How AI Is Reshaping Modern Data Teams and the Future of Platforms with Xavier Gumara Rigol

0:00Benjamin Wagner: I think that the most exciting thing that we are working at the moment is this natural language to SQL to Insight type of tool.

0:07Xavier Gumara Rigol: The Data Engineering Show is brought to you by Firebolt. Firebolt is a high-performance, open-source, analytical database. It runs as a single binary on your laptop and scales to hundreds of nodes and petabytes of data in your cloud. Whether you're dealing with real-time analytics or batch processing, Firebolt lets you unlock the full potential of your data without compromising on performance or cost. Hello, everyone, and welcome back to the Data Engineering Show. Today, I'm super happy to have Xavier on. Xavier is currently the head of data at a company called ManyChat, who work on basically kind of marketing automation, content creator automation, and so on.

0:45Xavier Gumara Rigol: So really great to have you on the show. Do you want to quickly introduce yourself and also what you guys are building at ManyChat?

0:52Benjamin Wagner: Yeah, I have to do so. Thanks for having me in the podcast. My name is Xavier. I've been in the data ecosystem throughout the whole of my career, started as a BI developer, then transitioned to managed teams in the data platform data space, doing so for the past 10 years. Now at ManyChat, I'm leading the machine learning and the data platform team. So we help the business with the data that they need and making sure that this is on time and accurate. And we also help with providing some personalized experience to our customers, the content creators. and in the past, maybe like a nice things that I did in the past also is before joining ManyChat, I wrote a book on the divide between data and product functions.

1:37Benjamin Wagner: That was a fun experience in my career that I'm very happy to have done that because many companies have done functions of data and product in many companies that don't work very well together. So that's why I wrote the book.

1:48Xavier Gumara Rigol: Very nice. Now I'm actually super curious because this seems like something that's changing a lot with AI. And like, I think a lot about like engineering and product and how these things are changing. How do you see it for data teams, right? Like my general perspective on software engineers is that they need to become more versatile and that these roles are like in many ways blending, right? Because building the thing is no longer the hard part. It's like, it's usually like building the right thing, prioritizing, focusing on the things, kind of then taking it to customers and getting them excited.

2:20Xavier Gumara Rigol: So curious how you see that evolving with data roles now that AI and agents are getting exceptionally good.

2:28Benjamin Wagner: I think that the type of work is changing. It's not going anywhere, but it's changing. For example, in the past, or in many companies still currently, like the work of a data analyst is to provide the results of a SQL query to somebody in the business, right? So it's a Tableau dashboard or a visualization, whatever. That's changing with AI. We have been able to implement an analytics agent that is able to answer natural language questions with high accuracy quite fast, I would say, this year, right? So we have much more data in the business being self-served and being able to cover the data themselves, etc.

3:07Benjamin Wagner: If you think about the work of data scientists, for example, the past maybe was to build machine learning models. Now these machine learning models, probably they have some kind of generative AI flavor into it, like prompt engineering that extracts some information from your data, and then you want to run this pipeline every hour, every day, whatever. And I think that when you go into that space, yes, maybe you do things now with generative AI that it took months to build before, but this new work with generative AI, you need to also properly do it because what happens sometimes is that, yes, You build a model, but then it's not properly evaluated.

3:46Benjamin Wagner: It's not deterministic. And when you think about these things and put them into production, you need to actually, or the business would like some predictability and deterministic results, right? So it's a different type of work, but the work is not going away. Now, what we are seeing with data scientists is that they are working with evaluations, they are working with LLMs as judges, and more things are becoming LLMs or pipelines that incorporate foundation models, LLMs. I think that the type of work is switching, and we are just doing more things faster with less amount of time. And for data engineers, for example, the data platform they have to support, it's a little bit more complex because of these questions that we need to answer with AI, etc.

4:34Xavier Gumara Rigol: If you think about just the organizational structure of data teams, especially on large companies, what you historically had in many cases is that data is a separate organization. And I think then, okay, you had teams which were saying, okay, we want to now do self-serve analytics for data. And actually, we turned into a data platform team. And then we start giving self-service analytics to every function of the business, all of these things. And then in many cases, you have data product people and data engineers or data team that was focused on just helping the C-level execute. If you zoom forward to, let's say, 2027, it's hard to say what's going to happen then.

5:13Xavier Gumara Rigol: But how do you think, especially in large companies, also like the structure and organization of data teams has to change to adapt to how teams build today, how fast teams could move in theory and so on?

