Alright. Let's make a start. So thank you very much for joining.So today's session, I'm going to be here to talk you through about how you're going to close the gap between AI experimentation and also having that real business outcomes that everyone has been looking for. The majority of the organization have already run lots of AI proof of concepts. But the question is, why does so few of them manage and made it through to production or roll out to the rest of the of the organization to deliver that sustainable value? So a little bit introduction about myself. I have been working as a data scientist and recently turned into a consultant. I have conducted, advisory and also technical delivery projects, on key government infrastructures, city planning programs, portfolio risk, and also public health.

I'm currently working in Solke, which is a Swiss consulting organization that brings together IT strategy, digital solutions, AI data platforms, and also cloud engineering implementations. So what I'm going to really talk about is to draw on the patterns that I have come across, during my project delivery and also, client engagement every day. And over the years, we have seen AI generating a lot of impact, but what does it actually take to, close that gap and also translate it translate that into a long term business performance? K. And in fact, many of the organization have no problem launching any AI POCs. And the enthusiasm is real. But then the vast majority of these initiative, they just never actually get rolled out. So let me start by sharing some of my observation why it fails. And the failure and also the disappointment actually came from both sides.

So for the technical team, it was really demotivating to see the POC just stays as a POC and that the code never actually get used. And for the senior executives and it was also frustrating for them to see the resources being invested and without actually having any real impact. So as you can see here, this is this really isn't just my own observation. The IDC, that's the, that stands for the International Data Corporation, they highlights the exact same issues in their reports. And that goes the same for Gartner and the same with the Capgemini data executive summary. And that the struggle to roll out POC is very common. And as you can see here, nine out of 10 project did not succeed. And this struggle shows up in so many sectors.

As you can see in this chart, if you have experienced similar, a problem or issues in the past, you are really not alone because there is a, big disconnection between the POCs and the reality. And the disconnection isn't just technical. It happens across the organization level where any of these five foundation layers, either they don't exist or they don't work together, which is about having this clear framework and strategy with the with the right way to select your use cases, having the right data and AI platform plus the cloud infrastructure, and be able to embed your AI solution across your operational processes, and then putting that solid AI and also data governance in place.

So organization, they tend to under invest into these foundational, capabilities or really underestimating the connectivity across all these five pillars. This is because the responsibilities, they tend to spread across many departments. And when these foundations, they aren't there, and this is exactly why POC does not get the support that they need to scale. And this slide really just sums up what I said earlier. So most of teams start by building the models. They have very good model performance, and this is often their first big win. Their first step to get out the door of their department to the wider organization. But then a lot of AI projects still failed, not because of the performance or the accuracy, because of the adoption or the scaling and some of these integration issues.

The team start celebrating the accuracy without asking if the data foundation, the data governance, or the integration pathways are put in place, which is essential to scale up. And what the team needs to do is also be able to define that business value of your prototype quite early, showing that tangible outcomes through the technical solution. And this is how the teams will be able to move away from a one off experiment using AI or just as an optional capability. And that brings us to the first key takeaway, the use case to value funnel. The idea itself is fairly straightforward. And this is because not every AI idea is worth pursuing. So you need to be very disciplined about how you're prioritizing the AI opportunities and focus your resource resource onto developing them and avoid spreading your investment across so many interesting or plausible ideas in that sense.

So what we need to do is start by anchoring the AI initiative to define your business objectives, whether it can be, for example, reducing your operational cost or improving your customer experience, generate revenue, or just in general, organization digitization. Then we can start following down through to your use cases based on the outcomes that you are targeting onto, as well as the technical and operational feasibility, and also the strategic fit to your organization. You might start looking into your use cases and start by asking, does it drive the outcome that you're looking for? And then run that technical feasibility and risk check. Are we ready? Can we integrate it into your operations? Do they have any regulatory issues? And, obviously, last but not the least, you prioritize your use cases with the strongest value and also having that most realistic execution pathway. And this kind of, funnel does fall through the tough prioritization decision right at the start and making the link between the AI development efforts and also the business outcome really clear right at the beginning.

We also need to align this business value very early because what we have seen a lot is that the team treat the value estimation and the alignment as an afterthought, something that they tend to do once the model has already been built, but then by then it's already quite late.

But then in reality, you may not have all the information, at your hand to make a start. So let's find out what we can do. And that's absolutely okay if you don't have all the information at the beginning of that funnel use case process because this quantification does not need perfect precision. So what we need is to create that shared understanding of that impact with your decision makers. And without having at least a rough estimate, you cannot prioritize initiatives or having gaining that executive support or even just to justify your scaling and rollout. And to be able to reduce the unknowns of this estimation, we can ask what really matters and frame that value across multiple dimensions and understand where each of your decision sits within your stakeholder ecosystem. What does each of your stakeholder cares about, and what kind of decision are they trying to make here?

