A first-hand look at building an agentic platform at Splio

Transcript

Ken Mugrage: Hello, everybody. Welcome to another edition of the Thoughtworks Technology Podcast. My name is Ken Mugrage. I'm one of your regular hosts. Cool episode today. We're actually talking with a senior Thoughtworker from Europe, as well as a couple of folks from a client, about actually solving real problems on the ground. It's nice to get past some of the theory. If I can, I'll start with my colleague Carolina. You want to do a quick introduction?

Carolina Allende Amen: Yes, sure. Thank you, Ken. My name is Carolina. I'm a lead machine learning Engineer in Thoughtworks Spain. I was working with Philippe, François and the team on this project. It's great to be here today.

Ken: Great. Then I'll go in reverse order of rank, being a former U.S. military person. [laughs] François, would you like to introduce yourself?

François Bonicel: Yes, hello. My name is François. I'm VP engineering at Splio. Worked with Philippe, Carolina on this great project on our agentic platform.

Ken: Philippe?

Philippe Donon: Yes. I'm Philippe. I'm CTO at Splio. I'm very glad to be there too. Just a little introduction about Splio. It's a SaaS tool, CRM specialized in retail. What differentiates us from other CRMs in the market is our large customer model. It means we are doing deep learning, and if we do the analogy with LLM, LLM is predicting the next token in a sequence of text. Our large customer model will predict the next action a customer will do, what he is willing to pay and that's a bit of who we are.

Ken: Great. I appreciate that. Thank you for the intro to Splio, because that's really the core of what we want to talk about. AI and machine learning, and those types of things are obviously not new to you. I know you've done some acquisitions lately and so forth. What is it in the last year or two, or whatever the time period is, that you could tell us, that changed that made you say, "Hey, we need to take a leapfrog forward"? What problem are we trying to solve here?

Philippe: You speak about the timeline. I will come back to AI in the world of deep learning. This is a pace we took, and we are experts in Europe. End of 2024, we saw huge progress in AI for coding tools, so coding assistance. This is where we did our first hackathon in order to help our developers to be empowered with those tools. This hackathon has been amazing. This is where we really realized that our work totally changed, and that now we are no longer going to be writing code. This is a perfect transition about the agentic world.

Now, what we realized, thanks to this hackathon, the second one where we built an agentic platform, is that now we are no longer going to write code, but we are going to write agents that will speak with each other. This is why during this hackathon, we discovered a brand new world, and we discovered that our job totally changed.

Ken: How did the engineers take that? What's the conversation like with teaching people? Is it a new skill set? What was that like from a personnel perspective?

François: Like Philippe said, this is really a new world. The hackathon we did in December 2025 was really a turning point for us because we already had some engineers that started to use Cursor or Windsurf or everything. We just scaled up the experience and gave them the right AWS tools to get more power they can have to code. How they react to the first reaction is just okay. The code generated by the agents is crappy. It's not as good as I do [it]. We are six months later, and now we can say that 90% of our engineers are using AI daily, and it's growing and growing and growing.

Now we have other problems to solve, like merge request bottlenecks, like things a lot of other companies are trying to solve. The adoption is quite amazing, in fact.

Carolina: I was going to say I think you're doing a great job at that as well. Your developers seemed very savvy in the use of agents for coding, and also very targeted. Instead of giving a full task to the agent, they would say, "Hey, I need this specific problem solved." I think that's a great sign of how much adoption you've had for AI tools in the past. You said December 2025, so roughly what? A year, a little bit more? A little bit less?

François: Yes.

Philippe: Also, an impressive stuff is that we are speaking about generating code, but now it's really everywhere. The level of abstraction is pretty amazing. In the past, we will be speaking about dashboards, about observability, about how we connect with business, about our user stories. Right now we have only one tool. This tool is our harness that we use above the models. Everything is done through there. Now I don't even remember the world where I was building my queries in Grafana or the world where the developer was getting his tickets, understanding them and after writing them, we have everything connected in one sweet spot. That's very amazing, and that unlocks so many capabilities.

