Groupon's Matías Coumet buys commodity, regulated, global-scale infrastructure like message delivery, builds the context layer that encodes how the company makes decisions, and waits on agent autonomy until that layer exists.
Build the brain, outthink the competition
Groupon used to need two weeks of engineering to change a campaign. Now it takes under two hours.
Build vs. buy is martech’s oldest debate. AI made it harder, because every company now has access to the same models and the same tools
That shift is one piece of Project Foundry, Groupon's company-wide push to rebuild itself as an AI-first business. Matías Coumet leads revenue operations and BI at Groupon, including AI operations, and decides what to buy and what to build. His team bought one platform for global messaging and built nearly everything else in-house: the context layer, the agents, the workflows.
He's also skeptical of most of what the market calls AI. His hottest take: "Most of the AI features are actually prompts with a price tag." He argues for focusing on the data and context, because models change and you can swap them.
The line between buy and build
Matías buys anything that's a commodity, regulated, or needs global scale from day one. Message delivery is the clearest example. Deliverability, carrier relationships, and compliance would take years to rebuild, and none of it differentiates Groupon.
He builds everything the AI needs to make company-specific decisions: business logic, definitions, context, workflows, strategy, and tactics.
The test between the two is one question.
"If I buy something and my competitor buys the same thing tomorrow, do they get the same edge? If the answer is yes, then that's not something that I can create value in the company. It's just having the short-term edge because I bought it first."
There's a third option too: wait. Matías leaves fast-evolving model capabilities alone, because building them now means rebuilding them later. Early in-house LLM projects across the industry were outpaced by OpenAI, Anthropic, and others. He also leaves alone anything customer-facing without a human review step.
Why Groupon bought the one thing
Groupon's legacy stack was 10 years old and ran across more than 80 internal services. Buying a messaging platform was both a technical and a political decision, and Matías is candid that the change management was brutal.
The deciding argument was about who should be able to change a message.
"Marketers shouldn't be opening a ticket to change a message or modify an asset."
Buying the plumbing freed engineers from maintenance so they could build what actually makes a difference.
A mandate buys speed, not belief
The CEO named Foundry in a pitch to the board. Matías reads the name as a metaphor: melt down the old processes and recast them. The goal was a full rebuild, not a tweak.
Because Foundry came from the top, it had three things a grassroots effort rarely gets: speed, top priority, and permission to shut down outdated internal processes. That last one matters. Matías's tell for a company faking AI-first is to ask what it actually switched off. If it's only adding, that's a pilot, not a strategy.
The mandate had a cost.
"What it cost us is people complying before they believed in the process."
The first reaction to AI, he says, is defense. The mindset shift took eleven months. Now, when he asks his teammates whether they could work without AI, they say no.
What broke wasn't the technology
Groupon ran data unification and its AI buildout in parallel, on purpose. Matías doesn't believe AI should be separated from the data, and the data wasn't structured yet.
"What didn't break was the technology. It was fine in that matter. What broke was the definitions and the ownership."
Each department had its own agenda and its own KPIs. Once everything flowed into one place, gray areas appeared wherever two departments shared a metric.
The data fix was consolidation into one warehouse, with a CDP as the single source of truth for customer data. The bigger fix was the context layer, built on BigQuery. It holds every team's strategy, tactics, definitions, and rules for communicating with customers. It's model-agnostic by design, so Groupon can swap Claude, ChatGPT, or Gemini without rebuilding anything. Locked definitions change only by cross-departmental agreement.
AI doesn't resolve disagreements over definitions. It exposes them. The only fix Matías has found is to put every department head in one room for a full-day workshop, agree on the definitions, and codify them in the context layer. Reporting is merged the same way. Department heads review shared numbers together, which makes it impossible to spin a story at another team's expense. Collaboration clicked when teams stopped optimizing for department KPIs and started treating company OKRs as the shared goal.
The rule that never bends
From day one, Matías set one rule. Nothing touches a customer or moves money without a human approving it. Analysis agents can run unsupervised. Execution agents cannot.
The reason has less to do with trusting the models than with fit.
"It could be technically right, but it could be wrong for the business."