5:28Benjamin Wagner: I think, yeah, I'll talk in the book about this. And I think that the data roles that are value creators, let's say like data analysts and data scientists, they need to work in a cross-functional team with a product manager and with a specific. problem to solve. User activation, retention, onboarding, checkout. They need to be part of a cross-functional team responsible for one business problem or domain. And then they need to get help of part of the cross-functional team. There needs to be software engineering, UX, product management, and data. And if in the past this used to be like persons working in these roles, I think that now Now we can think of this as just roles and maybe different or people have different roles right here, or you don't need the person for each specific role, let's say.

6:17Benjamin Wagner: So in the past, you had backend engineer, front-end engineers, a PM, but now the PM maybe does the analyst type of work with things to your AI. And you don't need a data analyst specifically for that specific team because the product manager with the tools around and AI is able to self-serve what he or she needs. So I think that we are going to see more of that. People that is able to do more than their original role, right, or their classic roles. All of this needs to be supported by a strong platform team that enables the rest of the organization to be self-served on these type of things so that actually like these cross-functional teams can work autonomously.

6:57Benjamin Wagner: So one of the things that I've seen also is that in general, well, it depends on the cases, right? But one of the things I'm seeing at ManyChat, for example, is that it's not that we are doing less because of AI or cutting roles, etc. In general, it's not the case. It's just that we are repurposing things in order to do more things at the same time. And this is the benefit of AI and how we are taking advantage today.

7:20Xavier Gumara Rigol: Maybe, and this is a natural point to chat more about what you guys are doing at ManyChat, right? Take us through the work of the data teams there. Are you mostly internal to the organization? Do you do customer-facing analytics for the content creators? What systems do you use, etc.? So we'd love to learn about your stack.

7:39Benjamin Wagner: So we do actually both. Everything started for internal customers, right? So I joined a year ago, more or less. The platform team was there from before. It was decided to use the Snowflake plus Tableau. We're still using that. And we have a data platform team that is responsible for maintaining the infrastructure and maintaining the capabilities on top of the analytical database. And this includes, for example, data quality checking, privacy compliance for the rest of the data team or analysts and machine learning scientists to be more autonomous. And then what I mentioned before, like we actually have the data analysts embedded in the product teams and the business teams so that they are close to the business and they are responsible for a specific problem.

8:23Benjamin Wagner: And they partner with DMs, which they are very close to them in order to run A-B testing, perform deep analysis, do idea validation with them, etc. And then more recently, maybe in the last year, we have started also to explore customer facing analytics. So that's something that we are iterating on this. Started with providing data from Snowflake into operational database, let's say. And it has been working pretty well so far. And this is something that we are trying to get more and more. Because when you are a content creator, you want to have statistics on how well your posts are performing, etc.

9:04Benjamin Wagner: And this is something that we have been exploring and we will continue.

9:08Xavier Gumara Rigol: Gotcha. And there you basically create summary metrics, aggregated data, etc. And you push that actually back from Snowflake into your operational database in order to then serve high concurrency summaries to your customers. Yes. Cool. So if you look forward throughout the rest of 2026, right? Like what are you excited about? What are some projects you guys are driving that you think are really exciting?

9:32Benjamin Wagner: I think that the most exciting thing that we are working at the moment is this natural language to SQL to Insight type of tool. because this is a new capability that we didn't think it was possible just six months ago. And I think that late last year, beginning of this year, I think the foundational models made like huge improvement when it comes to writing good SQL, for example, and everyone in the organization started using more cloud code, for example. It came naturally, right? So like people wanted to connect to Snowflake MCP in order to ask questions and be self-serve with the data because they were using that other tools like Notion, Slack, whatever.

10:07Benjamin Wagner: So this is kind of the new way of working for them. But the thing is that in order to provide high accuracy on the answers, you need to have very well structured and built a context layer on top of your data warehouse. And we had, but we also have improved this through the last months. And I think that we need to continue because at the moment we've made available this capability for subset of the data that we have in our warehouse. But some other areas, we still have low accuracy and we are building the right context, modifying the right tables so that the LLMs can get the query right, let's say.

10:50Benjamin Wagner: And then we foresee a bit of a problem here in the sense that now people is using this type of tool in cloud code as an output to bigger outcome, let's say. So they just want to ask this or get the answer to a specific question in order to continue the research. So I think that now the problem that we have is that there is a lot of artifacts that are starting to float around the organization and around data, etc. And we do govern a little bit more on that front. So we will be exploring some tools around these or processes and see how we can do that a bit better.