Also, having tasks we would not even think about doing them in the past, and now they are easy and feasible. This is more about code with AI, about the agentic world. What we call in Splio, more of the AI for internals. This hackathon that was very amazing and gave us so much results, was more about the AI in our product. This was more the subject of this podcast and the outcome of this amazing hackathon.

Ken: Actually, that's a great segue. Thank you. It sounds like it was all going well from an engineering tool perspective, but then something this year led you to– I saw a post that said something like 4 days, 1 room, 20 people. It was yourselves and AWS and Thoughtworks, if I remember right, and I'm not sure who else was there. Everything's going well. What brought you to that? What you said, "Oh, hey, we actually do need to do some course correction here."

Philippe: You are seeing every company brand showing stuff about AI they build in their product. You don't know if it's fake. You don't know if they fake it until they make it. You don't know if it's real. For us, we know that in SaaS, we speak about the SaaSpocalypse, we speak about the end of SaaS. For us right now, we see a lot of opportunities. At that point, we saw something clear, is that we have a lot of data in our software, and this is gold.

A lot of the CRM managers or the marketing teams, they are using this data to do audience, to do targeting. They also use the data lake in their own company, and they have to have data analysts to do complex queries where we have a lot of this data in our data lakes that they share with us. The thing is that we have 500 customers, 500 different brands with different objectives, different knowledge items, I would say, but different wording, different objectives. How can we make a marketing team have all this insight, all this analytics, and define a strategy for their marketing? Through AI, we saw that allowing them through natural language to query and ask their data was pretty amazing.

We already had in production an agent where we learned a lot about the usage, about tracking conversation, about memory, about a lot of stuff. Our customers were at the wow effect. We know we spotted the right agents to begin with. When we displayed this to the CRM teams, they also said, "We want to show this to our CEO. Wow, this is amazing. Wow, this was something that was taking me three weeks in the past, and now I can do it in a few prompts." Also, what is important, I say in a few prompts, but it's still something that is loading because we are loading a lot of data.

People understand now through the new AI, they understand that there are different kinds of reflections. Sometimes you can expect an answer to be very fast, and sometimes, when it's getting a lot of outcome and it's very clever, you can expect this to be long. This was one of our first loads in production because our first agents, we thought, were too slow. It was at a point of the beginning of ChatGPT where the answers were very fast, and thanks to them also bringing the reflectioning board. We could bring this, and the customers were used to it.

Let's come now to this hackathon. Our customers were saying, "Okay, that's great." We knew how to do one agent, and one agent with a semantic layer asking data. This was mainly our expertise, but our customer was saying, "Okay, that's amazing. I will save a lot of time, but I don't have enough teams after to do actionability. Now I have very good insight, a very good strategy, but I have to build many campaigns, many audiences, and many other things. How will I do that?"

This is where we started to think, "Wow," maybe we need some help about this agentic world, this orchestration, how to bring all those agents into production." This is where some tools proposed this framework. Several days together, where we do one day where we gather everybody, the product, the product marketing team, everybody to see what is the pain point of our user, what can bring the much ROI, can have a wow effect to our customers. Two days of building and one day of demoing it to our board.

François: Getting all those people from different teams in the same room, it helped us a lot because you are at the office everybody is on his daily tasks and you never get this just a higher view of the marketer journey. For example, the first day we did this great exercise about ideation and just rebuilding our customer journey is really an interesting exercise, just to all align on the pains of our clients and all align on what we want to tackle in those three days. This is really, really important.

Ken: As I understand it, especially François and I think Philippe as well, you've both been at Splio for quite a while, and seeing different acquisitions and transitions and what have you. Why is this the right moment to make this bet on agentic? Is it because of technology? Is it because of the personnel? Is it the market? Or is there any one thing, I guess?

Philippe: Yes, Splio went through several acquisitions. We are very glad with those acquisitions because they all bring different services to our product. What is very important to our customers is this one-stop shop. They can have everything in our platform, marketing automation, CDP, predictive loyalty, wallets on the phones, notifications, several channels, and this is amazing.

