A conversation with Michelin: Scaling AI in software delivery - S2 E4

Building a plane while flying it one thing. But what happens when the blueprint redraws itself every month? That is the reality of software delivery in the age of AI. 

In this episode of CGI's From AI to ROI podcast series (season 2, episode 4), host Fred Miskawi, Vice-President and Global Applied AI Lead in CGI's CTO organization, is joined by Julien Millau, Distinguished IT Engineer and Architect at Michelin, and Kevin Beaugrand, Director, Consulting Expert at CGI. 

They explore what AI in the software development life cycle (SDLC) looks like in practice, not in theory, at Michelin, one of the world’s leading manufacturers. The conversation moves beyond tooling to examine the organizational, cultural and measurement approaches that determine whether AI ambition translates into real business outcomes.

To learn more about AI in software development and delivery, visit our SDLC page. 

Discover how AI is transforming enterprises and government organizations. Visit cgi.com/ai for insights, resources, and news on AI-driven strategies.

Transcript

Introduction 

Welcome to From AI to ROI. Now imagine building a plane while you're flying it. Now imagine the blueprint redraws itself every month. That's what software delivery looks like today in the age of AI. The tools you mastered last quarter can be obsolete. The processes that define your team for a decade suddenly up for renegotiation. And the target? There isn't a stable one anymore.

I'm Fred Miskawi, Global Applied AI Lead in CGI’s CTO organization. And this is season two, episode four of From AI to ROI. In the last three episodes, we looked at the hype, the promise and the ambition of AI in software delivery. Today we're doing something different. Today we're going to go inside one of the world's most iconic industrial companies to hear what it looks like and what it sounds like when the rubber meets the road, literally. 

My guests are Julien Millau, Distinguished IT Engineer and architect at Michelin, and Kevin Beaugrand, director consulting expert at CGI. 

So we're going to talk about why the technology is easy at this point, or at least easier. And why it's no longer about the technology anymore. Why upscaling at scale is harder than it looks. Why accepting that AI will make mistakes is one of the most important leadership decisions you'll make this year. And why in Julien's words, as we were preparing for this podcast, AI still needs a pilot.

So if you're a leader trying to turn AI ambition into measurable ROI, this conversation is for you. Julien, Kevin, welcome to the show.

To set a little bit of context, can you briefly describe your role, the teams you work with, the products you build and maintain, and the tech stack you use today? Julien, let's start with you.

Julien Millau (02:06)

Yes, thanks for the introduction.

As IT Distinguished Engineer at Michelin, I have a transversal role across multiple teams. I am primarily focused to help and support teams to adopt AI and enlisting them to transform the way they work today.

Today, we deploy AI on Michelin, we use GitHub Copilot for software and for software developers. And for that, I work closely with them, and the engineering, product and platform teams as technical advisor, to help them and to give them the best practices to develop and to develop with AI. In parallel to that, I shape and promote a long-term vision

of the software developer of tomorrow and how we can leverage AI today into the software development of the cycle.

Fred Miskawi (03:07)

And Kevin, what about you?

Kevin Beaugrand (03:08)

So yes, I'm Kevin Beaugrand, I'm a technical architect working at CGI mainly at Clermont-Ferrand. So very close to machines and to Julian's team. I'm part of the directors that are in CGI to develop and deploy AI across our organization. I work with many customers in my business unit and I'm also as a developer, as a technical developer, especially interested about the transformation around the SDLC by using AI. And I'm teaching and I'm also helping some organizations and teams to embrace AI for their organization and to be sure that they can use AI in order to provide more efficiency, more productivity. But in a way, they can continue to host their applications and to maintain their applications in the long term. So we will discuss that today.

AI in the SDLC: The Journey So Far  

Fred Miskawi (04:13)

Effectively, setting the foundation for what needs to come after. That makes sense. Julien, maybe you can start us off by explaining a little bit about the software development lifecycle and how you're leveraging AI within that lifecycle.

Julien Millau (04:31)

Yes. So today we are mainly using tools to use AI. We are focused at the primary level on developers. So we use copilot, code assistant, test generation, documentation. And after that, we realized quickly that tools alone don't scale. And for that, we need to embrace everyone. We need to onboard everyone to ensure that everyone that acts and work on the software development lifecycle also works with AI. So for that, we deploy a structure, we'll talk about that just after, a structure and foundation to ensure that we can develop easily with the tools, with the business, with software developer, with SRE. So all the people that act on the software development lifecycle and make a story, a journey of working with AI on the software development. So for that, we have a lot of things to do to scale that new way of working with AI and something that is continuously evolving.