Something that works for a competitor could seriously damage Groupon's revenue. Accountability works the same way. Every agent and project has a documented RACI with a named DRI. Traditional on-call has been replaced by an L1 agent that triages issues and notifies the person accountable for that workflow, the one who can pull the plug
Every agent earns its keep
Every agent at Groupon has to prove its ROI. Anything that isn't ROI positive gets fixed or killed.
The value side counts time saved, speed, quality, and backlog reduction. The cost side counts tokens, infrastructure, and human review time. To keep token costs down, the team builds with a frontier model, then switches to a cheaper one for trivial day-to-day work.
Matías is honest about where the math breaks. Sometimes human quality beats AI speed, and sometimes token consumption costs more than the value an agent creates. Those processes stay human on purpose.
The most expensive mistakes aren't dramatic failures. They come from confidence. An agent pulls from contradictory sources, presents an invented or out-of-context answer with certainty, and a human acts on it without catching the error.
Groupon doesn't run a central team chasing every new model release, either. Each stakeholder runs their own research agent for their role, and teams share what worked and what turned out to be slop.
Are vendors becoming dumb pipes?
For many, yes, Matías says, depending on how open they are. The winners have the best pipes and stay open to the customer's own brain through APIs, real-time events, and support for MCP and customers' own models.
Vendors who wrap the same models their customers already use offer no inherent AI edge. Matías puts the share of "AI-powered" vendor features that are wrappers at more than 50%. What matters is the data, the workflows, and the distribution around the model. Or, in his metaphor, a saw is only as good as the lumberjack using it.
Some things he'll always buy. Asked which tool he couldn't replace in a quarter, he answered immediately: messaging platforms.
What should scare a vendor CEO is intelligence moving to the customer. ChatGPT-style agents can already make purchases, scrape interfaces, and compare competitors on their own. When customers can access everything themselves, the vendor UI becomes optional.
So what should vendors build in the next year? Agent-operable architecture with APIs, event streams, and support for bring-your-own-model. Pricing on outcomes and usage rather than seats,” or finish the last one: “And they should be world-class at the hard parts, like message delivery, that companies won’t build themselves
His prediction for what disappears within three years: standalone AI copywriting tools and dashboard-only analytics layers. They'll survive as features inside broader platforms, not as standalone products.
Hiring for when the agent is wrong
If a company builds everything in-house, how do its people show their experience on a résumé without the brand-name tools? Matías hires for adaptability over brand-name experience. Structured thinkers from large companies can struggle with the flexibility Foundry demands.
His interview question: What happens when the agent is wrong and you didn't catch it? He's looking for accountability and root-cause thinking. A candidate who says, "I fixed the agent," and stops there hasn't diagnosed anything upstream.
He also rewards failed experiments. Someone who safely tests a new model and reports it doesn't work for Groupon saves the company time.
Build, buy, or wait in 2027
For a company a year behind Groupon, Matías's advice is three steps.
Buy the plumbing. Infrastructure and architecture the business can't run without should come from vendors who already specialize in them.
Build the context layer first. Put the definitions, the ownership, and who does what in one place, so agents rely on a shared brain instead of working in isolation.
Wait on agent autonomy until that layer exists.
"Don't go into the autonomy of those agents before building the context layer. Why? Because it will just create faster mistakes."
Three takeaways
If a competitor could buy the same tool tomorrow and get the same edge, it isn't your differentiation; your advantage lives in your definitions, context, and workflows, not your vendor list.
What broke at Groupon was ownership of shared definitions, meaning a context layer is first a governance project and second a data project.
Without that context layer, agents just make faster mistakes, which means autonomy is the last thing you turn on, not the first.
Matías Coumet leads Revenue Operations, including AI operations, and BI at Groupon. He started as a product engineer at Ford and worked in manufacturing and freight tech before moving into revenue operations. Find him on LinkedIn.
Subscribe to Making Sense of Martech wherever you get your podcasts. Follow us on TikTok, LinkedIn, and don't forget to like and subscribe on YouTube.