When you think about such a big platform without agentic, it's also very complex. A lot of screens, a lot of different products to orchestrate through your team, and a learning curve, that is important. What we strongly believe is that our product needs to be simple, and it needs to be effective and productive, and we really think that through agentic, first, our customer will be much more effective. They will be able to understand the platform faster, or maybe not understand it because they don't need to understand it, some part about generating an audience or some complex part that we were not able to simplify. Now it's very feasible through the agentic world.

Also, as you said, the market. The market is going agentic. I think we need to go agentic, otherwise, our customers will churn. This is where the fact that maybe in the past having such big products was maybe hard. Now it's even more a quality for us, even more a differentiator, because through agentic, they will be able to have everything interconnected, and this is, I think, a very powerful thing that we can ask through the agentic World.

Ken: You said something there about making it easier for people to use it and so forth. In my experience, the easier it needs to be for the user, the harder it is for us to create it, because there's more conditions to think about and that sort of thing. As I said, you started with a CRM agent. Not to confuse agent with agentic, but then ask a CRM agent where a marketer could speak to the CRM in plain language. That's a great experience for the marketer. There's a lot of things I do in plain language, but it becomes very hard for you. Why start there? That seems like a pretty hard place to start.

François: Yes, what we wanted to show to our customer is, as Philippe said before, all the data they have on our platform. This is our treasure. Let's say when they come to Splio, they give us a lot of data and they want to access it, and they want it to be easy to access it. In fact, this is a pain point we have for many years, because we are able through UI to give them access to build campaigns and everything, but there are little, let's say, tiny clues, for example, in the data and opportunities in their data. They are not able to see because, for example, our dashboards lack customization because they want to do a specific KPI. They don't have access to the platform.

With an agent, they can really build in natural language complex query to just ask our warehouse and get the response they really want. That's something they cannot have in fact today in our platform without using the agent. This was really the point, and also to show them our predictive capabilities, because those predictive scoring and everything are accessible via the agent. It was also the point to show the power of our data, in fact.

Philippe: Yes, and it's also where everything starts. Through the CDP, what we really want to be is not only a CRM, it's also like giving you all the insights about your customers, so it's really where you should start. What do I know about my customers? What customers I'm forgetting and not targeting? From there, where do I have a big opportunity? From there after you can create and you can do everything about actionability. Or, "Will I reach them?" Meaning which channel, which content, which object, which product? What is the right way to address those people? I think this is very important to start there.

I think this is also where we can address the major part of our users in our platform, because it will help everybody in our platform. Not only the designer, not only the campaign planner. It helps everybody. This is why we started this journey here.

Ken: A lot of times the hard part about any interaction with an AI agent is trust. We've all seen the stump your LLM things where simple math problems, count the number of letters in it, et cetera, so your customers are now using this voice thing. How do you generate that trust? How do you make sure that if I'm asking the Splio product, I'm asking the CRM thing, and I'm getting an answer, that I trust that answer enough to take action on it? How do you build that kind of trust there?

François: We have a really great team of data scientists at Splio that are already experts in deep learning and for years that are able to build pipelines to evaluate the results of the models and everything, so the idea was to how we are able to build an evaluation framework on top of those agents to be sure of the qualities that will be given to our clients. As you said, nothing is perfect. We all know AI hallucinations, and when we give answers to our clients, sometimes it's not perfect. After that, we have an evaluation framework. We are able to say, "Okay, we are giving 80% of the time the right answer," and after, it's also a question of feedback loop."

How do we get feedback from our customers? How can they just say to us, "Okay, this answer is not okay." After that, we have many ways to improve the answers by playing with, Philippe mentioned, it's a knowledge item or improving the company context around all that data to be sure we answer better to our clients, but this is quite a tough part of this journey.

Philippe: Earlier, you said that when you can make it easier for our customers, it's much more complex for us, and that's totally the case with what François said, so we had to build this agentic platform. It's a totally new architecture. Before, we had the database, the backend, the API, the frontend. Now we have the agentic platform. Inside this agentic platform, it's so many tools. What François said, but it's also about giving the orchestration between those agents, having all those frameworks about LLM as a judge, about feedback loops, but also about knowledge items.

