Name: From Pilot to Production: A Practical Guide to Scaling Agentic AI
Uploaded: 2026-09-16T15:16:30.282Z
Duration: 1 h 12 s
Description: From Pilot to Production: A Practical Guide to Scaling Agentic AI

Transcript for "From Pilot to Production: A Practical Guide to Scaling Agentic AI": Hello, everyone, and welcome to today's webinar. It's great to have you with us. I'm your host, Matt Hai, and today's session is brought to you by AI Magazine and AWS. Today, we're looking at scaling agentic AI in production. Most enterprises are past the why of agentic AI. The hard part now is the how, developing successful approaches for moving from a few pilots to a scenario where agents are running reliably across the business. That's what we're focused on today, exploring practical steps to scale Agentik AI in production, drawing on real world experience from Just Eat. I'm extremely lucky to be joined by experts who live and breathe these challenges and opportunities day to day, and I think we have the makings of a fantastic session with us. So please join me in welcoming Jagdeep Soni, GenAI specialist solutions architect at AWS, Anton Garvanko, senior GenAI sales specialist at AWS, and Jeroen Nius, technology manager, AI platform, Just Eat. Thanks everyone for being with us. How how you all doing today? Are you ready and primed to go? Let's go. Yeah. Yeah. Let's do it. Then let's do it. Look, quick introduction. I know we've a lot to get through. We we wanna dive straight in, but brief brief intro to each of you, if you don't mind. We'll we'll go around the room. I'll pick someone and thirty seconds to a minute. It's always nice to have some context in your own words. Jeroen, maybe you'd like to start as I'm looking at you first. Yeah. Well, you got my name right. So, yeah, I'm Jeroen. I'm a technology manager here at justthetakeaway.com. Currently responsible for the AI platform team, which is tasked with building out the Gen AI platform. I'm just the takeaway. I've been at the company for six years overall. Been in the industry for almost eighteen years now. Obviously, not always in AI, but this is my latest latest latest venture. Great. Fantastic. And, Jagdeep? Yeah. This is Jagdeep. I'm GenAI specialist solutions architect, based in Amsterdam. So I've currently worked with our enterprise customers across Benelux and Nordics. Been within this industry from last, sixteen to seventeen years, dealt with different areas of AI and now helping our customers build reliable agent platform and and architecture. Wonderful. Thank you. And Anton, last but certainly not least, Dane? Yes. My name is Anton. I'm based in Copenhagen in Denmark. I've been with Amazon for about seven years. My path was actually not so direct into AI, so I spent about fifteen years in finance, and then about five last years, in in AI at Amazon. Wonderful. I think we're ready to go. You for setting out the scene there. I think as you you individually described, we've got a really good mix in in your respective roles and I guess your experiences bringing a real good kind of complimentary business technical lens and more to the conversation. So I'm gonna throw you straight in if you don't mind and begin with a quick question that that should help frame things for for our listeners and for our viewers. Around the room, Anton, you spoke last. I'll start with you if that's okay. I guess in one or two sentences, just for some context, why don't each of you tell us why so many Adjenic AI pilots stall before production? I think, really, in my experience working with the enterprise customers, why kind of the the the project stalled before reaching the production is that they were really very well designed for, being actually the the pilots. So people really thought about it. There was a good brainstorming sessions. There was really good maybe hackathons and everything. It sounded exciting until, actually, kind of the the music stopped playing, and then people really need to to see, okay, how is this going to work in our enterprise environment with all of the tooling we have, with all of the processes we have, with all of their security requirements, and so on and so forth. So I think in, in my experience, it's really, it's almost like a like a largest wave, like, in this Castaway movie, that you really need to get over, to to get to production, which most of the projects never actually make. Thank you, Anton. Jack Dave? Yeah. Based on my experience, no, with enterprise customers, so one of the key challenges that customers are facing is around security and governance aspect. Right? So, yeah, pilots are always exciting. As Anton mentioned, a lot of hackathons generate excitement, use cases, and everything. And people in the boardroom are like, okay. Yeah. This looks really good. But when it comes to actually scaling it, rolling it out, then there are safety concerns, and there are concerns around governance. And it's not always easy to have governance implementation for these agentic workloads. We'll talk more about why that is the case. But these are some of the fundamental reasons I see that a lot of these