CEO of Alteryx on Why AI Agents Need Real Business Logic

0:00I think the most overhyped is that you can simply take very large volumes of data and throw it at an LLLM and have it actually make sense for the business. It's sort of like a parlor trick. It looks interesting. And then you dig in and find out, yeah, but none of this is actually how we run the business. Welcome to the Talking AI Podcast, where we talk AI with both experts in the field and early adopters. I'm your host, Matt Page, and we're here to demystify AI for you so you can get some value from it. Let's talk some AI.

0:29Everyone assumes that AI will make data analysts obsolete. But what if AI doesn't replace them? What if it actually makes them indispensable? Today on Talking AI, I'm joined by Andy Van Millen, CEO of Alteryx, a former developer turned product leader turned CEO, and is now steering one of the most ambitious pivots in enterprise analytics. And under his leadership, Alteryx isn't just building tools, it's rethinking what analysts do, how teams govern their data, how AI agents get the business logic right. And we'll dive into why analysts may become the linchpin in this new AI era. What skills will matter, how governance and trust will shape this future, and how AI might actually rewrite the enterprise data strategy.

1:12But Andy, welcome to Talking AI. Thanks for having me. It should be a great conversation. Yeah, this is going to be a good one. I'm very excited about this. But I want to start with your thesis around analysts, right? Your view of the analyst role is, I'd say, contrarian to the mainstream of AI will automate everything narrative. Lay out that thesis of how this analyst role is changing and evolving in your life. Yeah, I mean, I think there's a set of expertise that comes with how a business runs, how it adapts, how it's implemented, that often sit within these operational teams that analysts are often part of.

1:52And the way to think about this is when you want to know how sales commissions work, there's someone on the RevOps team you go talk to. Or when you want to know what the audit and accounting policies are, there's somebody on the FP &A team you go talk to. When you want to understand how the return policy operates, there's someone on the merchandising team that knows that information. And I don't think that not only that knowledge, but the specificity of like, how do we want a business to run completely disappears? I think a big part of what we're trying to figure out is how does it scale better?

2:23How do you make that knowledge, that ability to work with the data be something that the AI interacts with? And that's where I think analysts have this unique new role that they can be playing as an enabler to really make that happen. And I think we see that today in how AI operates. AI operates without a lot of that specificity. And what we really want to know is like, what does it mean in my company with my data when I do this thing? And I think right now the people that answer that question are analysts. And so how do we help them, you know, sort of sit a little bit behind the scenes with AI to be the enablers of that technology to give the right answers?

2:56You're reminding me of one of my early roles in my career. So I was at Cricket Wireless on the team of data scientists, their advanced analytics team, right? And I was working with these like super smart PhD data scientists. And my niece essentially was storytelling with the data, like pulling the insights out of the data. I remember there's two key things that just stood out to me. A, what they were doing was super advanced. And they're always felt like this kind of glass ceiling between what I could do and what a PhD data scientist could do. and it just took so long to actually do the things they were doing get to the data like you're talking about you got to talk with this team and that team and that team like nightmares are starting to pop up in my head from my past life there and there was one stat i think this was earlier this year all tricks ran it said in the typical work week analysts spend 10 to 11 hours just collecting and preparing the data, like just getting it ready to even have at a starting point.

4:02I'm assuming AI is going to be playing a major role in that aspect of it. Quick break in the pod. Our State of AI 2026 report just dropped and it breaks down what actually is changing in AI, what's hype and what leaders need to be paying attention to this year. You can grab it right now on our show notes or at hatchworks.com. Yeah, I think analysts are on both sides of the AI equation. So on one hand, I think giving an analyst access to AI capabilities helps that analyst move more quickly, helps them prepare the data more quickly, helps them gather the right information, maybe even find the right information.

4:36But I also think it takes the work that they do and enables it to scale out exponentially as well. So maybe let's take a simple example. If I'm a salesperson and I'm going to close this big deal, what salespeople do is they get out Excel and they start trying to figure out what will my commission be if I close this deal. And what they end up having to do is reach out to the commission's person. You know, they send them a Slack or a Teams message. You know, what will my commission be if I close this deal? And that person has their own spreadsheet that they're operating on because sales commissions change every quarter.