Knowledge items that are common for an industry, but also being able to have a brand content for a brand, being able to know the acronyms for our brand, or the way they calculate their financial record, or everything. It's very highly personalized. It's great we are in marketing, so we know about market personalization, but it's also about the fact that what we feel, and I will speak about internal AI and AI in our product, is that we really feel like we have another teammate.

It's really what we want our customers to feel. They have one teammate that can analyze the data, another teammate that can be a designer, and that he will learn by working with them and learn about the brand content, the tone of voice, everything that these brands want. I think this is where it's also very, very, very complex.

Ken: That's interesting because that's where we go from standalone agents, so that I can have a chatbot to actually do agentic, autonomous things making decisions and so forth. I guess, Carolina, what kind of processes did we probably bring to this to get this platform and get agentic versus agents? What's that look like from a consultant? There's a lot of things at Splio that are very unique, obviously. Sorry, no such thing as very unique. Unique is a binary condition, but anyway. At the same time, we do see recurring themes in some clients. I guess, Carolina, what did we bring to the table to try to help with this orchestration?

Carolina: I think both Philippe and François pointed out something very valuable of the process that is sitting everyone down in the same table in the same room and having a discussion that is cross-sectional between both the C-suite, product team, the data scientists, the machine learning engineers, the software engineers. Everyone is on the same table discussing the same topics.

Any side of the problem that in another occasion might have been overlooked, let's say if it was just software developers putting together that agent, maybe they haven't thought about some side of the product type, that gets sorted out on the spot, or at least we have someone to ask the question to and spark a nice discussion. Then that means that whatever is prioritized to be built is something that is a consensus among the entire team. It's not something that tech said, "Oh, an agent here would look cool." It's something that tech said, "Hey, we can build this." Product said, "Hey, there's value here." Someone from the C-suite said also, "Yes, our clients would love to see this working."

I think the main value of the program is the alignment that it brings and the fact that we're all building towards the same goal together, and everyone is trying to unblock each other. Of course, in the process of building these things, new questions pop up that one can't foresee when you start designing. Everyone is very ready on the spot to help out, unblock, gather any data that might be necessary, any insight. We need to ask a question to some other member of the team that is not in the room, is in a different city. Don't worry, we'll figure it out, we'll find the information, and we'll build what needs to be built.

Ken: What I'm hearing here, which I love, is that it's not a proxy for those businesses. It's not like somebody who thinks they know what this other person is doing, but the actual stakeholders.

Carolina: Exactly. I think that's a very valuable point of the process as well. It's not some software engineer thinking they know how a marketer would think or what the process of marketer follows looks like, but it's actual people who work with the marketers and people from the marketing teams who know what the process looks like, who will tell us, "This is the correct way. This is not how we do it. I would've never thought of that." Yes, it's very much people who are very much involved in the process, let's say.

Ken: I know you did a lot of value stream mapping. The things that you actually do differ greatly sometimes from what the document says you should do. I guess, Philippe, as CTO, how do you get those people in a room? That's an expensive meeting, right? If it's all those stakeholders, how do you get organizational buy-in and get those people to show up for multiple days?

Philippe: That's a very important question. First, I did it with François. We thought it was a very good idea. We did it without asking for any permission because we have a schedule, we have sprint planning, we have stuff. We decided, okay, that amount of people are going to disappear for four days and do some things that we didn't work on the prioritization and we don't know yet what it will be. It was very important because, on the other point, without having this hackathon, I was not clear about what I wanted to put all the engineers to build in the next 6 or 12 months, knowing that we are in a race with other competitors and SaaS.

One of the objectives is to have a clear point of where we want to go. Thanks to this hackathon and thanks to Thoughtworks, we had it. The first success was the fact that, yes, we had a lot of people in the room, so a huge cost for four days, but we had to do a demo at the end. With this demo, we had to have a wow effect, because I also had to convince some people that were skeptical. They were saying, "Why do you want to spend so much time or so much money in agentic where we need this feature that we can sell, this feature that is missing, or this feature that we would like?" People were like a bit, "Are you sure you want to spend so much money in agentic?"