exciting pilot projects does not run into production. Great. Thank you. Jeroen, would you like to come in there? Yeah. And I I fully agree with Jagdeep as well. So we we do a lot of experimentation, with agents early on, three three years ago. We noticed that you can get results very quickly, but when you start thinking about putting these agents into production, you know, having real customers interact with these these new types of software, and being able to sleep at night knowing that your agents are, you know, operating as expected. So observability, guardrails, but also cost, especially now that AI is becoming increasingly more expensive, are all kinds of concerns that if you don't have a proper platform to kind of manage that and show that those those insights to your your stakeholders, it's it's really hard for them to really trust these new new AI agents. Thank you, everybody. Some some great answers there and hopefully setting up some of the key challenges that we'll talk about during this conversation. I guess one of the things we're going to talk about isn't isn't is that the roadblock isn't AI's capability as such, but those kind of structural foundations supporting it, which gets us into our nice first focus, really, the role of architecture in scaling. Now we said earlier that that everyone's first agent is bespoke build. So maybe tell me, someone can jump in and and start the conversation off. What has to change to to to scale across business units? I can start. I think, whenever I think about it, it's, three things really come to my mind. So the first one is, people. The second one is data, and the third one is technology. And it's really kind of in that order, that, that we usually see the kind of the problems occur at our customers. So if we start with people, I think important that there is a culture in every organization that, kind of braces for, the big change that, Ajentic AI brings. Many businesses might have to change. Many, teams, the way they operate need to change. And so as long as people are open to change, I think all of the next things go much easier. So this, kind of people change management is at least, like, I don't know, maybe in the half of the cases that we work with is actually popping up as actually the this is where the problem is that something to solve. So really people. The second one is actually not even technology, but really is the data. So, for all of these agentic solutions, or even any generative AI solutions in general, for them to be useful, they actually need to have the access to the proprietary data of each company or each organization. Whenever this doesn't happen, then then there is really it's like becoming like a generic chatbot, for example, or something that really does does not work in the enterprise environment. The same here actually at Amazon, that, we we get to test our, own services first. And, we see also a big difference. Like, for example, we recently implemented Amazon Quick, as a desktop version, And it was implemented pretty well. And when we kind of look back at it, what how it was implemented, well, what made actually, people, like, hundreds of thousands of people use Amazon Quick, across Amazon is because somebody thought about all of the data that actually gets fed into it through the MCP connectors, through the skills, and so on, that, whenever the data is there, this tool becomes generally useful. So, really, the second one is a a lot around the data. And, really, the technology usually actually comes last. Sometimes it could be that the customer says, yeah, show me the technology. And then we're kind of in the process of conversation. We find out that maybe some of these foundations, like the people and the data, is is not yet there. Yeah. So for the technology, it's, it's, also within it. It's three, classical parameters. It's the cost, it's the latency, and it's the quality. So for the cost, Jerome mentioned the, AI solutions, they become quite expensive unless you really properly manage it, unless you set up the budgets, unless you really, tag them who is using, the AI solutions to what extent. Then there is also the the latency. Is there enough capacity to to process large, gigantic use cases? And then there is also the, the quality of all of these, let's say, responses or the actions that AI is doing. Is it really the right actions, or is it really the right information that we are getting? So are those, let in that case, LLMs, they are they good enough to actually to perform this, this task, as an agentic, yeah, kind of basic for this agentic platform. So lot lots of things, but, really, it's first people. Second, it's, the the data, and the third is the technology, at least in my, humble opinion. Wonderful. Thank you, Anton. Jagdeep, Jerome, would would either of you like to come in there and, I guess, expand on some of those or or maybe explain in more detail why some of those challenges or or or those those kind of themes that Anton discussed are the critical difference between scaling and and and staying more static? I I agree with Anton on on data. I think, you