5:08There's sales programs and SPFs and things that are out there. So someone sort of does that math. What I want to be able to do is ask ChatGPT or Gemini, what will my commission be if I close this deal? The reality is today, if you do that, you're going to get this really broad answer of like, well, here's how software companies tend to generate their commissions. And ChatGPT is going to give you this interesting, long answer about how commissions are calculated. But the thing it won't do is tell you what your commission will be if you close this deal, which is really the only thing you want to know the answer to.

5:39And so how do you go to that commissions analyst and say, hey, what if you built a little calculator that you maintained that ChatGPT could talk to that would tell you the answer to the commissions question? And next quarter, when that gets updated or changed for the business, that that same, you know, commissions analyst could update the calculator so it gives the right answer again. I think that separation of the technology from the analyst is important. And the analogy I use is a company's website. So a lot of people like to compare the AI boom to the dot-com boom, right, when there was an internet version of everything.

6:18And if I was in the web content management business back then, and there was this big debate, like, should IT be building the website or should marketing? And the answer became sort of both. Most large companies, the IT team, run some of the infrastructure for their website. But what we're not doing is have marketing send all the content over to the IT team that's converting it into web pages. The marketing team owns the message and the images and those things. And I think that's kind of where the analyst fits into AI. AI needs to answer all these questions about the business. I don't think IT is going to own the audit questions, the supply chain questions, the comp commission questions, the marketing funnel performance questions.

7:02Like that's not what IT professionals do for a living. They're going to own the infrastructure that says, oh, we have this big data platform and we have this large language model and we have an infrastructure that makes it all work. But I think they're going to have to go back to those same analysts and say, but could you make sure it gives the right answer to commissions? Or could you have it actually encapsulate our marketing funnel or our merchandising rules? And again, that's where I think there's this sort of and answer, which is like, yeah, IT is going to own a bunch of that. and AI is going to be amazing.

7:32And the analyst is going to have a role to sort of make sure that it answers the question the right way and that it represents the business and things they know about. Yeah, I think that's where the analyst has that superpower. They have the business context, which is so critical in essence. But the commission example you give, I feel like that's such a, it almost feels like a distraction. A, because it's like, oh my God, I got Johnny in sales asking for what this commission is going to be again And when I should be working on like these more strategic, impactful questions in the business, I guess, essentially, you're empowering others to do this in essence, right?

8:09Well, but you could, it is a simple example, but you could also have, for example, that same commissions agent working with your annual planning agent to figure out based on what comp plans you put in place, what your operational expenses will be. So again, you're going to have to have a commissions expert involved in that. But you might be part of a, maybe you're running scenario planning with AI, we're going to run through 100 or 1000 scenarios. Sort of like what we do with McLaren, they run 1000s of simulations of every race, you know, using data that they pull together. Maybe you're running 1000s of scenarios for financial planning, which is something we don't do today in most companies.

8:48But you still need the analysts involved to say, well, you know, this is our business model. Here's how commissions work. This is our business model. Here's how professional services work. This is our business model. These are our operating margins. Like, those are real things that exist. What you don't want to do is run a thousand scenarios where you go, well, you know, for a generic software company, you know, we ran a thousand scenarios. Like, your board is not going to be impressed with that. They're going to want to know, well, what about our company? Like, what about our margins? And every business is a little different.

9:15We've all joined a business and you come in and you think you know about how that industry works and you learn like, oh, we do this a little different. And this is interesting in our product. And so I think there's a reality that we're trying to bring to AI about how any given business operates, right? About the realities on the ground and what your policies and your rules and the things you've told your customers and how you've built your products and how you deliver your service. that again, I think your operators, your analysts in your company, the ones who know those rules, they know how that works.