Full Episode Transcript
00:00:01.600 — 00:01:11.070 · Jacqueline Freedman
Welcome to Making Sense of Martech, where the rabbit hole goes deeper than the headlines. This is a show that pressure-tests ideas, not just platforms. On the Hot Seat today we have Matías Coumet. He leads Revenue Operations at Groupon, including AI operations, where the company is betting its future on an AI-native redesign called Project Foundry.
His team bought one platform for global messaging and built nearly everything else in-house—the context layer, the agents, the workflows. Each agent has to prove it makes the team faster or it gets switched off. So the question running through the conversation is really simple: When you build almost everything, what exactly do you buy? And are your vendors actually becoming dumb pipes?
A little bit about our guest first. Matías and his team moved Groupon's global email, push, SMS, and onsite messaging into a single platform, cutting campaign changes from two weeks of engineering to under two hours. Now he's building the data and agentic AI layer underneath it—shared infrastructure that CRM, marketing, product, and BI all build on. Welcome. Thank you for being here.
00:01:11.590 — 00:01:12.470 · Matías Coumet
Thank you for having me.
00:01:13.430 — 00:01:23.270 · Jacqueline Freedman
Well, I'm excited to dive in with some rapid fire—some quick answers, no dashboards to explain things. First off, what was your first martech tool?
00:01:23.310 — 00:01:37.790 · Matías Coumet
I would say it was Salesforce, and I'm happy that the first marketing technology that we're actually using was customer data. Nevertheless, I'd say that spreadsheets still are the most used martech tools around the world.
00:01:37.790 — 00:01:44.950 · Jacqueline Freedman
Without a doubt, for sure. Build, buy, or wait. Which one do you secretly enjoy the most?
00:01:44.990 — 00:02:08.750 · Matías Coumet
I prefer building. I consider myself a tech-savvy person and also a self-taught professional, so I prefer building things. I like to get hands-on. The only thing that I have to mention is that waiting sometimes saves the most money. So that's one of the issues when considering just building things directly, for sure.
00:02:08.789 — 00:02:15.990 · Jacqueline Freedman
And so on that note, what is one thing you would never build in-house, no matter how good your team or the AI gets?
00:02:16.030 — 00:02:39.030 · Matías Coumet
I would say message delivery. We handle emails, push notifications, SMS—so many delivery systems to customers—because of the deliverability, the carrier relationships, and the compliance. I would say that message delivery is the only thing that will never be in-house, especially for scaling a business. It's not that big a differentiator.
00:02:39.070 — 00:02:52.270 · Jacqueline Freedman
Agreed. It's very complex. It sounds easy, and it is easy until it's not. In terms of agents, what is one that you're the proudest of creating and building, in five words or less?
00:02:52.390 — 00:03:08.980 · Matías Coumet
I created several agents. I have one that helps me triage the capacity and the workflows of my own teams. I have a lot of people right now, so it's easier to handle that way. I would say, in five words, one that answers questions before they are being asked.
00:03:09.020 — 00:03:21.540 · Jacqueline Freedman
I love that. If only we'd had that forever. So every exec these days is saying AI-first, AI everything. In your opinion, what is the tell that they're actually faking it?
00:03:21.580 — 00:03:40.620 · Matías Coumet
In many cases, I would ask one question: What did you actually switch off? So in terms of the processes they have, what did they implement in terms of automations and workflows that allow them to switch something off? In many cases, they're just adding. So that's a pilot program, not really a strategy.
00:03:40.660 — 00:03:50.020 · Jacqueline Freedman
Agreed. And I think a lot of those pilot programs are coming to an end for numerous reasons. Okay, last but not least, what is your hottest take?
00:03:50.060 — 00:04:07.420 · Matías Coumet
That could be contradictory, but most AI features are actually prompts with a price tag. So in my consideration, you need to focus on the data and the context, not on the model, because the model evolves over time and you can actually change them as well.
00:04:07.460 — 00:04:36.460 · Jacqueline Freedman
Wholeheartedly agree. I do not enjoy folks overspending for very simple things. Groupon is in the middle of an AI-native transformation, and the whole company is watching to see how fast it pays off. And you're the one drawing the line between what gets bought and what gets built. So could you break it down for me? When does something get purchased and when does it get built? And what is the guardrail—the line that you choose?