mentioned, Anton, without the data, every AI agent becomes a generic chatbot, doesn't really provide value. And a lot of companies, you know, under have underestimated that data. Data is always the tricky part, and having that data set up in a way that it can actually be used by AI agents is something that takes time. And that's why scaling these AI agents at a rapid pace becomes a problem because that data is not there. So, I agree there. Data key in in in scaling AI agents. Sure. And I think the fundamental nature of, these agentic workflows are also very different. So that's why the shift in terms of, so it's a it's a pretty steep technology shift. So we also need to, take care about how the people and process aspect is structured around the shift. Thank you. Thank you, everyone. Let's let's move on. Let's talk about, governance, permissions, guardrails from the outset, building that in right from the beginning. Obviously, the cost of not doing that properly is is is clear as we've all kind of discussed. You also talked about the importance of the of of this strong one time foundation, particularly for something like guardrails that that I just mentioned, which which can't really be brought in later. Now you might want to talk about your white paper, which which gives some pretty blunt kind of advice on this. Governance security built in from day one, not bolted on. Anton, Jagdeep, perhaps one of you would like to start. What what what does day one look like on a on a practical level when we're talking about something as as as serious as governance? And why is that bolting on such a such a problem later on? Yep. So, I think from the day one point of view, it is really important to think about agent identity, governance, observability, evaluations, as a as a part of the architecture blueprint itself. Right? Agents, we need to treat them like a human, identities. Right? So when you log in to some system, you have your identity propagated. Similarly, you need to think about agents as well. Right? Because you need to have right permission boundaries for your agents. Now same it goes for, Cartwheel as well. Right? You need to protect your, information, leaking out through your agent applications, and then you need to have a combination of probabilistic and deterministic controls. Now here, deterministic control also becomes really important, which has to be part of your architectural blueprint itself. Right? It is it is the one requirement. Right? Because the fundamental nature of large language model is that they are probabilistic. And that's the reason, I mean, the whole application stack that we have seen so far before large language model is now very different because large language models are probabilistic. So we need to have deterministic controls on on these. Right? Now, if we try to if we think about that okay. We we build an agentic application and later on we try to bolt in some controls, it's really difficult. Again, going back to the fundamental nature of how these applications are are created because last language model, they decide which tools to call. So, traditionally, we always had a a predefined path for application. But with last year's model with agentic applications, it's not a predefined path. Right? Model is deciding which tools it has to call and with sequence and and all that. So with that, if something goes wrong, then we have a pretty big last radius. So we have to, confine that last radius, and that's why bolting it later on becomes tricky because, yeah, then you are doing the, the batch work. There are other aspects as well which which makes it even more difficult because, let's say, today, you, already have your agentic application running. Right? But you have not managed, your memory. And if your memory gets poisoned, then your agents are going to work on a poisonous memory. Right? So, you are never going to get the desired result of it out of it. Then there are also, also risks and and frictions, which will come with the adoption. Because let's say you don't have observability in place. Right? And people are not able to audit what your agent is doing. Right? So that is directly connected with the overall adoption or rollout. So that's why these fundamental things, really have to be there in your architecture blueprint on your day one. You can think about, okay, let me, you know, roll out my agenting application, and later on, we will keep on adding observability. No. This will not really work because this is not like your traditional application. So these are some of the things that we have to think from the day one, and these are the things which will also make it difficult if we try to, you know, bolt later on. Thank you, Jagdeep. Anton, Jerome, a a follow-up, I guess. You know, governance, guardrails, excuse me, conversations that that are so important for for for risk leaders, security leaders throughout the organization. When you're in an initiative like this or a program like this, how do you make sure that that that you bring them along and and you you get that buy in, I guess? I think I can bring one of our latest examples here