And here, I've got a couple of example questions that can help with the business value framing. And the key message I want to bring out here is that you don't need a precision or perfect numbers at this say at this stage because what really matters is that the business team, not just the technical team, is involved from the day one and also owning that value story of your AI POCs.

And you you just need an estimation, and that is very clearly focused on your business outcome so you can build that, shared impact or shared view with your stakeholders. And moving on, I want to talk about how we can operationalize AI into the business performance. And shifting from this experimentation, in your lab or maybe in your, data science team to performance really means that we need to change our way to design and also measure and also maintain the AI solution away from that POC phase. And there are differences between prototype that works and also a system that delivers a sustainable value. People tend to mix these up and assuming whatever that makes a prototype successful will by default drive that business outcomes. And on the left hand side, you can see we have the prototype world. This is where most of the organization and the team started. They build a model.

They start cleaning on the data. The model hits a great accuracy and also performance, and everyone's quite happy with that. But here's the reality. The demonstrated accuracy is achieved in a very controlled environment, with limited variables and also having no real users. It is a POC, a proof of concept and not a proof of value. And the real test is on the right hand side, which is on business performance. And that's where AI actually earns its place into that operational process within your organization. And having that performance means a model and also a solution being adopted by users having the measurable return of investment. And it's not just about the model accuracy itself. It means that the model and also the system is resilient. It can handle messy and real world data, edge cases, model drift all of the time.

It also means having the governance in place with the clear ownership risk control and also the audit trail. And also that your AI product you're creating isn't just a stand alone piece. It is embedded in your core business processes. And an error in the middle, so that's the gap. And this is where many of the AI initiative, they tend to slow down or failed. And and also the fact that many of the projects do not make pass through this pilot because they fall into this gap. And this is because the, the POC solutions that is being produced have no clear ownership, and they also don't gain the operational support that they would need and also no plan to scale from both technical and also business operational perspective. So the question for the, leadership team is not about whether we can build this POCs or that technical solution because majority of the organization, they are able to do it.

The real question is whether we can sustain the value, the performance of the solution itself, and that requires a fundamentally different set of capabilities, which I'm gonna talk about in a moment. And in order to turn this working model into the operationalized business capability, we need to focus on integration as Ivan keeps saying. It it needs to be a solution that support decision makers. And rather than replacing the, accountability because that will help to drive the organization adoption and also the trust of the user itself. And as for the solution value, we are not just looking at the model accuracy. Because, therefore, organization, they track things like user adoption, decision latency, and also lots of business KPIs and your realized value coming from the product and the solution itself.

And these metrics can be multidimensional, and it also can be evolved over the time. And all of these needs to be backed by a having a very strong technical foundation that enables productionization and roll out. So from the start, the team needs to design a solution for production and reaches to think about data pipelines and also how everything fits into the existing workflow, operational processes using the MLOps infrastructure. And, also, think about the responsible AI practices that allowed continuous improvement and while staying in control. And when all these comes together, AI just becomes a truly operating capability instead of just a bunch of, disconnected pilots as well how we called it. And finally, I want to touch on governance. So what we see governance is a enabler to build trustworthy and scalable AI.

But, in other cases, people also see as a blocker or a gatekeeper for innovation. But then in many of these successful AI programs, having that governance structure or framework will allows you to scale your solution in a safe and also sustainable way. And what happened is AI fails when everyone builds but no one owns, and this is where the governments come into the organization. It means having that set of roles, processes, policies, and also the technical controls to ensure your AI system and solution can be executed and carry out in a safe, compliant, and usable way across their life cycle. So it does provide a clarity on the roles and the responsibility and making sure that all the teams across the departments are aligned. And it allows you to apply that agree, and also agree on a principle and also any any kind of, like, risk based control for your use cases.

Mandate your solution and transparency, explainability right from the beginning, and also allowed your organization to evolve from gathering individual models or projects to looking after the entire AI portfolio and to be able to create that shared standards end to end and having that consistency in the in the risk reporting.

So I put together six pillars to execute AI governance in practice. Like I said earlier on, without having that clear ownership. So, the model might only just perform on a technical level, but it never gets adopted or not even be able to maintain that model quality, over the time and defending the risk when it comes. And what we also need is having that framework that classify use case by risk and also business impact so we can apply controls proportionately. So it could be, for example, giving that lightweight controls for low risk use case and then giving them much straight to control when the stakes are high. And then we also put the right guardrails in place so that we talk about both at the organization level and technical level, using MLOps infrastructure and also the responsible, AI practices to making sure that both, ethics and also the technical versioning such as on data model, automated pipeline throughout the, throughout the development life cycle.

And last but not the least, we talk about looking after the AI portfolios across the organization and not just on individual models or project, because that gives you that centralized view on all your use cases and create that shared understanding and reusable components end to end.

And also be able to having that consistency, risk, and also value reporting. It also helps you to focus on prioritizing your investment decisions over the time.