00:04:36.620 — 00:05:19.290 · Matías Coumet
Okay, so in terms of buying, I would say anything that is a commodity, that is regulated, and that we need global scale from the very beginning of implementation—I would buy it, because it's way easier than building it from scratch. We have the experience to build it and make it scalable from day one. So in those cases, I would say buy for sure.
In terms of building, I would say the business logic, the definitions, the context, the workflows—everything that the AI needs in terms of information on the strategy and tactics of the company and how we make decisions. Everything needs to be in-house. We are building on top of it.
00:05:19.330 — 00:05:21.130 · Jacqueline Freedman
I wholeheartedly agree.
00:05:21.170 — 00:05:42.850 · Matías Coumet
In terms of the line itself, usually I ask one question: If I buy something and my competitor buys the same thing tomorrow, do they get the same edge? If the answer is yes, then that's not something I can create value in the company with. It's just having a short-term edge because I bought it first.
00:05:42.890 — 00:06:37.530 · Jacqueline Freedman
Hmm. That really flips a business case on its head in terms of the approach. I really like that idea and that kind of devil's advocate component. Because previously, I am a huge proponent of "don't build what you can buy," and if you can buy it and you know your company will never white-label that product internally or actually sell it and maintain it, the cost of doing business is better to purchase. And it's interesting to see where that switch is, because philosophically, the company has to invest in maintaining these products and tools and systems. I think you're in a particularly unique situation where the company is focused on doing that. And to your point, that is something your competitor can't quite compete with unless they really go all in with the same philosophy.
00:06:37.570 — 00:06:38.250 · Matías Coumet
Indeed.
00:06:38.570 — 00:06:51.080 · Jacqueline Freedman
So you've said that the hard part was deciding what to build and what to buy and what to leave alone for now. What's on the leave-alone list that you would really like to pick back up soon?
00:06:51.120 — 00:08:06.750 · Matías Coumet
There are some features that are fast-evolving in terms of model capabilities. So those models that are evolving faster—if you actually build them, you will build something now and rebuild later. It is a duplication of work and a misapplication of the resources you're using. So in some of those things, I will leave it alone.
As an example, by the very beginning of the AI implementation, companies wanted to create their own LLMs or have them locally and build something on top of it. And the features that OpenAI, Anthropic, or other models were creating were faster than the value they were actually creating inside the company. Therefore, I would prefer leaving that alone for sure.
Also, anything customer-facing that has no human review. One of the variables that I put on every agent that we are creating here is that we can have the creativity of the AI, we can have industry best practices, and we can have a lot of context, but never in terms of execution. The AI should work by itself, always with human interaction for giving the green light. Okay, this is something that we can put directly to the customer's eye. So that's the way that we are measuring if it makes sense or not to have the agent as well.
00:08:06.750 — 00:08:35.870 · Jacqueline Freedman
I wish more folks would think like that. You had a ten-year-old stack built on over 80 internal services, which to some would be a nightmare or a really great opportunity. And yet the people who built those internal services were still in the room. Was buying a messaging platform a technical decision, a political one, or was it just the most obvious because of the regulatory components?
00:08:35.990 — 00:09:14.430 · Matías Coumet
Certainly both. I kind of pretend it's only technical, because the change management process that we faced during that implementation was brutal, so I have to be honest about that. From my point of view, the main reason why we walked and implemented it is because marketers shouldn't be opening a ticket to change a message or modify an asset. From my point of view, they should do it themselves or with the help of AI. So we wanted to free those engineers from the plumbing, from the maintenance, and get them creating new features inside of the internal services, and wanted them to build what actually makes a difference.
00:09:14.510 — 00:09:45.020 · Jacqueline Freedman
I am so glad you're able to help change the thinking, because to your point, if everything was built in-house, it seems a bit foreign to be like, "Ah, well, abandon what we've already created and built," and yet it's maybe not—to your point—up to what your competitors would be doing. And it's like an interesting kind of flip on its head: Okay, what are our competitors doing that actually we should consider, and that might be one.
00:09:45.060 — 00:10:41.940 · Matías Coumet
The only thing that I wanted to highlight on that one is that by the very moment that you're starting to say AI, it was like the implementation of the internet back in the day. I'm an old school guy, so I know in many cases how it was implemented as well. So it's part of my experience. And most of the people were saying, "This is going to take my job, so I don't want to implement it." They were putting some blockers into that implementation.