to the table. So, yeah, we have a customer in Norway called DNB Bank. It's our public reference. So the the bank, in discussion with our account team, found that, there are some processes in the heavily regulated industry like FSI, financial service industry, that are so manual that it takes, actually, like, the whole, huge department to process, in in a most manual way by just checking the documents one by one. So that process is, is, is one of the most famous process in all of the banking, industry. It's, know your customer, KYC. And, so so, the bank was kind of quite blunt in saying, like, okay. We actually, this is an internal process that we can absolutely, use the agentic AI, in the most, governed, in the most, way that that allows for observability about what the agents are doing to check the, both the corporate and the the private customers, when onboarding them. So as a result, the, the work of hundreds of people have been saved. So these people can actually do much more value adding work, at the bank. The cost of processing, the, each case, went down from hundreds of dollars to about $56 per case. And, the, the result is actually that the quality of, of, of of these checks of of now your customer has actually also improved. But just, just to show that going through these guardrails, checks, going through the all of the, like, observability, requirements, it was possible actually to launch the the real use case in in Agentech AI that, both hit exactly like the most regulated industry probably, along maybe with health care, and also it allowed to both to save and to improve the quality. So just just to give us one example in this. Thank you, Anton. Fantastic. Jeroen, we're going to come to you, I guess, some some some practical insights, if you can, from from kind of boots on the ground perspective. You you've been in the thick of it, I know, and and there's plenty to talk about. Let's set the scene, if you can, with with some examples, please. Mhmm. What did you set out to do with with AgenTic AI? I know you you talked about the importance of of of the platform in that and and platform first. Correct me if if I'm wrong in in relaying that back to you, but why don't you kind of give us the the context around that journey and where you started? Yeah. So so here at Jet, it actually started out similar to, I think, everywhere. And it said, you know, companies started experimenting with AI agents. But these were all one off projects, basically. So every time you wanted to create a new agent application or an agent, they started from zero. So you have to again, you know, Jagdeep and Anda mentioned, observability guardrails, all these things that, you know, are important to scale an AI agent had to be built up from the ground up. And at Jet, we recognized that that agents were the next paradigm, but these were still yeah. Like I said, they were building them as one offs, and they were reinventing the plumbing every every single time. So here at JET, we turned it around a bit, and we said, like, well, you know, what if we build the platform first? So if we have a platform that provides all these foundational capabilities with observability, governance, guardrails in place, that would allow teams to no longer worry about all these all these these, well, major concerns and could just focus on actually implementing, the AI agents. And we had a couple of use cases, obviously, that we we kind of, like, powered from the get go. But nowadays, we have a platform, that's powering more than 50 use cases across JET. Fantastic. And and, obviously, you talked about the the challenge there of of going from from, I guess, one agent to to many, to to scaling. Can you give us some kind of insight as as to what broke might might be the wrong the wrong word for you, but but what kind of didn't work? What what what what you found went wrong or where those bottlenecks or barriers or or or challenges came when when you tried to really reach that scale? Yeah. So I think the the the main challenge that we we ran into, was the observability part. So, again, agentics, you know, convincing leadership that that agent is actually going to provide value. That that's maybe a bit of the easy part. Right? I mean, we can show that an AI agent is automating a process, and therefore, you know, the people that worked on these processes can now do more valuable work like Anton mentioned as well. However, how do you get that trust from them? And I think you can only get that trust by showing that you're tracking your agents' work and that you basically provide these insights so that even though we've automated everything away, it's no longer a black box, but leadership knows that, you know, processes are still being followed. We're still taking all the boxes, but people are no longer involved, basically. So observability. And then, like I mentioned, trust. Right? So how do you how do you earn trust, from your stakeholders? And I think by showing that you have these partners in place, you know, that an agent is operating within the within the boundaries, not crossing them, and, teams being alerted