Dashboards to decisions: How AI is closing the gap between insight & action | The Airwallex Podcast

My thing I said all the time then, and I still say all the time now, is dashboards raise more questions than answers. It's almost this paradox. It's like if you make it cheaper to go get a data insight in time and energy, you're just going to get a lot more because the demand is infinite. There's infinite demand to know things. We are curious creatures. You talk to a lot of data teams and it's like, "Well, we got this request for this dashboard. We'll get back to you in 3 weeks." And for a lot of things, 3 weeks is as good as never. I mean, the decision's been made, dude. Like, we've moved on, you know.

Hi, welcome to the Airwallex podcast. I'm Fiordenza Bojolo, Head of Strategic Data Science and AI at Airwallex. Today I'm speaking with Barry McCardel, the co-founder and CEO of Hex, a leading AI analytics platform that is reshaping how teams interact with data and make decisions. We're going to talk about how you can incorporate AI analytics into your business to make informed decisions faster with real-time data, what are the best practices to deploy, and what are the pitfalls to avoid while doing it. Barry, welcome to the Airwalks podcast.

Thanks for having me. It's such a beautiful setting and a beautiful day.

Beautiful San Francisco day as well.

That's right.

Great to have you here. So from Palantir to Hex, you've worked on very different problems, very different scales of problems. What prompted you to solve how teams interact with data?

Yeah, so I always say it's like very selfish. I've been a builder and user of data tools my whole career, so like I was a kid who coded a lot in high school and then in undergrad I had no idea what I wanted to do. Wound up doing social network analysis as a kind of a focus in my last couple years there. I took this class, it was really cool, worked in a lab and kind of without realizing it, I was doing data science. So I went to consulting, did that for a couple years and I was the guy on all my consulting projects that just was so nerding out on Excel and I was building these crazy, basically data apps in Excel and illicit access databases on a purloined desktop PC under my client's desk. Don't tell them. And so just that, I was really passionate about this. And then I had a friend who was at Palantir, brought me there. That was a great time, learned a ton. And then fast forward 5 years, 5 great years, left to go to a healthcare startup. It was an amazing company, really cool mission. How do you use data to help improve clinical trials? But what was interesting to me was our data stack, we have this modern data stack and "We have a traditional BI tool on top, so dashboards." And it was like, "Fine, dashboards, they're fine." But it was like all the interesting work was happening elsewhere. It was like random SQL queries and Jupyter Notebooks and scripts and spreadsheets and a screenshot of a chart pasted into a PDF of a deck that was sent in an email was the way to share. And I was like, "This sucks." But that was state of the art. And so I started as a fire. I was like, surely someone has solved this. We tried a bunch of different things and tried to cobble together different open source stuff and there was nothing that was a great product for this. I was kind of bopping around New York where I lived at the time asking people on data teams, what do you guys do for this? And one person said, well, maybe you should build it. You seem to have a lot of ideas for this.

Why don't you start it?

And I was like, nah. And then I woke up one morning and I turned to my wife and I was like, I got to do this. So.

That's so exciting.

Two of my closest collaborators from Palantir, we got together. I think we all had a lot of passion for this space. And so there's a joke at Hex that the project is good enough now that one day I'm going to quit because I've like built the thing I needed and like, thank you.

It's always evolving though. It's always evolving.

It's always evolving, which is good. But yeah, that was like the genesis.

I totally agree with you. The problem with dashboards is that they're static. It's always about you're seeing a number and you need to do the analysis to understand why that number is following that trend. What else is going on?

It's like people are like, dashboards are dead or whatever. I'm like, they're fine. My thing I said all the time then, and I still say all the time now, is dashboards raise more questions than answers.

Exactly.