Fred Miskawi (05:45)

And Kevin, from your perspective, not only what you're seeing in the account, also what you're seeing with other accounts. How is that software development lifecycle evolving and how AI is being introduced in the various tools that we're using? 

Kevin Beaugrand (06:15)

Yes, we saw that we mainly two ways in order to use AI. The first is the vibe coding. We saw many organizations that are testing the vibe coding and using that for prototyping, for example. And this kind of usage is very interesting because they can use vibe coding to increase their productivity into prototyping, but this kind of project does not pass the production phase. The main difficulty is about ensuring that the code is production ready, the code is maintainable, the code is observable, and so on, before getting that to production, and they are struggling in that phase.

As a second part, I saw many organizations that are testing and embracing slowly, mostly slowly AI into the development phases because it is in early stage or it is quite a little bit in production already tools like GitHub Copilot. But they are also looking about using AI in the first phase of the SDLC, like the user story wrapping, task chipping and so on. So this kind of usage is more slow for now but I see an increase in the investment on that part.

Beyond Tooling: Organization & People 

Fred Miskawi (07:46)

I guess it's not just about the tooling, it's about the approach. Julien, you were mentioning the methodical approach that you're taking to make sure that everyone within the organization understands how to leverage these new tools. What else can you tell us from that perspective?

Julien Millau (08:02)

We need to onboard everyone and it's also fundamentally a people and organizational topic. So for that, we focused at the beginning on software developers because it was the more simple way to deploy that. As Kevin said, it's very simple to deploy AI for prototyping, for UX research, for having something very quick, but when we want to scale that and to improve our production, improve our application and deliver more features to our business, we need to take care. And for that, we need a pilot. This is why GitHub Copilot, I love the name. We need someone that drives the AI and also reviews what is done by AI. We have the same thing with Anthropic, with Cowork.

We need someone to work closely with AI. At this point today, we need someone to help AI and we need someone to drive AI. And so we focus at the beginning for software developers and after that we say, okay, it's not only a subject for software developers. We have a tool for software developers that is GitHub Copilot.

Now we need to onboard also all people around the software developer, like functionalists, like product owners, like project managers. And for these kind of people that are less technical than software engineers, we need to give us tools that can interact with many tools that we have. So we put an open source tool, LibreChat, that has ChatGPT-like interfaces where all the less technical people can directly interact with many tools that we have like work source code repository in GitLab or the incident management system that we have and with that, it can also use AI to refine for example its issue, its epic that you want to create and you can also help to understand what the developers do for the code. So we take that for the whole lifecycle and not only per job profile.

Stakeholder Relationships & Evolving Expectations

Fred Miskawi (10:35)

And, Julien, what has changed with your relationship with stakeholders? So your teams, the groups that you oversee, they have a certain amount of value that they have to deliver. And because of these changes and expectations as well, we've had to rethink a little bit how we interact and report to stakeholders. How has that experience changed for you?

Julien Millau (11:02)

It's a big thing because it's an organizational project, so we need to deploy that. So the stakeholders are very wide. When we speak about AI, the stakeholders are the security team, the legal team, the human resources team. So it's not only the projects that work with the software engineers that are on board. Each time someone wants to use a new model, we need to review the security. So the stakeholders in this deployment project are quite wide. If we come back to the software development lifecycle, all the stakeholders of the development lifecycle, we need to prove to them what is the value of using AI in their jobs. So I take an example, we create tools called Bibops, that is an agentic way to interact with many of our systems and to collect data to answer to an incident that we have in production, for example. And for that, we need to prove to that AI is a good help for them to make their day-to-day jobs. Each time we want to deploy something, like any change project, we deploy agility, when we deploy DevOps, when we develop any transformation into the organization, we need to prove what will be the gain for this job with AI.

Fred (12:42)

And given this relationship that you're building with your stakeholders, I would imagine that expectations are changing or have started to change. Have you seen this so far in terms of expectations with your stakeholders?

Julien Millau (12:58)

All want to use AI because currently we don't say try AI, we say we should use AI and we must use it. Now we need to use it, not just to try, we need to use it in a way it will produce value for us in the it works and at the end produce value for our business. We need to use AI to be more efficient, we need to use AI to produce code that is in better quality at the end. So we need to be careful that when you use AI, we need to use AI for something and something that has a value at the end. So that's the main thing that we speak of when we speak about AI to stakeholders. The main thing that we have is that, okay, I can use AI, but for what and what will be the gain? And this is what we talk about every person that act on the SDLC. It's about the gain and the value of producer.