in time when, you know, things do go south. So these fundamental things, governance, observability, guardrails, baked in into the platform, not bolted on, are are things that that well, at first, you know, we we didn't prioritize them. But quickly after the first couple of use cases, we had to prioritize them because we we knew that, you know, we have to have these at the core. Sure. And, I mean, you talk about what you changed. Was that how challenging was that? I mean, to give an example, what Okay. Was that a complex journey? I I'm I'm assuming so. Right? But but was there anything that worked particularly well along there or any kind of things that you would point to during that journey that that were particularly kind of brought particular benefit or a real leap forward or anything like that? So so I think the the leap forward for us was that as soon as we had this at the core, you know, we were not unlocking one use case, but we're basically effectively locking unlocking all the use cases. The challenge, though, is that how do we build this in such a way that any type of AI agent can actually leverage up these tools? So, that's the hardest part. We we didn't know what use cases we would be powering, let's say, a year from now. Right? So how do you make sure that you're ready for that? I think and then another challenge is that when we started out building this platform, I think we started out roughly a year ago, and this was basically the era where we had proper LLMs, but not the the LLMs that we know today. Right? The the proper agentic LLMs that the proper agentic models. I mean, Opus 4.5, for example, wasn't there when we started. And then it moves fast. Right? So with us and the and the engineering organization basically following along with all these involvements, and we we are now building a platform. How do you make sure you stay ahead of what's going to happen next? And I think that's the main challenge. And it, especially on these these these important topics is something that that is kind of challenging, and it's it's still today because it's moving so fast. Sure. And and, I mean, I guess people people listening, people watching will be interested in outcomes. We we always are. Are there any particular results that that you would point to, any metrics that you can give that that really show the the the the the kind of scale and the outcome of of of that work? Well, in in terms of scale numbers, like, again, we we we're building the platform. Right? So I don't have numbers in terms of, like, how many hours saved or how much money was saved by having AI agents. But what I do see is that a year ago, we had some AI agents, you know, on production. And nowadays, basically, everything is becoming agentic. And I think that's that shows that that our platform really enables, engineering teams at Jet to to to now, you know, be AI first, you know. And I think that that shows that we're on the right path, and that's the direct result. Without without our internal Gen AI platform, that wouldn't have been the case. Fantastic. And and and look, let me ask you. I guess people will be watching and listening to this because they're interested in adoption and the journey and and embarking on that themselves. I mean, an important point from all of this, if if you're giving insight from boots on the ground is, I guess, having been through that work, knowing what you know now, what would you advise? Is there anything you would do differently? Is there any kind of key takeaways that you would give for for for the audience, for example? That's a good question. I would I would say I I would I would really focus on, you know, building that foundation once and then scale out instead of rebuilding that foundation every single time. I see a lot of examples where they still start from zero, simply because, you know, they don't have a foundation that's up to date. So investing in a proper foundation for your whole organization is, I think, still the right strategy. Don't reinvent the wheel. I mean, we we work very closely with AWS. AWS has a wide offering of, you know, agentic services that can help you solve these problems quicker, but make sure that you don't push your your engineering teams or your your teams that are building agent applications to to start from zero every single time. Provide them as a service. And that also allows you to stay in control because if governance and observability is at the core, you know that every agentic application in your company, you know, at least takes some of these, like, well, high high impact, concerns, basically. Great. Thank you. And thank you for segueing me so beautifully into impact. Of course, proving value is is, I guess, the key part of any impact. Right? We we we need to prove that value to the organization and beyond. Obviously, the white paper we're here to talk about, the AWS white paper says that that most organizations struggle to prove ROI. A quarter, for example, cite unclear value, as as being one of the main reasons the projects get canceled. Anton, Jagdeep, maybe you want to kind of dive into that and