Game Day at Global Scale: How PepsiCo Foods Uses AI to Feed the World Cup

Summary

PepsiCo runs 125 million demand forecast combinations weekly on the Databricks Data and AI platform, achieving over 80% DSD accuracy while cutting compute costs by 60% after migrating from an on-premises solution.
This video demonstrates Pep Planner, an internal application built with Databricks Apps and Lakebase that lets supply planners model real-time what-if scenarios and investigate demand spikes — such as volume surges during FIFA World Cup events — with root cause analysis and in-app adjustments.
PepsiCo's architecture, built on Agent Bricks, Genie, and Unity Catalog, is enabling agentic AI to begin automating anomalous demand detection across Sales, Marketing, Finance, and Supply Chain at global scale.

Game Day at Global Scale: How PepsiCo Foods Uses AI to Feed the World Cup

FIFA World Cup comes to the Bay Area this June. 70,000+ fans at Levi’s Stadium will consume millions of servings of Frito-Lay products, and millions more will be watching across North America. PepsiCo Foods doesn’t get a second chance to stock the right SKUs, in the right quantities, at the right locations. Get it wrong and you have bare shelves. this video shows how PepsiCo Foods runs 125 million demand forecast combinations weekly on Databricks, achieving 80%+ DSD accuracy while reducing compute cost 60% after migrating from on-prem solution. We’ll demo how supply planners model what-if scenarios in real time using Databricks Apps and Lakebase, query demand drivers through Genie, and how agentic AI is beginning to automate anomalous demand detection across Sales, Marketing, Finance and Supply Chain. You’ll learn: How PepsiCo runs driver-based forecasting at 125M combinations/week on Databricks How Lakebase and Apps enable real-time what-if scenario planning
Talk By: Michael Cleavinger, VP Global Data Science and Analytics, PepsiCo ; Venkata Wunnava, Sr. Solutions Architect, Databricks

Chapters

FAQs

How does PepsiCo manage demand forecasting at scale?

PepsiCo runs 125 million demand forecast combinations weekly on the Databricks Data and AI platform, spanning multiple product lines, geographies, and customer hierarchies. After migrating from an on-premises solution, the company achieved over 80% DSD (direct store delivery) accuracy while reducing compute costs by 60%.

What is Pep Planner and how does it help supply planners?

Pep Planner is an internal supply planning application built on Databricks Apps and Lakebase that allows supply planners to model real-time what-if scenarios interactively. Planners can investigate demand spikes, perform root cause analysis, and make adjustments directly within the tool — for example, analyzing a volume surge in Lays chips tied to a major sporting event.

How does PepsiCo use Genie for supply chain decisions?

PepsiCo uses Genie as a conversational interface that lets supply planners and analysts query demand drivers and investigate anomalies using natural language, without requiring deep SQL expertise. The Genie space is backed by Unity Catalog, ensuring governed, reusable data assets across Sales, Marketing, Finance, and Supply Chain teams.

What results did PepsiCo achieve after migrating to Databricks?

By migrating from an on-premises forecasting solution to the Databricks Data and AI platform, PepsiCo cut compute costs by 60% and improved direct store delivery accuracy to over 80%. The platform now also powers agentic AI capabilities that are beginning to automate anomalous demand detection across multiple business functions.