Fred Miskawi (14:04)

Kevin, what about you? How has your relationship with stakeholders evolved and have their expectations also evolved?

Kevin Beaugrand (14:14)

Yes, I will talk about the developer experience in using GitHub Copilot, for example, because I lead the training path for the developers that are working for Michelin, not at least, but for many customers. And I saw a change into their mind. We had many senior developers that had the ability to reject using AI into their jobs because they didn't want to use that for example but they also experienced some mistakes in the AI output two years ago yes the model was well not as performant as good as now, sorry. But if we see now, we see that these developers have the ability now to accept that AI can make mistakes and accept that they have to review what is going wrong into the AI when they are using that. It is not obvious to say that because we know that we should change our habits in order to use AI and to experience some mistakes, in order to see and to “smell” when AI can go wrong.

That's a change I saw around these senior developers and now the senior developers are really sponsored about using AI into their development phases. And it is very interesting to see that. 

Overcoming Resistance to AI 

Fred Miskawi (16:03)

And Kevin, we do see some pushback driven by fear, by concern, sometimes even pride. How do you communicate with people that have a concern about the technology, what this might mean to them or even to the industry moving forward?

Kevin Beaugrand (16:22)

Yes, we designed a specific workshop for them, for all people in fact, for senior developers, regional developers and people that do not like using AI, in fact. In this workshop, we designed a specific application that is close to the daily applications they are managing in order to see how they can use AI. We teach them about the limitations of AI, the ethical limitations, specific limitations, performance limitations about using AI. And we teach them in order to be more efficient in using AI where it is relevant for them. And also to be able to see where they didn't have to use AI because it is not relevant for this task, for example, because it is easy to do it yourself. It is not relevant. So  we teach them during this workshop in order to see the limitations of AI, to experiment the limitations of AI, and to still continue to have their own human eye into specifying what AI should do and what AI should not do during their project.

Fred Miskawi (17:47)

So have a conscious decision on what to automate and what not to automate. And Julian, you've seen and heard Kevin, you've seen him in action, seen our teams in action, seen others in action. What's your take on that? How do you deal with when you have that kind of pushback or individuals that still are either resenting or fearful of the technology?

Julien Millau (18:13)

What we are doing for now is that we are deploying multiple things. We are developing what we call flying teams. We have some teams that are composed of many experts into the organization that are here to help all the teams to deploy here. So we work closely together to review what they do for their product, and how we can help them to use and to push AI into their projects. 

And we also have put in place what we call pit stops. It's every two weeks, teams that want to ask questions around AI or ask how we can do something with AI go to a team of experts and these experts are here to orient them in one way to implement that.

And we have also an enabling team, a team that where experts come into a specific project for one day, two days, and work together to define their objective on how they can put AI into the development of their projects. So we have multiple ways to address that, depending on the maturity of the team, the maturity of the people in the team. But in fact, when we convince a big part of the team, every person that has a little fear or little misunderstanding about how they can use AI, they're okay at the end and they come and use it.

So the main thing is that we need to explain. We need to, not convince, but we need to explain the value at the end and how it can work. Because when we speak about AI, we can speak about a lot of things. And we need to recenter that to the day-to-day job of the person.

Measuring Value & Impact 

Fred (20:21)

Yeah. And Julien, I love your approach. So speaking of value and the importance of measuring value, Julien, how do you measure and demonstrate that the value is produced? How do you measure and demonstrate that?

Julien Millau (20:38)

Yeah, so the important thing with AI is that AI evolves at a speed we haven't experienced with previous technology before. It's fundamentally changed the way we measure the success of that. We can no longer define a fixed objective at the end, say, in one way we want this outcome. It's not possible because the reference point moves around months and years. It’s a little bit different than what we can do before. We have a start but we don't have the end of that. We just know that we need to increase the way we adopt tools, we increase the way we develop things. So for that we measure, we adopt accelerate on our development. So we use the accelerate metric and we use also the DORA metric to measure how software is built. 