explain why why that's so hard. It is hard primarily because, of, like a no a lack of comparison. So, in order for you to calculate ROI, and my background is from finance, so I can speak about this for four hours. You need to have some base, and you need to, to calculate what is the benefit and and find the delta, between the two. Some cases are easier than the others. So let's say, for example, there is, two sides of it. There is a cost, to the ROI, so cost of, your agentic solution, both implementing and running it. And then on the other side, there is a benefit. And the benefit could be quite straightforward. Could be, let's say, you saved the up on the staff costs. So instead of, people manually processing specific task, now, Agentic, solution does it at the fraction of the cost. So you calculate the difference, you compare it with the cost, and you you have your ROI. So this is like a success and a happy story, but this is quite rare. In reality, the if we talk about agentic AI solutions, often these are the solutions that bring the new business. And then it becomes more difficult. Like, okay. What if, we did not have the solution? How would the business evolve then? So kind of even mathematically, it's kind of quite challenging to calculate, the these deltas. One example I can bring is actually now that we are at AWS, we have our kind of main part of the Amazon is the, the fulfillment network, so the the supply chain, company. And there, what, what our teams are doing is that they are doing they are going back to basics to really to AB testing. Simple things. Let's say you have a number of products, where the, Agentic AI solutions are driving, the decisions on where to place these products, how to deliver them, throughout the entire supply chain. So you apply this, Agentic AI solution to part of the products. And then in the other part of the process, you do not apply it. And then you compare the results one with the other. And important is that in the selection of those products, there is a real mix so that there is no bias between the a and b groups. So then then, kind of in a really scientific way, you can prove, okay, this is actually my ROI, of the implementation, and this is the cost. And, and this, let's say, the payback period is, for example, a month or two. Often, the agentic AI solutions, their payback period is is really short. But that's that's a good example where it works. But, yeah, like I mentioned, not all of the cases are straightforward. So in in many cases, companies simply do not have, let's say, this muscle built, to, to do the measurement to offer ROIs of Adjentic Solutions. So my background is from finance. My, large part of my network is, people working in finance. And what I hear from them is that a lot of finance teams nowadays are actually there, asked to to to make these calculations, and they are really scratching their head. Some of them, they actually found the solutions using the Agenctic AI, so where AI supports them in in making these calculations. But important is that, also, the business understands, okay, the all of the drivers behind the, let's say, this, benefits from agentic AI solution. So it's it's it's a complex topic, but I think if we sit and speak one year from now, I I believe that there will be some really good solutions where people can just plug and play and say, like, okay. Yes. There is this ROI, solution where, where I can, just, you know, deploy it, and it calculates for any sort of, agentic yeah, solution benefits. For now, we're it's a it's back to kind of really calculating and finding a way for each of our customers. Thank you. So so so, I mean, Jagdeep, feel free to come in here. So the path from, I guess, promising pilot, being in that pilot stage to to really reaching those numbers, those metrics that are going to make let's be honest, our boards sleep happily at night. Could you define that maybe, Jackie? But could you kind of build on what Anton was saying there and and set out some of those key steps or that path for for for those listening and watching? Yeah. So, I think it is also important, to have right technical implementation for this measurement. Right? Because, now we are beyond that point where we are, you know, comparing which what is the token cost for each model and and all that. Right? Now it's like we really need to have technical implementation where we can actually measure, how much we are spending for for a particular task to achieve a goal. Right? So and and that is also a a tricky piece. Right? It's it's not that straightforward to do this because this will again be linked with our actual ROI because now it's like, okay. A company could say that, okay. We are spending so much on on, the models, but actually the return is not that much. Now to counter that, I mean, we need to have a clear measurement so that we can again refine and redefine our strategy, what to use, how to use, and and what to change when when we have new model capabilities and so on. Okay. Thank you. And, Joreen, let me come to you on on this point of of proving value. You you you've been through this. You