It's like, here's what happened. I look, I'm the CEO of my company. I, I'm an executive stakeholder. I, I get a lot of dashboards, but I'm, well, of course I use my own product for this, but like, I'm always like, well, I wanna understand and go deeper on it and ask more things. And so we built Hex around that initially. Like that was the initial thought was like, In fact, we weren't really focused on dashboards and what you'd think of as BI. We were like, how do we go help all these people who want to do deeper dive work? Data scientists, data analysts. There's also a lot of product managers, engineers, people who were working a lot with data. And so the first leg of the company, the first couple years, we really focused on that and we built something great that a lot of people liked. And then now of course we'll talk about now with AI and all of that, but that was the genesis. And what's interesting We'll get into more of what we're doing about AI, I think. But what's interesting is it's kind of the same mission, like this massive technical shift that's happened.

Exactly. That enables the mission, actually.

How do you understand what's going on in your world with data? We now have like computers can talk to us now, but that same idea and the same unsolved problem feels very evergreen.

Yeah. So I'm noticing a shift in companies' willingness to adopt AI analytics in the past 6 months. Obviously smarter LLM models is part of the puzzle, but are you seeing that shift as well? If you— so why do you think that's happening right now?

Well, I mean, obviously the models have gotten good. I also think they're just— people are seeing more applications in more places where AI tools are able to make massive improvements for them. I think data is this glaring thing. It's like, I should be able to ask data questions, yeah, you know, using AI and get accurate answers, but that's really hard to do. For, for us, it's not just the last 6 months. It started really last year when we launched our agent product. It seems people saw it and they were like, that. And now the demand is really high.

Tell us the state of the product today. What does Hex do today utilizing AI?

So Hex makes it easy for anyone to ask and answer questions with data. If you're a more advanced user, technical user, we have a notebook-style interface where you can see code and sort of really get into the guts of a more complex project. All the way to if you're sort of a more casual data user, self-serve data user. We have an interface we call Threads. It's basically just chat your data. It's really nice. Or you can use it in Slack, you can use it from ChatGPT or Claude via MCP. You can use it from anywhere. And it kind of just lets anyone ask and answer questions of data. But then because they're all able to do it in one place, because you can share and collaborate on projects, because you have shared context and the collaborative context layer is just so critical. It makes it so everyone has trusted and accurate answers and there's really good observability and context configuration around all of that.

So—

Yeah, that's so important.

We don't think it's like one point thing. It's like, look, if you're looking for great notebook interface with AI built in, we have the best version of that. It's awesome. If you're looking for a great data chat experience, we have the best version of that. It's awesome. But those aren't the reasons that I think the overall thing is compelling. It's because they all work together with the context layer. And that's something we're I think the integration is kind of the thing for us right now and why teams are adopting the product the way they are.

Yeah, let's talk about that. Let's talk about for a company that's interested in adopting AI analytics, what are the best practices that you're observing from the businesses that adopted your product and also from your own experience?

Yeah, so I think there's a really interesting pattern we're seeing, which is people really want to use AI for data. But it's different. Folks have an intuition for this now because of what we've seen with AI coding where it's like you have a population of expert users, engineers who can pick up these tools, integrate them in their workflow, start to build trust and intuition on how to use them. And then there's all this infrastructure, the governance, collaboration, and deployment rails around coding that has been built up over a long time. And there's kind of this already infrastructure and guardrails around software engineering that you can just swap in. Is it a carbon-based or silicon-based life form doing the keystrokes? Yeah, it's the same place. Data, I think a lot of people see that thing happening in coding and they're like, "Great, we're going to do this with data." And they kind of jump all the way to, "We're just going to give everyone in the organization ChatGPT overnight and it'll just work." And you can try that, But what we find the most effective and best practices is starting with your expert users the same way you would in software engineering, by the way. Start with your expert users, have them use an agent to do their work, build up intuition. And then a concept we're really focused on right now is what we call context exhaust. Basically, as you're using a product, you're emitting a lot of context without thinking about it. If you sit down and you use our agent to do an analysis, you might say, I want to look at all the new customers who have signed up in our priority region or priority countries, e.g., you know, Singapore, Japan, I'm making this up, US.

You're giving the context to the agent in your notes.

Yeah, use this table, join with this other table on this key, all these things you know.

