Predict, Optimize, Decide: Enterprise Decision Making For The Simulation Era - AUTONOMOUS summit

All right, we'll be kicking off the session shortly. Or just wait for folks to continue to trickle in and then we will jump in. All right, I think we are ready to go. So good to see everyone again. Welcome to day two of the summit. And I am Amir uki. I'm the managing director at BOI here in our New York office and I lead our applied AI and technology practice. And today I am very excited to chat a little bit about how AI based simulation is going to dramatically change how the way that leaders make decisions. So before we jump in, I just want to chat about this word simulation and kind of what's in a word. So simulation means so many different things in so many different contexts. And some people may think of NEO in the Matrix. I personally think of Truman Burbank in the Truman show, one of my favorite movies. Go see it if you haven't seen it. But today we'll be sticking a lot closer to home and we'll be talking about simulation that improves enterprise decision making. So think of your marketing leaders, your innovation and insight leaders, your distribution leaders, supply chain leaders, and so on. And the simulation is an area where we've seen a really big increase in investment as an in house capability, which is pretty unsurprising as companies continuing to, to kind of vie for competitive advantage with AI in 2025. So a couple of things that we want to come away with today in terms of what we've learned. So ideally we'll cover kind of a broad knowledge of what AI simulation is, what a simulation engine looks like under the hood, and what changes once you have a simulation capability. And then we'll talk about some high level ways to get started and we'll kind of squeeze that into the, the 15, 20 minutes that we have. So we'll start with a very simple definition to kick us off. So simulation is a virtual model of a complex real world system that analyzes, that predicts and that optimizes outcomes before you take action. So simulation is basically allowing decision makers to explore lots of different what if scenarios. And then by doing that, they can leverage data, they can leverage machine learning, they can leverage computational modeling at scale to develop much more effective strategies. And I would say an important caveat before we get into the nitty gritty is that today I'm just covering the enterprise field of simulation in terms of decision making. I won't be covering other fields of simulation like medical simulation or military simulation, which follow very different rules than enterprise decision making, and that I am very much not qualified to cover. In those domains. So now, knowing that, let's talk about how we got here in terms of the evolution of AI in decision making. And AI in decision making has actually followed a pretty clear progression in the pretty recent mainstream adoption of enterprise AI. We've gone from generative AI that is allowing for retrieval, synthesis and generation, to, in a nutshell, help you extract and summarize knowledge, then to agentic AI that is moving beyond retrieval to allow the automation and the improvement of complex workflows. But AI LED simulation is actually a very different paradigm. So rather than automating decisions, it actually lets you pretest, it lets you optimize, and then decide between different paths before you take action. So the fundamental shift here is that we're going from predicting what might happen in a business scenario to actively simulating and shaping what the best possible outcome could be. So simulation as a concept in enterprises actually isn't that new. So traditional simulation models have been around for decades and are generally static and rule based, which means they require predefined conditions, they don't tend to have the ability to adapt to real world changes, and they usually can't react to multiple scenarios at scale. AI led simulation, on the other hand, is probabilistic and much more dynamic, which means it's learning from data, it's exploring different high dimensional decision spaces, and it can simulate lots of different markets, competitive and consumer behaviors all at the same time. And that is something that we couldn't really do before and we definitely couldn't do at scale. So let's look at supply chains as a comparative example between the two. So traditional simulation would be basically testing supply chain performance under fixed conditions. So an example would be warehouse inventory levels based on preset delivery schedules. Whereas AI led simulation would simulate supply chain disruptions with lots of different dynamic variables. So fluctuating demand or weather or geopolitical events. So two different approaches to the same use case there. So to put it simply, with agentic AI, we're broadly establishing what is happening and what action we should take. And in AI simulation, we're establishing what will happen and therefore what action we should take. There are lots of different value drivers behind why a client comes to us, usually to build a simulation capability. And some of the most common questions that we're trying to solve are whether they should acquire or merge with another company, what price they should set for maximum profitability or maximum market share, what their national or their global inventory strategy should be. They may want to answer which sustainability initiatives will have the biggest impact on them, while still maintaining profitability. But I would say that by far the most common question that this is trying to solve in the last six months is how do we make our new products or our new market launches successful? So let's run through an example from the lens of an imaginary, an imaginary CPG pet food company that we're just going to call Nibbles Inc. So they're about to launch a brand new high protein functional pet food product. But they have some really key decisions that they need to make before launch. So they need to establish pricing, they need to figure out marketing, they need to figure out positioning. But they're nervous because when they've done this previously, they've relied on their