But I just have to say on this: this is not a tool that is short-term. This is here for staying. We are changing the paradigm, the way that we are implementing tools inside our companies. So the first reaction from anyone indiscriminately is protecting themselves in terms of defense. The next step is: this is not for replacing humans—it's for giving them tools and making them more efficient in their jobs.
00:10:41.980 — 00:11:12.500 · Jacqueline Freedman
I've been in those shoes where you have to change everyone's mind. But I was fortunate in that the people who built it had moved on already, and so there was less personal feelings involved with a business decision. It can make things really tricky because you sometimes are asking people to get rid of their own creation and child that they are proud of. And that is beyond politics. That's a very tricky change management situation.
00:11:13.140 — 00:11:13.940 · Matías Coumet
Indeed. Indeed.
00:11:13.940 — 00:11:25.060 · Jacqueline Freedman
The big initiative that you and your company have right now always has a name. And the team picked Foundry. Who named it and why?
00:11:25.220 — 00:12:18.890 · Matías Coumet
I think it was a pitch from our CEO for the board of shareholders. The main idea on the naming of the initiative was to set a foundation of everything that we wanted to implement in terms of AI. Why Foundry? I take it as a metaphor of melting down the processes that we had before and recasting them, because we didn't want to create a tweak inside the company. We wanted to remake everything.
And by implementing AI again, we wanted to make sure those processes were updated. And as we mentioned before, the architecture was outdated as well. We wanted to build a new infrastructure for AI as an AI-first company and also make sure that those workflows that we were creating in the future were part of this new foundation that we wanted to set for the long term.
00:12:18.930 — 00:13:14.170 · Jacqueline Freedman
I am so grateful that that is actually how things began, because to your point, oftentimes these things can come from the top down, but they don't think about what it actually means. And oftentimes it's an add-on—an initial add-on—and it's just building upon tech debt that is weighing the entire company down. But instead, if you—I mean, to the definition of Foundry—like really melt those metals down and start anew with the same ingredients, but with a fresher set of eyes and a more... You learn so much about what a foundation of architecture can be after doing it and outgrowing it. You learn a lot of the mistakes along the way.
So this came from the top down. What is the benefit or perhaps the negative of it not coming from the bottom up?
00:13:14.290 — 00:14:49.229 · Matías Coumet
What it allowed us to do—the implementation itself—in terms of pros: it was the speed at which we wanted to implement things, because the processes that we were building before were taking long. In terms of priority, it was a complicated situation and discussion among all the peers. But now, if the champion is coming from the very top and this is going to be the top company project, everyone needs to build on top of it.
So when we're talking about speed, it was way faster. In terms of priority, it was top priority. So we need to make it happen as well. And then the permission to switch off other internal processes that were outdated, that were very inefficient—so inefficient that we needed to make it happen as soon as possible.
In terms of cons, like what it cost us: people complying with it before they believed in the process, like what we discussed earlier. When we are trying to implement AI in any company, the first reaction is defensive: "I don't want to implement this." Those kind of compliance that I am referring to—before they actually believed what we were actually implementing is a new strategy—if I ask one of my teammates and the people that are working with me if they could work without AI right now, they would say no. So we changed the mindset of those people in, let's say, eleven months for sure.
And in terms of the mandate, it opens a lot of doors.
00:14:50.230 — 00:14:57.070 · Matías Coumet
So the value is that the champion itself decides what stays and what's not. So you can build on top of it.
00:14:57.070 — 00:15:49.470 · Jacqueline Freedman
I love that, because typically in these situations, something coming from the top down is met not only with apprehension like, "No, I don't really want to do that." And so it's nice to hear how it can be a different outcome. But it sounds like the reason—well, maybe not the only reason, but the main reason why I see it being particularly successful—is not only was it the top down, but the agency was given to everyone: "Hey, this is our primary company focus. You have permission to explore this and make things better, and you might fail. But it is with the permission of: we have to do this better. So let's try it with the ways and things we could do." Is that a fair kind of assessment of how it was received internally, after maybe the initial—I don't know if shock was the right word—but...