Yeah.

Or somewhere that, you know, our agent can say, great, you know, we'll go do that for you. But also while I'm working on that, do you want to go commit this into your organization's context?

Yeah.

Or do you want to suggest a change to your context? That kind of context exhaust can compound over time and grow and you start to build confidence in a certain domain. You say, great, this domain is really well covered with context. Now we can go unleash this, um, on more people in a way that's gonna be trusted. So to me, there's like, that's one part of the best practice. The other thing, and people are surprised sometimes I say this, is like no one's AI analytics product, uh, including ours is perfect. The models are stochastic. Surprise. They're also kind of bad at this still. They're clearly being very brutally reinforcement learned on software engineering, which is a verifiable domain. Our domain is less verifiable. A famous example, you can kind of pick up these examples, but in the most famous eval set that everyone uses for AI SQL, it's There's a question that's like, "Which country has the most fraud?" It's actually a financial services thing, so maybe it's kind of applicable to Airbox, but which country has the most fraud? Well, you can answer that by saying the country with the highest absolute amount of fraud or the country with the highest fraud rate.

Yeah.

Both are valid answers. What's right depends a lot on the domain, what you mean. Maybe it depends on the last turn you had in the conversation when you were asking about rate of fraud.

Your use case.

The use case. Maybe how does your organization define that? Maybe the answer is the agent should come back with saying, "Hey, it's kind of both. Which one do you want?" Yeah. They're not good at that. They're surprisingly bad at that. And surprisingly, another thing we see a lot is the models will leap to conclusions sometimes where there's incomplete context. And so I think the other thing that's really important is having the observability and improvement of that. It's something we're spending a lot of time on at Hex of how do we have an agent behind the agent? So every thread that happens in Hex, we have another agent review it. And it scores it and we get, we, we say, here's what this thread's about, here's the topic, and here are warnings like, hey, the user pushed back, or the user was doubting this answer, or there appeared to be incomplete context. And when we're building the system with that in mind and like really focused on that of like, how do you have this like compounding adaptive context layer that's learning, um, and helping you as someone who's running this and your data team sitting at the center of this. Understand what are people asking about.

Is the API to push the context and all of the queries and conversations back to your clients so that they can improve their internal context layers for other use cases? Is that coming along?

Yes. So there's some of that already exists. Tomorrow, well, it'll be in the past if you're watching this now, but we're launching this suggestions feature. And as part of that, that's all exposed in our CLI. So you could, as an example, get in the CLI or VR, API a feed of all the suggestions we have, where we have these things our agent has suggested. So you have the thread level, we go through and analyze and we say, where are these warnings? Where are these issues, potential issues? And then we cluster those and say—

That's amazing.

There's a suggestion of, hey, we've got 9 people last week that asked a question that there's not clear context on. We can suggest actions. We can say, here's what to do. Some of those actions we can do in Hacks, but you we also right in the UI have a little thing you can copy the CLI command to go paste into your terminal.

I think that's gonna be super key to have that continuous improvement loop. Uh, that's right. Uh, because unless we put that in place, uh, we're not gonna get to that like level of confidence that we want.

That's right. We, I think every great technology product is built around breaking a trade-off. Like if you're gonna give everyone in your org like ChatGPT or Claude with a SQL thing, you can do that and people do that and then they wonder why people are getting a lot of different answers and there's no observability into any of this, no ability to collaborate on it. Someone shares something and it has no sense of data permissions and security. Like, how do you unleash this while making all of those things true? And I, I think that's like the platform we're building.

That's amazing. And it, it kind of comes to what I wanted to discuss about the pitfalls people fall into when they're trying to build and establish an AI analytics product.

If you view data generally as just writing code, it's like, well, models can write code now, so everyone can do this. And you can do that. And this is the thing is that I think people don't realize where the problems are happening, because everyone's getting their data answers, right?

But are they correct?

But you don't know because it's like, who would, how would you know?

Yeah.

