge engineering burden, exactly as you’re describing, because every place we push to has its own taxonomy, its own understanding, its own structure, its own schema, so it’s quite complex. At the same time, that feeds back into the intelligence side, because that’s where all the search and AI activity is finding your information, in the places we’re pushing it. So we have a nice closed loop in how we think about it.

 

For a lot of businesses thinking about how to build this out from an engineering perspective, I’d say, Marc Benioff, the day before Claude Design came out, said the UI is now an API, and the API is now the UI. I want everyone to consider that Salesforce was the first SaaS company to ever go public, and one of the largest revenue companies in the world, essentially admitting that no one wants to log into Salesforce. You have to look at your own software and what you do, and ask whether nobody wants to log into it either. What they want is the data in context for what they’re doing.

 

So, opening up all your engineering to that mindset, your goal should be that access should be almost agentic-first, and then worry about your own UI afterward, because there’s a pretty good chance nobody’s going to use it. In many ways, I think it’s a renaissance for data engineering, to get more value almost instantaneously, versus a lot of data engineering being hidden in the back room, in service of what the software’s doing. That ship has sailed. If Benioff is saying, “don’t even log in, just use our API,” that’s what I think is happening.

 

In many ways, I think it’s the best time to think about data in terms of architecture, schema, shareability, and mobility. It’s probably not the best time to want to design everyone’s software. That would be the main difference.

 

Saket Saurabh:

No, very true. I was also curious, when it comes to engineering, and data engineering specifically, there’s an evolving question about what can be AI-generated, using tools like Claude Code, versus what’s still really hard and needs solid engineering. Any thoughts, given the complexity you work with?

 

Christian Ward:

Yeah, look, I’d say our engineering team, like most engineering teams, is very much “trust but verify.” They’re very focused on, “okay, this is interesting, let’s look at this.” The only caveat I’d throw out is, when the chief engineering folks at Anthropic, who have access to Mythos, are saying on stage that a hundred percent of their code is now AI-generated, I think none of us are quite there yet, but we also don’t have access to wherever the next models are going.

 

Put another way, if you documented, not what the benchmarks are saying, but how much your engineering team actually used and trusted AI in their code base, you’d see a curve that looks remarkably geometric over its next several iterations. What that means is your best engineers have a huge opportunity in terms of what they can output. Everybody talks about “10x,” I don’t think we’re there, I think there’s a lot of hype, maybe they are at Anthropic or OpenAI, but they have access to the next generation. What I’m saying is, for the rest of us, it’s not there yet.

 

Saket Saurabh:

But.

 

Christian Ward:

But, at the same time, if you don’t believe it’s coming, because of some political, religious, or other bias, I think that’s a problem. I think you should approach it with the joyous innocence of a child every time you try a model, and see what it can do, because what I’ve found, like all of us, is that engineers often get locked into, “I tried it, it doesn’t work,” and I’m like, no, you tried it two weeks ago, and now it can. You’ve got to give it time. For example, you’re probably using slash-loop, slash-goal, slash-workflow in some of these tools, they’re an incredible multiplier for bug squashing and identifying gaps in the code, they’re really quite good, and even those have improved a lot recently.

 

You’ve really got to accept that, retry things, and I know that’s exhausting, refactoring my code every week is exhausting, but it keeps your engineers in the right mindset, that this is going to get there, and you want to be the one who understands how it got there, and what to leverage, and when.

 

Saket Saurabh:

Yeah, I’d fully agree with that, it’s constantly evolving, you have to be continuously using the products, because that’s a learning curve you can’t just skip. You can’t say, “I’ll come back to it in two years when it’s all settled,” you’d have lost that learning in the process. At the same time, I feel like, in many places, AI code generation is really good for workflows, but when it comes to data flows, especially with the diversity, variety, noise, and errors in data, it’s a different class of problem. Not all software problems are the same class of problem.

 

Christian Ward:

I couldn’t agree more. When we do analysis, a lot of times we’re using AI to inspect the data, the definitions, what’s possible, but then we take that out and engineer the actual math of the calculations, create those as new columns in the table, as derived data sets, and then bring that back to the AI and say, “we built what you were describing, now show me more,” or “make this a dashboard,” or “tell me what I should do next.” I completely agree.

 

Especially once you’re into really large data sets, another mistake people make is crushing the context window with everything at once. You could give it twenty rows across two hundred columns with clear definitions, and it’ll give you a ton of ideas that you should then go engineer and store as actual values through time, because AI has no inherent sense of the data through time, and, frankly, if there’s any value in data, it’s through time. You have to build for storing and engineering that. Again, I go back to, your AI strategy is really just your data strategy, and you need to know how to use both.

 

Saket Saurabh:

Yeah, absolutely. As we come to a close, one last question, on behalf of CDOs who want to learn from you, is: what should they be measuring in terms of success, understanding it’s not just about cost? And second, how should they be communicating that to their execs?

 

Christian Ward:

Yeah, there are a couple of things, and it’s a bit industry by industry. We’re still living in a world where a lot of IT teams are telling their data teams, “you can’t use that AI here.” I’d offer two things. First, you’ll signal to your team members that they should leave and go to a company that does use those foundational models, it’s a retention issue. Beyond that, you’re hamstringing your data teams from being able to show the very value the executives are pushing them to demonstrate. You have to build consensus on how you’re going to approach this.

 

To some extent, the chief data officer role has always been about identifying data assets you can translate into value creation, and being the bridge between IT and the data teams day to day, and the executive team, on what should be stored, analyzed, improved, and tracked through time. So your role as a CDO isn’t just to sit in the data and analyze it, it’s to go around the organization and build consensus on the right way to measure things. For us, one of the biggest things we track is token usage, by department, by model, by the framework of the queries, not necessarily the exact content, but understanding, is this a marketing question, an email question, a calendar question?

 

What’s really interesting, to your earlier point, is that people talk a lot about AI, they’ve got their “AI committee,” and it’s Timmy from the mailroom because he’s young, and a bunch of other people, and I’m like, that’s not how you do this. You’ve got to convince people to adopt it, and the best way is to get them actually using it. Once they’re using it, and you’re tracking it, you can really push forward. But I think the chief data officer has to get right into the middle of it, to help drive the adoption cycle.

 

Saket Saurabh:

Awesome, incredible, I feel like I’ve been absorbing and learning so much, I could go for another hour chatting with you, Christian, but unfortunately we’re at the end of our time. Thank you so much for making the time, and for a fabulous conversation.

 

Christian Ward:

Thank you, Saket. It was great speaking with you, and I look forward to next time.