existing data alone. And what they found is that three months down the line, maybe market conditions have completely changed and all the hypotheses they held previously are no longer relevant. So this kind of pretty common situation is actually a really great opportunity to use AI powered simulation to test and identify the right strategy on the pricing, on the marketing and on the positioning before launch. So step one for Nibbles Inc. Is for them to define the simulation goals. So the most common decision areas that companies use simulation to figure out are the potential demand for the product, how pricing will impact consumer adoption and profitability, and which marketing strategies will maximize roi. And then of course, what are the competitive dynamics that could influence success. So for this company, that could be establishing what price point is going to balance adoption and profitability, it could be establishing what channels like social media or vet partnerships or even in store promotion could yield the best roi. It could be figuring out and establishing what the market is for functional pet nutrition or how their competitors will react to the launch. But once they've defined those goals, they need to start thinking about their data inputs, which Tiger mentioned in depth in the talk before this. So data inputs are basically the raw information that is gathered from either internal or external sources that is feeding into the simulation system. So those data inputs are providing what we call the foundational facts, the historical records that are really critical to model this simulation system. And basically they let us establish baseline assumptions for the simulation engine. So put simply, data inputs are really just building blocks that are transformed and interpreted by the simulation engine. They are processed and they become the system's baseline level of context, of contextual knowledge. So as with all data in AI, that data does have to go through a pre processing, a pre processing phase. And without getting, I'd say too bogged down in the details, what this broadly means is that we need to Clean the data, normalize the data, address things like missing or inconsistent data. It means that we will create variables that capture consumer behavior. So think about things like price elasticity or purchase frequency. And it will include things like live integration for really time sensitive data. So time sensitive data could be social listening trends, for example. There's a lot that goes into this phase. I've kind of top lined it. It's pretty important. But those are some of the main points to touch on. So what would this look like for Nibbles Inc. Team for their internal data? It could be CRM insights on past pet food sales. It could be historical ad performances, ad performance from previous launches. It could be loyalty program data on repeat purchases. On the external side of things, it could be competitive pricing trends from E. Comm or retail. It could be social listening, as we said, or it could be macroeconomic indicators as well. Step three, once we have figured out what our data inputs are going to be, is to bring in what we call simulation parameters. And simulation parameters are basically specifying the conditions for this company to explore their what if scenarios. So, so there are lots of different types of simulation parameters. But for this example for our pet food company, the Nibbles Inc. We are going to focus on these four, which are consumer segments which are usually represented as individual consumer agents in more of an agent based simulation. And each of these consumer segments will have their own unique behaviors and preferences. The second will be product variables and we treat these as modifiable inputs. And this can be things like pricing tiers or packaging formats. The third are marketing variables that we define as dynamic levers. So think of things like adjusting ad spend or testing different influencer strategies. Or they could be environmental factors that are integrated into the model. So think of things like competitor actions or even broader variables like seasonality. So four different simulation parameters there. The importance of those simulation parameters, put simply, is that they act as modifiable inputs or levers. So modifier inputs are defining the conditions of the simulation and levers are what? They're kind of the dynamic inputs that business users can adjust to test the simulation's outcomes. So one is defining the conditions and one is used to test the simulation's outcomes. So what do those simulation parameters look like for our friends over at Nibbles Inc. So consumer segments could be budget conscious buyers versus premium pet food shoppers. The product variables could be a small bag versus a bulk bag or it could be grain free versus traditional marketing strategies. Could be a TikTok influencer campaign, or maybe it's a Retail promotion campaign, and then for the competitive landscape, where we could be looking at how competitors respond with promotions or with new products. So step four is almost the final step. And this is really the powerhouse of a simulation capability. This is a simulation engine, and it's where different models are testing different combinations of these parameters that we just talked about so that they can predict lots of different outcomes. This step can really vary a lot depending on what we're trying to achieve for different businesses, different sectors, different clients. But some examples for our Nibbles Inc. Case study could be that we may use a demand prediction model. A demand prediction model is going to forecast demand based on factors like pricing, marketing, and competitor activity. On the technical side of things, it could work by using machine learning techniques like gradient boosting, which is basically analyzing historical and real time data and predicting future demand curves. A different model that we could use here could be scenario optimization, and that is identifying the best combination of strategies that are going to maximize the ROI for the launch. Again, on the technical side, this is usually using different