Like, how would you know? Like, you know, like someone's getting the wrong— we're already seeing examples of this, like in the wild where companies are reporting numbers and then having to issue corrections because they're like, this is a hallucination or whatever. So that's really hard. I do think that's a pitfall and like a mistake. And the space is evolving and changing. People are going to try to get ways to do this. The models are getting better, but you know, something tells me like we'll get to GPT-7 and it still won't be able to solve for if you don't have great context and you don't have this ability to collaborate on that context and work as a team.

That continuous loop of improvement of your context layer.

That feels really important. The other thing I would say, by the way, is I do think interface layer matters. Of course, a lot of people are starting workflows in these agent interfaces, which is great, but I do think there's also something to be said for a purpose-built interface for a purpose-built task. I think a lot of people think just software in the interface layer is going to die and everyone's just going spend all their time in terminals or that that's something they're moving toward. Something tells me that's not going to be the case, but we'll see.

So I'm a company thinking about enabling AI analytics at my company and I'm thinking, hey, I have access to LLM models. Should I just build it myself with my engineering and data science teams or should I go buy it?

I have a weird answer that my team hates, which is Go try to build it yourself. Go do it. I think it's great because the customers who have done that and then realize where the jagged edges are and why it's hard come back and they appreciate us all the more. And I don't say that in a weird way. It's like truly, I really mean it. We actually have a blog post on our site about how to vibe code Hex.

That's amazing. Do it.

It's great. It's actually really fun. The first part of that process is really fun. Because with LLMs now, everyone's a genius, right? You can tell Claude, build me this, and all of a sudden you're a software engineer and you built your own thing. It's awesome, man. It's really great. The first part of that experience is really fun. And if you want to solve the use case for you personally sitting there doing something, you can get somewhere. It can build charts. It's good. Okay, what if I want to share it? What if I want to start working together? How's the context managed? How's that done at scale? Is it secure? How are data permissions enforced? How are the integrations working? You start to get down this, and not only is it a lot of work and stuff to figure out, and maybe something, maybe you shouldn't be vibe coding, 'cause you start thinking of stuff like auth and permissions and security and not just auth for the artifact, but okay, it's connected to these two other data sources and both of the data permissions on those have to be enforced. I'm like, ooh. But also, is that what you wanna spend your time maintaining? That's an open question. And I think it's awesome. I really mean this.

It is fun to vibe code in those.

These models make it so easy to tinker. I do this all the time. In fact, a great example of this, I don't think we've talked about this publicly and now I'm going to and we'll probably get barraged by vendors, but we vibe coded our whole own system internally for doing all of our own agent observability and evals. We have a product we built internally called Shoebox, And we vibe code it and it's really fun. It's cool. And we're proud of it and we're gonna get rid of it cuz we should not have tried to build it. We're like so deep down maintaining it and like DevEx workflows around it suck. And meanwhile there's companies that just do this.

Yeah.

And we are easily spending more engineering time maintaining Shoebox than we would just like getting something off the shelf. And by the way, one thing I think is cool about these off-the-shelf products is they're becoming more malleable over time. Like they're kind of assuming you wanna work with them. They've got CLIs and APIs and maybe if you wanna build your own interface to do a special task that should work. And we're thinking about the same exact thing at Hex, which is like I mentioned earlier, like how you can make the system feel like your own to the extent you want to be able to make it malleable and configurable. You know, we want to really help you max out on that, but you shouldn't have to go try to build all of that up from scratch. I think that's, um, if you want to go for it, but I don't recommend it.

It comes to also exactly what you said. Do you want to spend your time very valuable time on building your own product or building the analytics product that will enable your teams.

It's not just analytics. It's, it's people are doing this for all sorts of stuff. I mean, people are talking about their vibe coding CRMs. Sure. I mean, that's a simpler product in many ways. It's just a database, but you think it's just a database and then you get, you get so deep into it and you're like, is this what we should be doing? Like what problem did this solve for us? But I get it because I think we're in this era of tinkering.