you've given us some kind of real world examples there. How did, I guess, proving that value, showing the worth, did that make a difference for you when when you wanted to carry on, you know, into next phases or or to continue innovation and and and keep things moving? Yeah. So so so I agree. I think I think process automation using agentics. You know? I I think automating processes the the the part of proving ROI is actually not the only thing that unlocks that. I think having that trust, you know, showing that we're still in control, you know, and that we can scale up rapidly, that actually unlocked that that further investment. So, from a platform perspective, we've unlocked one use case that's, you know you know, either unlocking, or, you know, the return on investment is worth it. And now we have a platform that can unlock 50 more or a 100 more. And I think once we've we've proven that once we did prove that and we had that trust, you know, getting that that extra investment was well, I would let's not say straightforward, but it was much easier. Sure. Thank you. Guys, I want to try and bring this together to to some summaries, to some key takeaways. It's it's always good. I mean, we've covered a lot of ground and and and the lessons here are clear already, I guess. Measure it. The funding follows. We we talked about proving value and so on. Probably a lot to take away for everyone watching and listening, but but if we were going to go around the group, what would you give as a as a key takeaway, a key piece of action if if if I'm watching this webinar, you know, I put my headphones down, I want to go in and start getting stuck into some of this. I'll ask you for a a key thing that any leaders watching should should do next. Who wants to come in first on that one? I can go first. I think from, from what we discussed, what I I I liked all of the inputs from Jerome, from. What, kind of what I would like people to take away from, from this discussion is really it's not all about technology. Technology is really there for AgenTek AI. There is really good guardrails. There is observability. There is everything what you need to to, to to launch and to successfully drive the agentic AI solutions. But, really, it's these three things. So it's people, it's data, and it's technology. So really remember that it all starts with people. Starts with the culture, starts with the willingness to change, understanding what, Authentic AI solutions are and what they are not. Then really important that, you have your data in order. There is also always a question like a chicken and egg, data quality. Shall we first fix the data quality and then launch the GenTKI solution? I would say, really, it's it's whatever, chicken or egg. But, it's important to make use of data. So start small and really scale fast, when it comes to data. And only then it's, the technology, and and thankfully, the technologies are are there. Thank you, Anton. Jacky, would you would you like to follow-up? Yeah. I think I think we have talked a lot about importance of of setting up the right foundation, right platform. But I think it is also really important to have a clear matrix and categorization of different agentic use cases and the risk involved in that. Right? There are some use cases which have low risk and and, you know, which are easy to implement. There are some use cases where we may need to have human in the loop, where we need to have deterministic controls, where we need to have more governance. So if we have a clear picture, a clear matrix of this use case, and then, the controls that we need for each of these use case, then I think implementing them, on a strong foundation will be easy for, for teams who are actually building these solutions. Fantastic. Thank you. Jeroen, last word. Yeah. I mean, I I feel like I'm repeating, but, the the I would say, you know, start with the foundations. Start now. No. Start building now. Build it properly. Don't reinvent the wheel, but start with small use cases and discover your foundation from there, then set that foundation in. And that that allows you to really start scaling up, agentic AI at your company. Amazing. Well, look, thank you so much. Some incredible insights there. I I feel we're scratching the surface. I feel there's a lot more to probably give, but thank you so much. Sadly, that does bring us to the end of today's session, brought to you by AI Magazine and AWS. It goes without saying that we would like to give a big thank you to our guests for such an excellent and insightful conversation and for setting out clearly a direction for moving past the why and and into the how as we set out at the beginning. To find out more, and you must find out more, read the white paper, AWS's white paper, Europe North Gen AI, Agentic Adoption. Then talk to AWS about our capabilities engagement if you want to keep things moving forward and really grasp Agentic AI and go for it. Thanks again to our panel and to everyone watching. It's been a pleasure being with you all. Goodbye. Take care, and we will see you at the next one. Thank you, everybody.