And what is important for us now is to measure the lead time. The lead time we use to push something into production. And we go so deeper into the lead time to understand the SQL time. How long will be the development? How long will be the review of your development? And how long a code merge has been pushed into production. And with that, with that metric, we check the trend and we see and we try to see how AI can improve that and how we can use AI to improve that. Our goal is to develop features faster for our business with a better quality. And for that, we put metrics and we ask a team and some technical metrics, we asked him to review them and to see the trends. And for that, this is how we can measure the AI for projects. And after that, we measure also the AI usage. So all the tools that can act on the SDLC, GitHub Copilot, LibreChat, all the tools that can act on that, all MCP that we deploy, we try to have a metric about the usage and we follow the usage in Michelin. But it's not a simple project like deploying something. We don't have the end date, in fact. So it changes a lot the interaction between the stakeholder and the world that we deploy that because many projects say, OK, what will be your goal with AI?

In fact, we want to use more AI to be more productive. But we want to increase productivity by percentage, but it's not a fixed number at the end. The evolution of AI into the last month or year is quite impressive speed. So we need to follow that. We don't have the end date, but we know that we need to improve our productivity.

Fred Miskawi (24:02)

So what you're saying, Julien, is you've built a culture of transparency, measurement, and reporting from the ground up. So as you're introducing the tools, the technology, you already have transparency visibility and you have the ability to report it. For those of you that may not know, DORA, or DevOps Research and Assessment, is from a team by Google. We had an opportunity to meet with a team in New York and it has become pretty central to a lot of what we do in terms of measurement. And then we add to that.

And kevin, what else would you add to that topic as it relates to measurement, transparency, and not just what you're doing with Julien, but also like across the accounts that we have.

Kevin Beaugrand (24:44)

Yes, I would say the main difficulties about measuring the performance or the impact of AI, because the impact can be positive or negative. We want to correct all the impact of AI around the SDLC. It's mainly to be able to collect the correct metrics around AI. If you are using, for example, GitHub Copilot, you can have some metrics that are already in place in GitHub Copilot, like the number of requests, the number of code acceptances we made per developer for example or per team for example, we can correlate that mainly, yes, with the DORA matrix we know, the MTTR, failure rate, code churn and so on into the project. But in fact, the main difficulties we will have in the future weeks, months or years is that AI will not be only in GitHub Copilot. AI will be also in your DevOps platform, for example, because it can impact the SDLC directly by automating some tasks into the SDLC that are running into the CI-CD, for example. And you will have to be able to collect those metrics in order to see the impact directly of each agent that can be human and can be also AI in the SDLC phases. Yet for now I didn't see one way to do that, one way to collect and correlate all the impact of all kinds of AI that are working into the SDLC and I think it will be a must-have for us.

Fred Miskawi (26:42)

Yeah, that brings us to that topic of harness engineering where we're spending a lot of time to kind of surround not just the LLMs, but the agents, these ecosystems to have the layering necessary to get the right level of output. Julien, do you see that as well with the infrastructure and that well regimented ground truth you've been setting for your organization?

Julien Millau (27:10)

Yes, it's changing a lot. We cannot say that it will end at that specific date. It's changed a lot. AI can be better and better every month. It can do a lot of things. If we come back just six months before, we start with a shared prompt library that we share into the organization with an in-house source approach.

After that, six months after, we have the skills that come. So now we are working on skills. We build also agents. So every six months, new things are out. What we know today is that we need to expose our application capabilities outside of UI or interfaces.

And we need to, for example, MCP, promote MCP or promote agent-to-agent to ensure that we can comply with the evolution of AI. So today we have APIs, we have many things. We think that MCP today is a good standard and quite robust and not a big change on what we have today with API. And it's a good standard to deploy more in Michelin. So we ask product team to deploy your MCP server on all your IT for IT products for now to ensure that we can interact with them and we can create agents easily to consolidate the data. And if the product, AI products change or evolve, we are still be able to expose our data and after that we can make many things accessible.

Upskilling at Scale 

Fred Miskawi (29:06)

And I know you're also providing time and space for people to be able to experiment and try things out, learn. And it sounds like as you're learning, you're putting that into practice and then you're scaling.

How do you make sure that your engineers stay up to date on what's happening in the industry?

Julien Millau (29:26)

For that, we have dedicated time for that in machine, what we call case end time. So every team in machine on IT have a case end time and they can work to improve their application. So they can work on many things, the things that they want, but the goal is to improve the way they work and improve their application and also be able to ensure they know the new tools or they check something or they make some trainings. 