00:15:50.550 — 00:16:16.990 · Matías Coumet
I agree. It wasn't totally like that. That was the way that it was implemented, actually. Because if it came from an idea below, but on top—as we were discussing before—and in the company, they had enough power for getting into those really hard conversations. And I'm not talking about the headcount; I'm talking about the ways of work, the way that we interact with other departments. As it came from the CEO itself, it was plainly easy.
00:16:17.030 — 00:16:39.580 · Jacqueline Freedman
Yeah, and it's unusual to me. I feel like normally when it comes from the top down, it's like, "Oh yeah, it's a mandate, but okay, anyway." So I appreciate that it was very much, "This is how we're going to do it. And here are the keys to the kingdom. Don't break it. But if you do, let's fix it."
00:16:41.100 — 00:16:56.580 · Jacqueline Freedman
So your CEO also said that this was not specifically about cost-cutting. It was actually about speed. But your ROI model measures every agent against headcount. From where you sit, which one is it? Is it both?
00:16:56.860 — 00:18:07.380 · Matías Coumet
Okay. So in terms of speed, the CEO was really clear in terms of that kind of priority publicly. The idea and the motto that we are using every day is: speed over comfort. We build things, we fail, we rebuild, and we go through an agile methodology for the iterations that we are doing inside the company.
So let's say that we are not talking about the measure of the ROI of the agent versus the person as a replacement, but as a complement. Every agent right now earns their place. So we have these, let's say, virtual stakeholders that we are building, and they have a human owner. So there's a person that is responsible and accountable for everything that goes well and wrong, especially the wrong side. Because the thing is that the real win for us was the capacity on the same team. Certain things that we were taking on—in terms of projects, for instance—were taking months to build. We're building them even faster. So in terms of efficiency, it went really well. But the most important thing is that we are actually getting harder problems and we are resolving them faster than we would have been doing them before.
00:18:07.420 — 00:18:27.690 · Jacqueline Freedman
That's amazing. And to your point, it sounds like RACI—so responsible, accountable, consulted, informed—is kind of the backbone. It's the spine of how it all works. Has there been a particular strategy around defining that and centralizing that information that has been particularly successful?
00:18:27.690 — 00:19:24.330 · Matías Coumet
So we have a particular framework that we are going to do. The idea is that any project that we are building needs to have certain structure and needs to follow the framework itself in order for it to be successful, but not only in terms of success, but also for it to be approved. So the way that we are working right now is in terms of the goals, in terms of the results of those kind of projects. But the most important one is the ownership.
So as you mentioned, we have a RACI for every project stating specifically who's the DRI—which is the responsible for the project—who's going to be responsible for certain other things inside the project itself, who's supportive, who's going to be informed, who's consulted. Everyone is on the same page inside that project as a document. And by the very end, when we are analyzing the success or the failure of a project, we also have everyone that already worked on it.
00:19:24.370 — 00:19:25.720 · Jacqueline Freedman
Sounds like a dream.
00:19:26.800 — 00:20:02.640 · Jacqueline Freedman
Sounds nice that everyone actually is aligned with how we go about this and the way in which it will be done or not done. So this is really helpful for figuring out the why, but I want to understand now the how of things. So let's open up the hood. Let's get into the plumbing and talk about the actual tech side. You ran the messaging migration. You built a context layer on BigQuery, and you set up an agentic operations layer all at once. A couple things, but a lot of things. Most teams would have sequenced that. What broke if you're doing it all at the same time?
00:20:03.280 — 00:21:28.910 · Matías Coumet
We did it in parallel, and that was a deliberate decision, because I don't believe that AI should be separated from the data. As long as you have the data that is polished and is structured, we can create AI workflows on top of it. Our data was not structured, so we needed to do a unification at the same time that we wanted to implement this big project from the CEO. And he was a champion—he was also the champion on this unification. So we wanted to build everything in terms of context as well as we will be implementing that in the future.