Another thing that I want to thank Carolina and her team is effectively, we are in a room, we know we cannot stop, we know we have to take shortcuts, and we cannot get blocked. Having experts in the room as go-to points where you are totally blocked by the infrastructure, and they can unlock you in a few minutes, it's pretty valuable. Also, about the value stream map and all of it about the different agents, also we had very interesting questions from Johnny or from Carolina. Johnny was a product strategist, about how much time do your customers spend creating campaigns, how much campaigns per month they create.

We got numbers that we were looking at, but not as, "Wow, our customer, if they are building 400 campaigns per month, they have to be spending a lot of time on the platform to do that. If the agent does it, what will be the willingness to pay?" Because they will save a lot of time, but it will cost a lot of token because we are speaking about complex tasks. It was not only about figuring the strategy or getting the how-to, it was also about finding the willingness to pay, the ROI, and giving to the executive team a clear strategy, a clear map, a clear investment, and a clear ROI on this project.

This was very, very amazing. This hackathon was a few months ago. We accelerated so fast. Now we have a clear vision, and we have been able to bring the full company with us because they saw a prototype working. They said, "Okay, this is what we need. We can put all the teams on that." Putting all the teams without having a framework, without having anything, was pretty hard.

I'm pretty happy with the result of this hackathon and also about the human experience. This was just before our party, but we also had drinks on the first day. The second day we had a meal in the office. It was pretty amazing. The energy in the room, you could feel it. Yes, as you said, strong preparation. Thanks a lot, François, again, for the preparation. It was months of preparation to ensure everybody has all the pre-requirements in their laptop, and have all the environments set with data that we can use. This was a huge preparation because, yes, we didn't want this to fail, but if it had failed, a lot of learnings come from failing also.

We are not scared, and we are lucky in Splio to allow people to try and to fail. We know that if this was a success, [chuckles] it's better, but if this had failed, we know that there will be learnings and it will be good enough.

Ken: I'm really glad you said that because it's the old cliché of, as I said a little while ago, making something look easy is actually quite hard, and it goes with these workshops and stuff too. François, spending all that time doing the preparation, I would encourage our listeners too, this isn't just put people in a room and start vibe coding. You may need to think about different things than you're thinking about today, but you still need to think about them deeply. To quote Martin Fowler, "Actual intelligence is still required to use artificial intelligence."

I guess to anybody, really, what about process changes? If we think about not the product side of it, but even the development side of it, I can't just take an agentic model and point it the same way I did a backlog management before, or when it comes to that customer interaction. If people are clicking on a user interface, I can predict where they can go from Step 1. If I allow them to speak, they could skip to Step 8. What does it take to think through the process improvements that have to be done here?

François: That's a really good point, the deterministic versus nondeterministic part. We are still learning a lot about that. We discovered it during the hackathon. Meaning, trying to use only one model, trying to tune it via system prompt. It was quite really manual and really hard-coded and trying to avoid all the complexity about multi-tenancy about security during the hackathon. All those parts were a bit avoided. We just took so many shortcuts to be sure to have a wow effect demo.

Now we are really making this happen in production, so we have to take all of that into account. Security is a big thing. Being able to validate certain scenarios, like if I give the agent this prompt, I want to be sure I will get this answer. Even so, that was so easy. When you build APIs, when you build a UI, you really enter a new world. As I said before, we have a lot of work on the evaluation framework to make sure when we make those small changes, it would be in tools descriptions, system prompts, connection between agents, moving the models because we think it's too costly or things like that to be sure we don't drift. We are still learning a lot about that. It's a new world.

Philippe: What we are trying to– because you have to figure out how you want to put the AI in your product. For us, we have a huge enterprise, we have SME. We think that they want to keep control. What we built, and this is also the idea we had during the hackathon, is that you have a chat where you can ask your question in natural language and everything. At the end, what it creates, it's still in the interface of our product. If you create a workflow, if you create a design, if you create whatever, you still have things.