And for that also we provide a learning path. So for each person or job that can interact with the SDLC, we provide a training on that. Training for AI but also for any tools or any technology that we provide.

And it's a self-service approach. So it means that any developers or software engineers that want to learn something can do that. We encourage them to learn. They can learn by themselves or they can also participate to some training that we develop. 

As an example, for software engineers, we develop with CGI, training to help developers to use GitHub Copilot. So we start the first session with usage of GitHub Copilot with the agent mode. And after that, we need to work to improve that, to push MCP and agent on top of that. yeah, we trust a lot of training. We know that training people is very useful for them because all our training has a lot of hands-on. We want that people use AI and not just a slide that explains what they need to do. We want them to try to test and to use AI.

Fred Miskawi (31:25)

Yeah, experimentation is key. We keep pushing that and the need for time to experiment and also providing a safe environment for people to fail. Fail small, fail fast, recover, learn from it and keep moving. Kevin, maybe a slightly different angle to that question, which is how do you stay up to date on what's happening?

Kevin Beaugrand (31:45)

Yes, we have the same kind of training paths that are in place for CGI people that are working also for Michelin and we share our learnings between Michelin and CGI locally. So this is interesting on that part. And we dedicated some training paths, some specific training paths depending on the job that are in place in the SDLC. For example, we have a specific training path for the developers or tech leads or architects that work in Michelin, but we have also dedicated training paths for business analysts, functional analysts, because they can use AI in different ways with different tools and we have to teach them about using correctly or effectively AI during their job, not only prompt engineering, technical and so on. 

We also have some workshops we can do with our customers or with our teams, which are called value stream mapping. It is not something created by CGI, but we are using these kind of workshops in order to see where we can dedicate or develop some specific agents into the daily work for some people in order to increase their usage of AI and their adoption around AI.

By doing that we have the possibility to also work into the adoption, yes I talked about that, but the performance or the confidence on using AI into their jobs.

Productivity Gains & What to Do With Them

Fred Miskawi (33:35)

Thank you, Kevin. And Julien, let's talk about maybe the time, the capacity that we're getting back thanks to this technology. So assuming that you are seeing productivity improvements or at least you're seeing your teams be able to produce more than they used to, what do you do with this additional capacity that you've gained thanks to those tools?

Julien Millau (34:00)

The goal is to produce more, to be able to answer to all the business requirements in fact. We have a lot of things to do, our backlogs are very full today and we cannot answer to all the needs of our business. If the productivity increases with AI, we hope that we'll have more features to create, more value that will be delivered. And this is why we want to use AI and measure if the productivity increases with AI.

 So we need to check the big picture, in fact. We can gain time to write code, but at the end, the review, the challenging, the fix can be longer. So maybe we can gain some time on specific tasks, but we will lose some time on others. And we need to see globally from the business requirements to the production and also to the support of the application, how we can gain  time and productivity on that. And we hope also not only about productivity, but also about quality. We can have better products, we can have some review by AI, we have developed some agents that can perform code review automatically. We have developed some agents that can review also the issue or ethics created by business.

Fred Miskawi (35:31)

And are you seeing productivity improvements across the SDLC, not just development at this point?

Julien Millau (35:54)

Yeah, the goal is to have productivity globally, not only for development.

Fred Miskawi (35:42)

Yeah, we're seeing the focus of our teams shifting where people, humans are spending time, effort and energy. It's been an interesting journey for us. And Kevin, what else would you do to make sure and ensure that the value that we produce is aligned with what our clients are expecting.

Kevin Beaugrand (36:01)

Yes, that's very interesting because, yes, you mentioned that AI can produce more code than a human can do in the same time. And once we have the possibility to increase our production and our productivity, we will produce more features into the same application in the same time. And we will have to review all the code. This is difficult for us because reviewing the code can take time, and it is really difficult to review the code that is produced by someone else. And it is also really difficult to review code that is produced by AI, because you don’t have the possibility to challenge the AI directly, you don’t have the possibility to see why AI is producing that, why it thought about this kind of implementation and so on. 

Regarding this difficulty, we have the possibility, we have the possibility to use also AI to help us to review the code. 