Yeah, there's going to be trial and error and figuring it out.

The new default is full codegen.

Yeah.

And it's infinitely malleable. We get feedback all the time. Our request list is probably thousands of tickets long of people wanting long-tail viz features or app layout features or customization or can I do custom fonts? Can I do custom colors? No way, I want to do a custom accent color for this thing. A tab bar, but no, I want it to be on the left and I want it to have the text like this or whatever. We would, we could spend a decade trying to build every one of those features in the old paradigm. Or we can say, hey, why don't we just let you code gen it? And you can build anything you want. But you're going to do that in an environment that's connected to your data, connected to your context with security and permissions and all these other great things we've built around this and make that amazing to do. And so I also think there's this synthesis of like, how do we help people embrace that malleability and tinkerability? And I want to go build something completely custom for me in a system that takes care of all of the stuff you probably shouldn't be vibe coding.

Yeah, yeah. Let's come to how you're seeing impact of real-time data analysis on businesses from your customers and clients.

I've always felt like speed really matters in data. You talk to a lot of data teams and it's like, "Well, we got this request for this dashboard. We'll get back to you in 3 weeks." And for a lot of things, 3 weeks is as good as never. I mean, the decision's been made, dude. We've moved on. We're already onto the next thing. And so the ability to just ask questions, or even if you are sending something to the data team, just the speed of them being able to get something back to you and iterate is so much higher. It's awesome. And so I hear stories about this all the time. My favorite recent example of this that kind of blew my mind was the executive chef at a quick service restaurant. We went on site to go see what their workflow is and their executive chef is a Hex power user. I was like, "What? We have a chef power user?" Yeah.

What is he using?

He's looking at timing for when they're putting things in the oven, like steaks in the oven and ingredient freshness and trying to look at their whole supply chain and really understand all of their data around how are their because they get a lot of data, they're a pretty modern company, prepping for board meetings or our last board meeting, our board members, they're good and diligent, curious board members, they'd be like, "Well, what about this?" Or, "Well, how is this affecting this?" Or these types of questions that the normal answer is like, "Okay, well—" Take a note, get back to you in a week. A couple weeks and usually do we ever really get back to you? It's just, okay, well I just pulled up the product and it turned into a demo. The back half of our board meeting just became a demo of me being like, "Well, let me just ask." That's the best way to impress your investors, I guess.

Bring the product to life.

And this is all leading up to a point I want to make that I think is really important, which is this is not zero-sum. The executive chef doing that work, it's not like that took a data job away because that's the type of thing that probably never would've gotten to the data team anyway because they're so bandwidth constrained. There's a line around the block to get answers from them and the follow-up times are weeks. And if someone can just ask their question, it's actually this thing. If you've ever heard of Jevons' paradox, do you know what this is? Okay, so in the mid-1800s, economist, I believe his name was William Jevons, observed the paradox, which was as steam engines were getting more efficient, net consumption of coal was going up. And that was very counterintuitive because they're getting more efficient, shouldn't we be consuming less coal? Less, yeah. But we're consuming more. Why? And economics is called a rebound effect. And the reason is kind of obvious to us now in the modern world, because it's just, we come up with a lot of things to do with watts. Like if you can be more efficient, if it takes less coal to generate a watt, you're not gonna get less watts. You're probably giving it a lot more watts. The cost of the thing went down and it, you got more of it because of that. Yeah. And you know, here we are with these high wattage lights. It's obvious to us now, but it, this is true in data too, where it's, it's, it's almost this paradox. It's like, If you make it cheaper to go get a data insight in time and energy, you're just going to get a lot more because the demand is infinite. In the same way there's effectively infinite demand for power, there's infinite demand to know things. We are curious creatures. And if you give a chef or a board member or anyone in between the ability to ask and answer questions of data, they're going to do it. And you're just going to get a lot more answers. And that's really cool.

Let's talk about two things. One is the role of data science teams. In this new world of AI analytics, and you're alluding to that, right? And then the second thing is decision-making in this new world where everyone is empowered with the data. Let's cover the first one first. What do you think is the new role of data science teams?