I would say the legacy page. Is not the legacy page. It's the SaaS page. You can still interact. You can see the results that the agent is building in Canvas, where we see our actual SaaS. You can see what has been built. You can double-check because it's not only saying, "I have done that." Because sometimes it says that, but it's not what it did. We can still take control. We can still have the clicks, the keyboard, as you said, that we had in the past. Also, what we need to understand now, before the mouse and everything, we had the heatmap. Now we have to do analytics on all the logs of the prompts.

It's very different because you will see how much prompt goes to analytics, how much prompts go to designer, how much prompts go to this agent, and the result and everything. The other thing that we have to get prepared for is that we are seeing many and many customers that say, "I want to do the actionability outside of your platform." To be connected through MCP, through whatever, but be connected to other agents, which is another kind of abstraction, or be connected to their harness, for instance, Claude Code, or Claude, or Codex, or whatever, or ChatGPT, but they want us to be connected to their harness and come to the platform for some usage. They need to see a dashboard and everything.

In the daily, have their agent be able to pick some data from there, pick some data from there, and after, act somewhere else. This is what we are going to see, and this is also losing a lot of control. This is also having less people actually browsing your SaaS platform, but you still remain very important. This is also where we want to go because we think that if in a company, you are part of this ecosystem and you are connected, a lot of people will be using our data without even saying, "This is Splio. This is this company. This is also our SaaS." It's going to be a very interesting world.

Actually, yes, as you said, right now we are typing, but I don't know what will be the interaction in the future. I am using much and much more voice to interact, much more agents working by themselves, and you just interact with them remotely. There will be a lot of stuff about long-running tasks, remote interaction, async interaction, so let's see. Yes, we have to be watching the trends and looking where it goes.

Ken: Carolina, I think our listeners know we try very, very hard not to turn this podcast into advertising at all. When we showed up at Splio, we had our 3/3/3 process that I know we've written about with Splio on the website and the Amazon platform. What is bringing those tools– How does that help? How does that accelerate? Why do people want to choose a platform that's out there versus just build their own?

Carolina: There's several parts to that question. First of all, you want to build your own platform to maintain full control of everything and customize it to the detail. You could just use a platform that's out there, but that will give you a limited set of capabilities. Let's say you might need a new one, your business domain might need a capability that a platform that's out there doesn't offer. That's why it makes sense for you to build your own. Then tackling the process of building your own platform is a huge problem in and of itself.

How do you know where to start? How do you prioritize? Sure, I want to build this platform, how will that platform be used? What is the journey of the user of that platform? Where does it make sense to start building? I think in the context of building that platform, gathering insights from the data, like Philippe was mentioning earlier, makes sense to be the starting point, but even when you have the starting point, how do you continue building? The idea with the Agentic Catalyst Program, which is the workshop that we run with AWS and with Splio back in May, is that Splio already had an idea of what they wanted.

They wanted this platform that was going to run the CRM in management, but we needed to sit down and break down, "Okay, what is that platform going to be used by? What are the phases of the usage of that tool? Fine, now we have that, who uses it? Who are the users of this platform that can say, 'Yes, this platform is useful. This is the way I used to work. No, this is not. You missed it here. You got this right.'"

I think we talked a lot about it already, getting everyone in the room to discuss the problem. Once that is done and we're all aligned, do we have the data that is needed for that? For Splio's case, yes, we did. We managed to get some data, we could build something. We also assess not only the feasibility of one of the use cases, because throughout the entire value chain, there might be more than one use case that could use agentic AI. Which one do you choose? We do a prioritization where we say, "Okay, this is really valuable, but how complex is it to build?" Because to bear in mind, we only have 48 hours-- well, actually two working days to build this.

Even if, in our case, we were about 10 engineers in the room, it's still a lot of work to be done in 2 days. The idea is we go for the use case that is a nice compromise between complexity and value delivered. The other valuable part, I think, and again, we mentioned this in the past, but I think it's a very important point, is the ownership of the solution, because the engineers from Splio are involved in the building of this from moment zero.

It becomes not only us helping them, but it becomes personal for them as well, like, "I started building this. I want this to work. What do I need to do to get this working?" Which also means when we leave the room and they continue building, they have the foundations of how that was built, what discussions we had. They also have an idea of how we can improve this in the future. What needs to be built to extend this project.