9:48How do you bring that knowledge to AI? Not just your data, but that knowledge. And I think that's the big gap. Like I can throw all my enterprise data at Gemini and it'll do some neat things, but I kind of need the analyst involved to go like, well, here's why this is that way. Or here's the version of revenue we use here. Or here's why we do partial deals in this region through these partners, right, for Gemini to build that scenario planning? Yeah, and that McLaren partnership is the neat one. I want to hear about that one later. But you hit on a point, though, because enterprises have messy labeling, acronyms, tribal knowledge.

10:28How do you actually help enable AI to interpret these things? Obviously, generative AI is good with unstructured data and these more nuanced questions. But I'm sure there's still at times where AI is like, what the hell is that acronym mean? For sure. Well, and sometimes even like, and why and when to use it? I think there's this big push right now in our industry, again, maybe to be a little contrarian around what's called the semantic layer, right? Like, how do you describe all the data in your cloud data warehouse? The challenge with that is it's often not the description that's the problem.

11:02It's when to use what and why and how does the business operate? And the framing I'll provide on why this is so important that the analyst is involved in this is, you know, I joke all the time when I go ask my analyst a question, they don't ever give you an answer right away. They ask you several questions back, right? When I say, hey, what will happen in this scenario? They go, oh, well, do you want to include this part? And what about that part? And there's a couple of deals and there's a partner in Asia that does this stuff. And do you want all that included? And so there's a knowledge about the business that isn't just a semantic description of the data.

11:34And so, again, that's where I go back to how do I empower those folks to say, oh, okay, this is a repeatable thing that I see. How do I build either a curated data set or a calculator or a small agent that answers that? But I've got to apply my knowledge to it. And this is important. I have to be able to go back in and change it because nobody runs a business to keep it the same. Right. I mean, I've never worked at a company where you show up and they go, everything's running great. Just don't touch anything. And, you know, the revenue will magically show up. Like you're constantly making changes and adjusting.

12:05So some of that will be things we learn from AI, but some of it's going to be, hey, we're going to expand into a new region. Well, the historical data doesn't tell us how to go into that region. We've never done it before. We need some people to come in and maybe AI is helping us make those decisions. How do we encode that logic into our systems and our agents? I think that's analysts. I don't think that's your IT team. That's going to be, hey, we hired somebody that understands merchandising in Korea and we're going to expand into Korea. And so the person we hired that understands merchandising in Korea is going to help us train the agent on how to go do that.

12:37And again, it kind of goes back to my web content management example earlier of this idea of like the technology pairing with the business expertise. And the business expertise, I don't think, just disappears off the face of the earth because of AI. I think we're trying to feel like, how do you scale that differently with AI? And you mentioned the part about the analyst asking questions first. That's a great contact for anybody listening. Have AI ask you clarifying questions. It just makes the response so much better. But back in the day, I remember that same experience where I'd get asked for a report, I'd deliver the report, and that would trigger five other questions.

13:12And then you have this cycle of, okay, well, that creates another ticket to get that put in the backlog to do the thing, which triggers five more questions. And it's this long, painful cycle. So the logical thing is empowering people to do things with their own data. And you're talking about building these agents and things like that. But what's the balance between empowering people versus giving enough structure to where, you know, the democratizing effect is not counterproductive in a sense? And I see that at the shadow IT and all that kind of stuff happening. I think it's actually a really well-framed question because it is a balance, but it's also an intent.

13:55So I'll give you a pragmatic example. I care a lot about Pipeline as a software company. You do a lot behind your pipeline and what visibility you have and things like that. when I joined Alteryx, I wanted to have more visibility into my pipeline by each of my sales teams in each of my regions by stage so I could sort of see, how are we doing? These folks have enough pipeline to work on. The old way to do that would have been for me to go create a dashboard and then have to ask my ops teams a bunch of questions about the dashboard. Why is this red? Why is that red? What happened here? And instead, what I can do now is I used Alteryx.