I think you're going to have data teams for a long time, and I think they're going to become something they've always been, but it's going to focus. And I think that the role, the teams and the roles are going to narrow in some ways and broaden in others. I think we're going to see data teams as these librarians, curators, arbiters of truth saying, where's all the information we have access to as an organization? How are we accessing it, bringing it together, keeping it secure? You're in a regulated financial services business. That's incredibly important to have great security and governance around all of that while making it accessible. How do we make sense of all of it, have the right context around it and make bring in platforms that make it available to everyone. I actually wonder in some ways if data teams will be better understood as almost like internal AI platform teams.

This is exactly what we're doing at Airwallex. We're shaping the data science teams to be AI agent builders. And what does that mean? You're building the AI agent, but you're also responsible for building the data foundations, context layers, the guardrails, the checks in the system to ensure that there is the model and the answers are accurate and improving, et cetera. So, we're now holding our data science teams to be accountable to improve the performance of these AI agents. So now the productivity of that data scientist is increasing too.

Totally. I think that's awesome.

So how do we become agent builders or enabling others to build agents in a reliable way? And then how do you also become agent managers so that you're able to kind of like do multiple things at the same time?

I think that's well said. I do think there are certain types of roles we'd assumed you have for data roles in the past that will sort of broaden away to maybe you don't think of them as data roles. I think of a lot of data analysts who really almost already were these almost BizOps strategy folks. They were really involved in decisions like that. You needed someone to be a specific data analyst because they're the one who knows the SQL and the thing. I think now you actually, I see a lot of these people actually almost getting kind of promoted in literally or figuratively, where they're now just taking on more and doing more things.

Yeah, they're becoming agentic go-to-market teams. Totally.

And it is interesting to think about how some of these roles evolve. I mean, it's happening everywhere too. Like, what's a software engineer anymore? What's a product manager anymore? These things are changing quickly and it's really cool. I think data teams are well positioned though, because these are people that already had to spend time understanding systems, understanding the businesses they lived in, thinking about context. That was already part of the job. One of our PMs on our team, Her name is Claire. She invented the term analytics engineering when she was at DBT. And we've been talking about this a lot internally because it's like, well, now we have this new role called AI context engineers. And she's like, no, it's just analytics engineers. It's the same thing. It's always been context engineering. Analytics engineering was always context engineering. It's the same thought. And so I think some of this is all new and crazy and different and really alien, But a lot of it's also very familiar. Some of it is the same with a different name on it. Right. And so we'll see how all of it shakes out. It is funny to watch people rediscover things. There's a lot of people who are building with AI systems for the first time and they're discovering tests. They think they invented tests. We call them evals now or whatever, but it's like—

They have a new name, evals.

We have a few decades of track rate experience on this one. Yeah.

It's a really good one. So now VR, everyone is empowered to make decisions with real-time data. What kind of organizations will thrive in this environment?

I do think there's a concept of decision metabolism. I heard this from someone, I forget who it is. If you're the person I heard this from that I'm stealing it from, I'm sorry. I did not come up with this originally, I'll admit. I was going to be honest. It stuck in my head a little bit about you can have a lot of data and you can have a lot of insights, but are you the type of organization that can action these things? We've all probably seen or familiar with or been in organizations that you can have perfect insight and information every day and they're still going to just be slow moving. They're going to take a long time to make decisions, be very fixed mindset, fear-based. I think it's just obvious right now that the organizations that are going to thrive the best are the ones that are able to adapt quickly and learn and change. There's sort of a bigger cultural thing that gets in the way. Yeah. That's going to be true for any number of AI things. There's companies that take a really long time to make product decisions or really bureaucratic with that. And It's great that AI can help write the code faster, but is your velocity actually going to improve? Which is an interesting question, by the way. I think a lot of people, including ourselves, see this AI acceleration happening in how we're building our products. But you kind of look around in the world and it's like, where is all the software that we're supposed to get from this? I use a lot of products every day that the pace of them shipping features seems exactly the same as it was a couple years ago. Disappointingly, I'll say. And it's like, where's— you guys are using AI tools, right? Why doesn't pace change? It makes you wonder where the bottleneck is. Because I think it's easiest to assume that the bottleneck was in how quickly engineers could type.