But the second question is about how we can increase the process in order to be sure that before getting to the review, we have the kind of assurance that AI will follow all the steps, all the constraints we had and we used to use by the humans and the developers because some of our constraints, some of our guidance are mostly in our minds and that's why it is also difficult to teach our developers into giving the AI the possibility to know what we used to have as a guidance. What are our frameworks? What are our quality assessments we used to do? What are our expectations about the security? What are our expectations about the features? How they should be developed and implemented and so on before getting to the review. In that way, it takes time to what we call to coach AI or coach agents into our project to build effectively the software. It takes time because we have to experiment with how the agent is reasoning and to see that it is more about not seeing directly the output of the AI but seeing which steps are followed by the AI when we want to develop something. And by seeing that, we can see if the reasoning correct or not, not only the output because the output will be very huge, but seeing the reasoning can help us to design how the agent should work and how we can teach it to develop correctly before getting the output. 

Looking Ahead & Closing Thoughts

Fred Miskawi (39:11)

Thank you, Kevin. It's a whole engineering discipline that you're articulating and communicating and it keeps evolving, which has been amazing for me. So as we wrap up, one last question, maybe two, we'll see. But Julien, what is something that you're looking forward to in the next 12 months? New technology, new techniques, new evolution. 

Julien Millau (39:36)

What I'm looking for today is the way we can avoid building web UI. Currently we build a web UI for every product. You use Amazon, use michelin.com, you need a client to learn how we need to use the website. I think in the next coming months and years with technology like Agent to UI provided by Google or MCP app that come into the standard. I think the only UI that the client will have will be prompting on a chat like ChatGPT interface or the voice. And after that, we need to expose our product capabilities and the client will choose and rebuild the interface they want depending on their needs.

And I really think that this will be the end in the coming years that we will not have many UI. We will develop our own UI depending on our needs on top of all the tools we use.

Fred Miskawi (40:50)

An evolving interface between the human and the machine that will be particularly personalized and customized for our individual needs. I look forward to that. It reminds me of the Minority Report movie. It was actually very interesting insight into the future of user interfaces. Kevin, maybe we'll end there with you on answering to that question. What do you look forward to in the next 12 months? What's coming our way?

Kevin Beagurand (41:12)

What I'm looking for is mainly about getting outside the experiment because we are still in the hype phase of using AI into the SDLC. I want to get out of that and to be sure that we still are some engineers that are working in the SDLC and we are building software like an engineer do. And that is very difficult because we, like you mentioned Julien, we have every week some news around AI. Since it is evolving every week it is difficult to build our foundations on the SDLC and it is difficult to see where we can be in production phase. And I hope and I want to be in part of that, be able to use as a production phase every tool we have in place and every AI that can be brought into the SDLC to provide more value in an effective way.

Fred Miskawi (42:36) 

And what I'll say is this for the next 12 months, I anticipate we're going to continue to see context curation, intent framing at the center of what we say, what we do and how we evolve our processes. And context thinking will end up being a soft skill that we all seek in the humans that we do employ in our teams. So, where does that leave us? 

There've been several moments from this conversation that I think are going to stay with me. The first, when you look at Julian's perspective and observation, there's no target anymore. We don't have a consistent target. can set a target evolving every month. So we have to experiment. We have to evolve with it. For a generation of leaders that have been raised on project plans and quarterly milestones, I don't think that's a tweak. I think that's a rewiring of how we approach leadership and how we lead. 

The second is you, Kevin, and the point about using AI in the SDLC, the fact that it's not a tooling upgrade, it's a process transformation. The technology that I think reminds us that the technology itself is often already in place and the struggles that we're seeing, some of the friction points that we're seeing is more around a human transformation, the rituals, the handoffs, the definitions of done and that human culture, that human experience in what we do in the value that we deliver, that's what takes the longest to change. I think the technology has moved sometimes beyond our ability to absorb it quickly. 

I think it's at this core of the story, which is this human transformation, the ability for all of us to move towards a slightly different path of producing value and adapting to these new tools that are coming our way. The organizations that win, they won't be the ones that pick the right tool necessarily. They'll be the ones that build a culture where experimentation is cheap, where mistakes are expected and learning compounds faster than the technology changes. We're still a ways away from that, but I think that's what we're all building towards. So Julien, Kevin, thank you for your candor, your expertise and thank you for sharing what it actually looks like on the ground. Much, much appreciated. 

And to our listener, if this conversation sparked something, share it with a leader in your network, please, who's trying to potentially move from AI ambition to AI ROI. Subscribe wherever you get your podcasts and join us next time from AI to ROI. We're having a lot of fun every time we make these episodes. So thank you so much. And until then, experiment, test, and remember the target will move. The only question is whether you're moving with it. Thank you so much, everyone.