14:36You mentioned I'm technical in my background. I used all tricks to create a workflow to go pull my pipeline data from Salesforce out of our Snowflake database. And I could create a very small data set that did all the summarization of our pipeline. And I could talk to my RevOps team and my FP &A team to make sure I was actually calculating it correctly. It was the right numbers, right? And there was some nuance, some very specific things that would tell me like, oh, this is categorized weird here, but I could build that. And then with this very clear data set that shows every quarter, every one of my sales teams, all their pipeline by stage.

15:08No extra data, no extraneous stuff, but well-defined how we measure it. And I can take that well-defined data and I can give it to a well-instructed GPT. And then I can ask it all kinds of questions. But I'm in this zone, right? I know that I'm operating on pipeline data. I understand kind of what it looks like. Again, I don't really understand the data. I don't have to know the table names or the column names or anything, but I can sort of go, how much late stage pipeline does my team in Asia have this quarter? and it shows me charts and graphs and it's all correct and it's got a scope knowledge.

15:39I think that's very different than when there's this aspiration of like, well, I'll just take a large language model, I'll point it at my Salesforce data and ask it questions about pipeline. Our opportunity table in Salesforce has 118 columns in it. Like it doesn't know how to handle that because we've been around a long time and we've run all kinds of different things through Salesforce. And so when I create a narrow data set, it's spot on, it's accurate, it doesn't hallucinate, it has well scoped and well good intention. And that GPT can be called by other GPTs. So it's not to say I always have to have a limited view.

16:14But I think the thing we've really missed is this, again, back to our conversation, this critical step of somebody who actually understands the data, building a data set that's purpose built, and building a set of instructions behind it that makes sense, versus saying, yeah, I've turned a large language model loose on my cloud data warehouse, and now I can ask it questions about my data. When we do that, what we find is either I get garbage because I don't ask it a well-scoped question or, and this is the hard part, I have to understand my data well enough to ask it a well-scoped question. And the reality is most people in your business aren't gonna understand the underlying 118 columns in your Salesforce database to try to ask it the right questions, to tell it the right way to pull the data together.

16:53This is the analyst job. This is where the analyst comes in and goes, oh, here's how we measure a pipeline at this company. out of this giant data set, we grab this information, we structure it this way, here's the calculations we do, here's a GPT that works with it, and now, to your point, I can ask all kinds of follow-up questions. I can dig in, I can look at pacing, I can look at why does this team have this, has this been true historically? It's incredible. I spent all kinds of time in this little GPT, all saving time with my analyst. I'm not having to ping them with all these questions, but I think that's the piece we're going to get right in the next 18 months is we're going to start to realize there's this role for sort of applied knowledge about the data.

17:33Someone who understands the business and the data who can build again curated data sets, little calculators, things like that, that you hand to a large language model. And then you go to everybody else and go, hey, this agent, this GPT can answer your questions about merchandising returns or marketing content or whatever. And when I do that, I find I get really good answers because it has the context and the underlying data is correct. Yeah, I love taking a meta approach. Like I'm still working with AI prompting. Like if you want to get a better prompt, talk to AI about what you're trying to achieve and have it help you create the prompt.

18:11And the same, it kind of triggered in my mind when you mentioned the messy data set of 118 columns. Do you see AI in that role assisting as well when you're trying to aggregate data? And again, 118 columns, like the best analyst probably doesn't know all the nuance. Do you see AI as part of that kind of discovery of, okay, best way to aggregate the data? What is interesting in here based on the outcome I'm trying to achieve as well? I mean, I think it helps. It certainly helps. So I've gone from that pipeline to now I'm looking at marketing funnel performance. Well, the first thing I did was I went to ChatGPT and I said, hey, I want to build a marketing performance GPT.

18:54What data should I gather to do that? Now, it gave me a generic answer. You're going to want this data from Marketo, this data from Salesforce. What it doesn't know is my Marketo data and my Salesforce data and why it is the way it is, right? And again, that's not a labeling problem. Those 118 columns I have in my Salesforce database, they're all labeled well. But what I'm trying to figure out is, well, which labels apply to the business process problem of, say, our pipeline? And again, that pipeline, what pipeline is it? Is it sales pipeline? Is it my consulting services pipeline? Is it implementation pipeline?