Philippe: Everybody got a lot of learnings. For François and I, we really wanted to see clearly the vision, see clearly the architecture, and see clearly what are the next steps, so we got it. The developers, they totally got the funny part of learning this brand new world with experts and some funny stuff they showed us. They also helped us with peer coding, so we accelerated a lot. PMs also understand much more about pricing, willingness to pay, a lot of stuff about the value that it can bring.

We were speaking about getting all those engineers in the rooms and the atmosphere. I really remember where 10 minutes before we had to close, we said, "Okay, we are closing in 10 minutes. You already worked a lot. You didn't want to take a break because you wanted to finish on time," and it was not working. We had a pain point. We were not making it work. One minute before the end, I remember this noise like, "We got it. It's working now." We were like, "Everybody behind the developer who found the last friction." This moment was maybe the highlight of the hackathon.

Yes, definitely if you are in this agentic journey, you have some parts that are still blurry, you don't know how you can bring all your teams or you can accelerate. Actually, the name is very good. The Agentic Catalyst Program. It has been a catalyst.

Ken: As we get close to the end of our time, one thing I always like to ask is whether these episodes go live on a Thursday for most of the world. I ask each person, for the listener out there, think of your role. Philippe as a CTO or François, VP, Carolina as a consultant. Monday morning, what could people do? What should they be thinking about? Reading a book is fine. Setting up a session is fine. Again, I'll go in reverse order here. I'll let Philippe go last because that's always the hardest one. I guess Carolina first, if someone's consulting, someone's working on this type of project, trying to solve this kind of problem, what's their first action?

Carolina: For me, the main take from this is looking at problems in a different way, because working in tech, one tends to be very narrow-minded. I don't mean that in a mean way. I see it in myself as well. Sometimes we all get really caught up in the exact branch of tech that we do. We just think about-- I'm going to talk about machine learning models because it's what I do, but it could be anything. I just keep thinking about how I'm going to build the next model and deploy it and use it. Maybe pausing to think, "Why am I building this for?" Not in a philosophical way, but rather how does this fit a bigger vision? Why are we building this? Who are going to be the users?

Does this serve a purpose in the sense of, "Is this usable by a user, or am I producing a model that I'm just going to run through the terminal that is not going to be friendly?" Are we building something thinking of the end user in mind, even if there's no experienced designer involved in the project? I think that shift in the way we think about problems is the main takeaway for me. It's an active effort day after day.

François: I totally agree with what Carolina said. I will just say one takeaway for me beside all the new technology we learned beside a lot of new frameworks, new aspect of our job we learned is after those three days, how you keep all this energy up to be sure you will have everybody on board when you click on now we go for real in production. This is really important. Just after those days, we were saying, "Okay, now we have to prioritize." Is it the right moment to put everybody on board? We are not sure about that."

We have to just schedule meetings with engineers to say, "What are your questions after those four days?" Technical questions, we have to deep dive together to check what we have to learn to gather a lot of information to be ready when we will push the button and say, "Now it's prioritized. We have to do it." This is really one important thing after those days. Keep the energy in your company, I would say.

Ken: Philippe.

Philippe: Thank you. Enjoy the journey because it's a new technology, a lot of stuff to learn. We are able to do so many things in parallel. It can be very overwhelming. Often, maybe on Monday, take five minutes to take a breath to enjoy the really amazing moments we are living as engineers.

Enjoy understanding that our life is changing, and try to keep a good balance between work, your personal life, because it can be very, very overwhelming to have agents doing 100,000 stuff at the same time, and be sure to take all your team with you. Because it can be very easy to have the ones that get it fast taking a huge advance, but to leave some people behind. You don't want that because when you want to carry and bring everybody with you on this journey, that is amazing.

Ken: Again, thank you very much to all of our guests. We really appreciate your being willing to share your journey and what happened here. I would encourage listeners to look at that. We've done a case study or white paper or whatever the right words are on the website about this, and do follow up if we have any questions. Thank you very much.