11:27Xavier Gumara Rigol: Nice. In this world, how do you approach build versus buy decisions, right? So like, for example, for the natural text to insights thing, like we just added a support for TextQL, for example, to Firebolt, which is one of these like cool newer startups doing AI analysts that you can interact with in human language. How do you think about that in 2026 now as a data team or leading a data team?

11:52Benjamin Wagner: Yeah, we are evaluating buying versus building. I think that and we are in some POCs for that. And the thing is that it has, but we couldn't wait for the POC to start to provide something. And also like with Cloud Code or other tools or generative AI in general, I think it has, the barrier to build has been lowered quite a lot. And we could provide something very quick, very fast to the rest of the organization. And one person in a month was able to build already something that could be tested at least with the data analyst. And then we expanded to the product managers, et cetera. So we have that, we are evaluating and we will make, it's something actually that we have been discussing this week is like, okay, after this round of POCs with this natural language to SQL tools, we'll need to make an assessment internally on what tool we use for what and what we buy versus what we build and make this like a conscious decision that we have sign off in the organization and then we move forward.

12:53Benjamin Wagner: So I would say like, for now, we are exploring both directions and we will make a decision in a few months because everything changes so fast that it's really hard to commit to something in a short time.

13:06Xavier Gumara Rigol: If you think about, I'm actually curious, how do you run a POC for a tool like that, right? I'm like a database guy. When you run a database POC, right, what do you do? Okay, you say, here's my cost. Here's the current performance I want to achieve. And then you can do like a performance bake-off or a cost bake-off. and then have a set of functionality. For these text-to-SQL things, it seems quite different because a lot of it is also about the quality of responses and these types of things. Take us through your POC criteria for evaluating a text-to-SQL type of tool.

13:38Benjamin Wagner: Actually, I think that the text-to-SQL is the solved part in a sense, right? So it's true that we value the abilities who self-serve. It's a given in this type of tools and they are very intuitive to use. In our case, I think that a bit specific, but some other people will be in the same situation, I guess. But we don't want to have different BI tools, let's say, around. In the sense that if we are leveraging and doing a BOC for a text to insight tool, which it's not only like the answer to the question, no, but these tools, they offer the possibility to create a chart, to create a dashboard and then to create, to share it with other people.

14:20Benjamin Wagner: And then we are kind of having a shadow BI tool and we want to prevent that. So one of the strategies or decision criteria is that we want to fully replace our current BI tool with such a tool. And if not, we will keep with a cloud code playing. So that's one thing. And there is cost also on the other hand, right? So it needs to be competitive when it comes to cost. And I think that what we want also is that everything should be manageable in a programmatic way, right? So everything that you create in the system, whether it's a chart, a dashboard, a thread in the conversation, we should also, or a metadata that you ingest into the tool, you should be able to create this in text because we want to work with this with LLMs and generative AI.

15:11Benjamin Wagner: That's our process. Awesome.

15:14Xavier Gumara Rigol: Appreciate you kind of like giving us the rundown of your data stack, the tools you're looking into right now. Any closing thoughts as we wrap up the episode on your thoughts on like modern data platforms, roadmap, AI, etc.?

15:29Benjamin Wagner: I think that in my perspective, like things are changing very fast and we need to adapt in this new world of working with AI. I have also experienced that it's very hard to promise autonomous agents building this type of things if you have kind of current, I was going to say legacy, but if you already have a data platform or like a data system, it's probably not 100 % ready to be improved and maintained by AI agents. you still need humans for that. For example, when you start something from scratch, we have done with the Cloud Code plugin, we see that it's very feasible to implement this with AI.

16:13Benjamin Wagner: So what I mean is that for the current platforms to maintain and for the current capabilities to maintain, you will still need humans for quite a few years, I would say. But then if you start from scratch, which you probably can get to a nice place where you need less people to maintain in comparison to if you were to build this like three years ago. We have not experienced this, but we will experience this kind of like new companies going on with less headcount, while current companies will still need a bit more headcount than them in the next year. So, and I think that this applies to data also.

16:50Benjamin Wagner: So we still need a data analyst in the company and we will need them for a long time.

16:57Xavier Gumara Rigol: Makes sense. Perfect. Well, so good to have you on the show. Thank you for dropping by. Excited to continue following your work. Thanks.

17:03Benjamin Wagner: Thank you for inviting me.

17:05Xavier Gumara Rigol: The Data Engineering Show is brought to you by Firebolt. Firebolt is a high-performance, open-source, analytical database. It runs as a single binary on your laptop and scales to hundreds of nodes and petabytes of data in your cloud. Whether you're dealing with real-time analytics or batch processing, Firebolt lets you unlock the full potential of your data without compromising on performance or cost.