19:27Like, what am I measuring? So there is a role for somebody who knows that about my data to be involved. AI can help me go, hey, help me build this GPT. What things would I need to go do? What's the best practice for measuring pacing? How do I make sure I bring in those other things? So it's interesting. I'm sort of using AI to help me shape up and frame what I'm going to do with that agent. I'm using AI to help me think through what data would make that agent successful. but I then still need to have a practical understanding of my specific business, not just the data, the business, right? What do my reps get paid on?

20:03How do we measure pipeline historically? You know, things like that. Do I include my renewals? Do I not? Is it pro-serve? You know, what about low margin things where we're selling? You know, like, what is it? And so I think if I can combine all that, what I really want is an AI-enabled analyst, right, who knows my business but is using AI to be more productive, that has a set of capabilities that they can build and importantly maintain, right? We change the way we measure pipeline in the future. I need to go change my data curation around what that looks like that can tell the GPT, here's the language we use at our company, right?

20:38Like what do we call late stage pipeline? There's no definition in my table of late stage pipeline. For us, that's a certain set of stages. But if you get on the forecast call, everybody talks about late stage pipeline. So how do I tell my GPT what those things are, right? and give it the tools to interact with my data in my business the way that makes sense to me. And AI is part of that, but it's not the whole thing. And the problem, this is that whole, you know, we've had this knowledge management conversation for 30 years about, like, how do you take the stuff that's in people's heads and put it in your systems?

21:08And the reality is, like, I don't know that you ever get there. You try to enable the people that have the knowledge to put it into systems. Again, I go back to, I think that's the analyst. Like, that's the person that knows that. Totally. One of my favorite things from this year was Darmesh CTO at HubSpot. He said at the beginning of the year, 2024, leading into 2025, so this is the year. 2025 is the year of AI agents. He just came back a month or so ago and did the same speech at their annual conference, and he said, I was wrong. This is not the year of AI agents. This is the decade of AI agents because it is so emergent.

21:43There is so much potential for this, and there is a very long runway of what is capable. We're still in its infancy in a sense. I've heard you mentioned several times, agents, kind of agentic workflows, GPTs. What does that look like in the world of Alteryx today? And where do you see that evolving? Because it is still very new and emergent. Sure. We talked quite simply about sort of three stages I think we're going to go through. The first stage is just automation. The reality is part of the challenges with AI is there's a lot of things in a company that today are not automated. It's really hard to take something that's manual and scale it into an AI universe, right?

22:23So I think we're going to go through a period of just what can I automate? I'll give you an example. We can sort of use the examples. We'll go through the three stages. So imagine your budgeting versus actuals process, right? So the end of every month, you're an executive. You get your budget versus actuals. This was my budget. You know, how did I do? In a lot of companies, that's still quite manual because companies run a bunch of different systems, right? Through acquisitions, different geographic systems. And so often you have some sort of partner on the finance team that literally is using an Excel spreadsheet to grab your budget and your actuals.

22:55There are two tabs and they're running your budget versus actuals, right? So if that's the way I'm operating, I'm not going to build an agent anytime soon that gives me real-time budgeting information because I can't even do BDAs in an automated way. So a lot of what we help companies do today at Alteryx is automate that budget versus actuals process. We'll connect to whatever system the budget is in, whatever system the actuals are in, and do that process. Well, now that I've automated it, so that's step one, I can start to bring AI into that automation. So one of the things I might do is use AI in that actual reconciliation process.

23:32So what things don't reconcile? Are there free text forms? Can I use AI to just make that process better. I think more interestingly is imagine at the end of that process, instead of giving you, let's say you were the executive I'm supporting and I'm your FP &A person, instead of giving you an Excel spreadsheet that is your budget versus actuals, maybe I give you a report that AI has run that tells you, hey, here's your last 12 months of BVAs. Here's where things have changed. By the way, I've looked at things like travel and expense externally. I can tell you that airfares are up 10 % globally.