So we wanted to shift the way that we were working on the data and create a Foundry, let's say, a foundation of everything for the AI afterwards. In terms of breaking—what didn't break was the technology. It was fine in that matter. What broke was the definitions and the ownership. We were used to working in a way before in which every department took into consideration their own agenda, their own KPIs or OKRs, depending on the situation.
But what we were changing was that there were so many gray areas afterwards that we needed to change the way that we were defining things and who is going to be taking ownership of things that would be affecting the departments at the same time. So deciding on who owns those kind of definitions came first. And we realized by the very moment that we were implementing this.
00:21:29.110 — 00:22:20.310 · Jacqueline Freedman
That is both so realistic and so helpful, because to your point, I think everyone gets excited by the buzzwords of AI and doesn't realize if you don't have your data, it's not even that it needs to just be clean. It just needs to be unified and organized and actually usable and legible. And also that in and of itself is a huge investment to make happen. And so it's pretty impressive how quickly that was transformed.
And I guess the question I have is: One, how did that get done at such scale? Because it is not an easy process. And two, what is the current process or guidelines to ensure continued data cleanliness and hygiene?
00:22:20.310 — 00:23:28.420 · Matías Coumet
So we decided to put a joint venture, let's say, in different business areas. Each of them were independent in the way that we were handling the data. We decided to just move everything into one warehouse, make it happen in the fastest way possible, but making sure that the data quality was not lost in the migration part of it.
One of the main implementations that we did was in marketing itself. And as you know, we implemented a CDP inside of marketing for making sure that we only had one single source of truth, because in some cases we were seeing like different reports sharing different information. That happens. We wanted to unify that, have a single source of truth for the customer data, and have everything in one place.
Once you build that, then okay, what can we actually do in terms of automations and workflows with AI? What are the processes that are cumbersome for our internal stakeholders, and where can we actually create value that the customer will see, and therefore bring more revenue, more margin, and create value for the company?
00:23:28.460 — 00:23:40.220 · Jacqueline Freedman
Really helpful to know. But now I'm curious: When it's 2 a.m., who is getting paged when something goes wrong as it relates to these core products that have been built?
00:23:40.220 — 00:24:41.900 · Matías Coumet
So we used to have a situation in which we have a specific team that was dedicated for the issues. That's the way that many of the companies are working with that. They have a team, they have on-call—passive or active—and they just work on the issues themselves. The only problem with that is that in terms of creating APIs, workflows, and the implementation that we are doing with agents, they are not owners or they are not accountable for anything that breaks afterwards.
So what we have right now is going through an L1 system in which we have an agent that triages the problem, analyzes who the main stakeholder should be involved, and especially notifies the accountable for a workflow in particular. That way, we can follow a more thorough process on resolving those issues with the specific person that can pull the plug if something goes wrong, and the one that is an expert of—if it went wrong, okay, what can we do to resolve it as soon as possible?
00:24:41.940 — 00:25:13.260 · Jacqueline Freedman
That is so clean and so well thought through. Where is all of this information stored? So we've talked about how there's basically a RACI for every single agent component. Is there a primary wiki that has every single aspect, feature, product line, and the hierarchy within it of ownership? Or where is that data being centralized, and how easily or difficultly is it to maintain?
00:25:13.300 — 00:26:19.880 · Matías Coumet
I will put you into context, because by the moment that we were implementing this—back in January—we didn't want to marry with a specific model. We implemented several ones to test them, to check which ones were the best for each of the suitable situations, and we were building on top of it. Then we realized that at the moment that we were implementing this, we needed a third project, which was creating a context layer.
So that context layer allows us to have the strategy, the tactics of every team and every department, and the ways that we have for communicating with the customers. So with any model, we can actually replace that model with one of the others—like the one from Anthropic, which is Claude. We can use ChatGPT, we can use Gemini—regardless of the model that we are actually using.
We built the context layer for making that assumption in the very beginning. So if anything breaks, it's actually the data flow. So in some cases we may have some transformation in the data that is not working properly. So usually the owner on that one is the data team as well.
00:26:19.920 — 00:26:35.080 · Jacqueline Freedman
Fascinating. I guess my question from all that is: Were these guardrails put in place initially, or only after a number of things went wrong did we realize, "Oh, nope, we got to go a little bit back to the drawing board and fully map this out"?