things like reinforcement learning to run thousands of different simulations at the same time and evaluate trade offs. So a trade off could be balancing higher promotional spend with reduced profit margins. And then just a third example of a model here would be a competitive response simulation. So that can incorporate competitive data and just simulate those market dynamics. They're often used, they're often kind of underpinned by game theory principles to try and predict those likely responses. So lots of different ways to use these models. But a hypothetical outcome of these simulations could be that we have futures that are indicating that $15 is the optimal price for the new Nibbles dog food product. They could indicate that TikTok ads can drive initial buzz, or they could say that, hey, there's a big threat from a major competitor and you need to consider a backup loyalty discount strategy. So with that knowledge, let's say that the team over at Nibbles, the marketing team, is going to allocate 40% of their budget to TikTok and maybe 60% to In Store Prom. Based on those simulation insights, the pricing team may set the launch price at $15 with a limited time subscription discount. And after human review of these outcomes, Nibbles Inc. Is ready to go to market with their new pet food. So I mentioned a couple of different simulation techniques in that example just now. There are really a lot of different ways to approach simulation, and those different approaches vary in complexity. So we just talked about a couple just now. Some of the models that we touched on are Things like agent based modeling that is simulating individual consumer behavior. Another is what we call system dynamics modeling, which is simulating much more long term scenarios. So things like brand adoption, pricing impact and marketing effectiveness. And a third is scenario optimization to run lots of different what if simulations that determine the best combination of things like price and promotions. So you'll see on this diagram, actually that last one for scenario optimization is a couple of different approaches. One with LLMs, one with more complex ML techniques. For our example, we would use one with more complex ML techniques, things like Bayesian optimization and reinforcement learning. There's definitely ways that you can create really effective scenario optimization models using just LLMs. So based on our experience with companies that have built and are leveraging the simulation capability, we do predict that in the next two to three years, AI led simulation won't just be a competitive advantage, it will be table stakes. So as these simulations become even more sophisticated, they'll replace or at the very minimum, they'll live alongside static forecasting models, they'll reshape strategic planning and they'll change how leaders make high stakes decisions. So ultimately they will become the default mode of enterprise decision making. And what we'll see is that we'll move from decision making based on simulated futures, or we'll move to decision making, sorry, based on simulated futures. So instead of relying on static predictions, we'll rely on AI driven simulation. Instead of pattern recognition, we'll move to shaping a future based on optimizing variables in advance. We'll move from asking what's next to asking what our best move is. And we'll move from past driven to future driven thinking to explore previously unknown possibilities. And what we'll see is that the role of leadership within organizations is going to shift quite a bit. So simulation capabilities are going to let leaders become just a lot more effective and a lot speedier at anticipating and mitigating risks. But it's also going to make them much more effective at identifying opportunities as well. So some important considerations just before we wrap up. Despite its power, AI led simulation does have many of the same challenges as other fields of AI and computational modeling. So think data quality, scalability, computational overhead, human trust in AI driven recommendations. We want to think about the impact of bias reinforcement as well. And there's lots of different proactive bias mitigation techniques that we use for that. And we want to think about things like decision paralysis. How are we implementing really clear decision criteria that are highlighting the most probable or the most impactful scenarios? And then of course, black swan Events, which we always talk about when we are considering anything AI related. They're hard to model. But there are some effective mitigation techniques, like using weak signal data sources that wouldn't traditionally be considered. So I'm going to wrap it here just with how to get started. If you are thinking about building an AI simulation capability, first we would ask what critical decisions does your business need to make? How could they benefit? Then ask what data sources you already have or need. Ask who in your organization would use these insights and how would they act on them? And then we always say start small, pick a single high impact use case, assess your data availability and then experiment. Run a prototype, run a proof of concept and then think about scaling. You really, really do not need a full scale system on day one. We always start small, learn and refine as you go. So try it, test it and scale it. So let's revisit our original definition. At the beginning of this talk we said that simulation is a process of creating a virtual model of complex real world system of a complex real world system to analyze, predict and optimize. But really we can simplify that to say stimulation is just a much smarter way of, of doing what you're already doing, but with much more confidence. All right, thanks everyone. Awesome.

Thank you so much, Amir for this great talk. A lot of questions in the chat, but unfortunately we're not going to have time to get through them. And so we have about 10 seconds left until we are booted to the next session. So if you do have any questions, feel free to reach out to Amir directly and he should be able to come of answer those questions for you. So thank you so much and see you in the next session.