You get the data, you get the code out.

It's like, I don't know. I don't know the answer to that question. It's an interesting riddle.

I think it's going to be how do we empower teams to make their own decisions in a decentralized way, but at the same time ensure a system where we know how each team is making decisions on, like what's the decision that has been arrived at the end. So—

That's tough. I mean, it sounds nice, but it's a cultural thing. It's tough to— it's a tough thing to synthesize.

Culturally and systematically empower teams to make their decisions, but build a system where the decisions that have been made are recorded and flow up to the top so that we're aware of what's going on.

If you can figure out how to do that, let me know. I do think there's a tension. This is getting outside of data a little bit, but there is a tension between wanting decentralization, pushing decisions down to the front, high ownership, high autonomy, and oh, desire for control, consistency, and predictability. We were talking earlier before we started filming about things like job titles and how companies grow and how you avoid becoming bureaucratic, right? A lot of that comes from this desire for control and predictability, and you get bigger and you get risk-averse and you add process. Process is nothing other than just trying to make things happen consistently every time. Yeah, exactly. And of course that's great if you want it to happen the same way every time. And companies just expand the circle of things that they think need to happen the same way every time more and more. Something bad happens, you ship a bug, you have an incident, you have a missed quarter or whatever it is, You start to implement process around it. You don't want that bad thing to happen. Yeah. The hardest thing in the world, I think, is to see something bad happen and not implement additional process around it because you're like, yep, that is a defect that is like, it's the necessary trade-off. I think a lot of things are about trade-offs, culture, companies, decisions, businesses, products. They're all about trade-offs. And I think really opinionated cultures embrace trade-offs tightly and they say, we're going to optimize for this thing. And there's this trade-off that comes with that.

And we're going to accept that trade-off.

We're going to acknowledge it and we're going to embrace it and love it. But there are some companies from, for which like they've chosen, and maybe it's just even their investor base or whatever, a public company where they're like, predictability and consistency is the thing. That's what we're going to optimize for. Well, the trade-off of that is you're not going to move quickly. You're not going to make decisions fast. That's okay. You can just say that that's you and don't pretend otherwise. Don't pretend that you're going to be able to do both things. We were talking about Palantir. One of the things I've learned from Palantir is remarkably decentralized culture. People at the front made decisions. You had a lot of things that didn't go anywhere, but because you were sort of covering the search space so much and you had so many shots on goal, you took smart people and put them in front of problems. The things that did work, really worked. There was a level of innovation and learning that would be impossible to direct from a centralized model.

It's very interesting that you came down to people. Maybe in that decentralized world, decision-making world, what it comes down to is hiring the right people. That's all that matters. Yeah.

This is Netflix. If you've ever read No Rules Rules, it's a kind of similar mindset. It's like, keep the talent bar super high, Bring in the best people and then get rid of all the rules. Because if you bring those great people in and you set them up with the right context, they're going to be able to go off and do a lot of things.

You recently posted an article about weirdos want it. So what type of people do you hire in this new era of AI?

Yeah, so I've, right, I posted this piece. It was actually an email I sent internally that I wrote one night called Weirdos Wanted. And it was something I'd been reflecting on a lot as we're hiring and growing our team, which is what are we looking for? It's easy to say we're looking for intelligence and intellectual horsepower and someone who's going to work hard and is nice to get along with and all those things. But I've been thinking a lot more about what's the role of humans in an era where the AIs can do a lot of things. When you're building an AI product, the thing you think about a lot is what's in distribution for the model. Basically, what's in the training set is kind of what you can predict to get out. And I think about the humans are here to do the out-of-distribution things, to have the weird ideas, to be creative and break that mold. And I think everyone we're hiring kind of needs something like that. I wound up asking this question a lot, what's weird about this person? And people didn't get what I meant. And I was like, what's different? What's interesting or strange? What's unique? It can be a weird side hobby. It can be a track record of looking at things differently. Something we picked up in the interview process, but I always want to find something that's out of band or unusual. And I think what makes our company special, if I may say so myself, is not, well, it's not me. It's not, it's this amazing band of people we have. Every interesting product decision we've had, every great brand expression, Every really special customer experience by definition has come from something kind of out of band and a little weird.

And those out of band thinkers, as you put it, I guess, would be able to actually ask the AI agent to think differently too.

It's an interesting question. Can you ask an AI to think differently? I mean, by definition they can only do what's in distribution and this becomes a kind of philosophical debate. But if something, you wind up hiring people every time where the thesis is like they're just going to operate a thing where it's just going to happen the same way every time. They're just going to operate a process and do something. And if it's not obvious today, it'll be obvious tomorrow that a lot of that can be automated away. That doesn't mean those humans have to be automated away, but it does make you think about what's the role for the human. And it's to set those processes up, but to think differently. I want to be a company that always has a very high critical mass of weirdos, which I think is, people may think is a strange thing to say we want because Sometimes you say weird is a bad thing, but I want to champion weirdness. I'm pro-weirdness. I think it's what makes things special. It's what makes things notable. It's what makes things different and what makes things human.

Okay. As we wrap this up, what's something that has been in your mind that you think a lot about that the industry is not talking enough about?

Um, yeah, we, we build a tool at Hex that makes it really easy to sort of, um, have a machine do a lot of things for you. And I wonder and worry sometimes about where the boundaries of that should be. It goes back to what we were just talking about with weirdo weirdness and the human nature of it. Because I spend a lot of time writing and I find it's a great way to think. There's this great line, "If you're thinking without writing, you only think you're thinking." I actually love writing as well. Yeah. But do you ever have this experience that I have that sometimes you're in a rush or you're feeling tired and you have the AI write something for you and it's pretty good and you're like, oh, and then you're like, wait, I feel like I just skipped the gym. This is a pretty well-worn topic. I'm not the first person to ever think about this, but I guess I think about it a lot and how we're building the product. And I don't know the answer to it. I'm very far from knowing the answer to it, actually. I'd love to hear ideas, but how do you build a product that both helps people offload a lot of things but also almost engages or encourages them to have a level of critical thinking on the things where they need to. I don't know what the answer to this is, but I worry about building an edifice of automation in our domain and in others that sort of almost turns us into the WALL-E people or something. Do you know that movie, the Pixar movie? It starts the long slide toward us just sort of being these endlessly entertained people who are massively context switching, dispatching little things to agents and not really thinking about it. There's a beauty to thinking and sitting with a task for a while and can we build systems that can help with that? And I hope everyone would develop the willpower to sit and do the deep thinking, but if the last 10 years of tech adoption has taught us anything, it's our brains get addled. I wonder how we can build systems that actually make that easier or more pleasant or more engaging or more focused. I have no idea what the right answer to this is, but that's a thing that I'm really curious about. I have weird thoughts like this. But again, weirdos wanted. If you also have weird thoughts about these things, please apply to work at Hex.

Well, Barry, I really enjoyed our conversation. It was really fun to kind of like get into the Hex story, learn the latest versions of the product, think about AI analytics with you, the best practices, the pitfalls to avoid, and also think about like how it's going to impact organizations and the type of people that we should hire. So it has been a really fun conversation. Thank you. Hopefully everyone enjoys it and looking forward to doing it again.

That's a wrap on this episode. We hope this conversation gave you practical insights, food for thought, and a clearer view of what's shaping the future of global finance. If you enjoyed this conversation, make sure to subscribe. Wherever you get your podcasts, and stay tuned for more deep dives into the technology, strategy, and people shaping this new era of global finance. We'll catch you next time.