Full transcript

[00:07] So we're excited to be here to talk to you today. As he said, I am Michael Clevenger. I lead our data science and advanced analytics at PepsiCo. And uh we're here to talk to you about signaled to decision. And my partner is >> Hi, I'm Venara Wunava. I'm a senior
[00:19] solutions architect in data bricks field organizations. Been there for four years with data bricks and almost a couple of years with PepsiCo. Nice to meet you all. So, we're going to talk about game day at global scale because you see 40
[00:31] miles south of here in less than 10 hours, Austria versus Jordan is going to be playing in the World Cup game at the Levi Stadium and that time is going to keep on ticking. It's going to get closer and closer. Now, the thing is
[00:44] inside the stadium there going to be people who are eating Lays and Doritos and Tostitos and popcorn and enjoying the game, enjoying everything that's going on at 9:00 p.m. There going to be 70,000 people in that stadium.
[00:56] Outside the stadium, there going to be 30,000 people who are tailgating around the area. Across the Bay Area, we're going to have 500,000 people. In California, there'll be two million people. Across the United
[01:08] States watching the game will be 10 million. And worldwide, it'll be 80 million. But the thing is that it's not just about this one match. It's not just about this one stadium. What we're looking at is it's about everyone
[01:21] watching. Because from PepsiCo's perspective, when you watch a game, you want snacks. When you want to have something that's fun to do, you want to get a drink. You want to get something that makes you smile. And if you're not familiar with all our brands at PepsiCo,
[01:33] just look up here and you'll see some of them. And I'm sure there's some that you realize that you love and that mean a lot to you. And that's what we like to think about because every day we connect to over 1 billion people. That's not
[01:46] every year. That's not every month. You probably heard Mesh talk about in the in the keynote just barely. 1.4 4 million people consume our products every day. And we like to think about that in terms of smiles. And every one of those smiles
[01:59] brings with it a lot of data. And we're going to talk about how we look at that from a forecasting perspective. Now, to help you understand why that's a lot of data that needs to be generated, we're going to take a step back. You see, we
[02:11] have three different audiences that we focus on inside of PepsiCo. The first is our customers. These are the retailers that operate and stock us. Second, we have our shoppers. These are
[02:23] the people when you go to a store and you're standing in front of an aisle and you're picking something to buy, you're a shopper. And then finally, we have our consumers. These are the people who eat and drink our products. So from our perspective, we look at over half a
[02:35] million stores inside the United States carrying our PepsiCo products. We look at 10,000 plus routes that are direct store delivery routes that are operating every day. Millions of people who are going in and buying our products. And we
[02:48] have to do this at scale. Because it's not just about these audience, it's it's also how you shop. So, if you think about it, you're shopping from different ways. When you go to grocery store, you're looking at a stockup trip or you're looking to do your weekly family
[03:01] purchasing. When you go to sea store, you're looking at impulse purchase, something you want to get right away. When you look at a club, you're looking at big bulk buys, multiacks, that party prep for the World Cup. And then when you look at e-commerce, you're looking at scheduled baskets or search driven or
[03:15] signals that are coming from the digital side. All of these combined give you a different unique perspective into what is happening. And we have to create a forecast to be able to get the products to all these places at the right time for those
[03:28] different needs. Because you want Seestore to be there right now today, you want our club to be there in preparation for your party that you're planning out in two weeks. And how to make that all happen is the difficulty that we've had. Because when we go back
[03:40] to it and we think back about that stadium tonight, the demand isn't stopping at the gate. It doesn't even stop when you think about the Bay Area. That demand goes beyond the Bay Area all the way across California and to the
[03:54] United States because game day for us isn't about in 9 hours. It actually started last week when we had to get our products in the stores in the right spots. It was about yesterday. It's about today. And that's what we needed
[04:07] to look at. So I want to kind of build this out now that you understand how we look at our our customers, our shoppers, our consumers. What's the size of this price? So, inside the United States for our snacks, we actually run 1.5 million
[04:20] DFUs. We'll talk about what those are a little bit if you're not familiar with forecasting units. And that results in 125 million different weekly forecast combinations we have to look at. At 104E
[04:32] horizon, thousands of different routes that are going through and delivering every day with hundreds of different retailers and channels. You see, at this size, the old way of working doesn't work anymore. It breaks. We can't do it.
[04:46] And so, we want to take you through how we thought about changing this and what we needed to do because the deliverable itself is changing. Forecast as a deliverable is no longer the goal. It's actually decision as a
[04:58] deliverable. We're going from modelers building a plan, planners interpreting it, and then your action lags to then moving to forecast feeds action with explanation in minutes because the
[05:11] handoffs is where the value erodess. It goes away. And so we have to figure out how do you actually get past that point because accuracy is now table stakes. It's not about getting a model that's accurate. It's about getting to that decision faster and being effective in
[05:24] the way you're doing it. The shift is not bigger models. uh and I know this is tough for data scientists to say hey let's not focus on the bigger models let's not make them all really fancy and deep and it's not about that the shift is what happens around the forecast to
[05:37] make it usable now the way that we did this is we decided that we needed to make a decision and that platform decision was going to be our decision and that's where we partnered with data bricks to be able to come together because we started out with one
[05:49] foundation that foundation is taking the lakehouse and really looking at what we have from our features and our models and our prompts and our AI all brought together tied into Unity catalog with governance. We needed our models to be
[06:02] sitting next to our data, not across the country, not in different clouds. We need to figure out how do we bring that together in an efficient way. On top of that foundation, we have agent bricks and ML flow and we have our apps and our lakebased and then we have Genie to be
[06:14] able to answer questions for people who are not technical but still need to get those answers in the right way. That dialogue with data that's important because for us it was really about how do we have one spot for governance? How
[06:26] do we have models and apps that are next to the data? And then how do we have reusability across markets, not just rebuild per market? And that's really that hidden multiplier that I'll bring up again. Now, as we built this out with our partners inside of data bricks, we
[06:39] still have things that we own on our site. So, we have what we call the PF plus, which is our PepsiCo forecasting engine. That is PepsiCo's IP. That's what we think we do really well. We build out this engine. We have that setting on top of the lakehouse on Unity
[06:53] catalog a IML. We look to deploy it on data bricks and they're a partner for the platform, but we own the engine and we manage that process because we're in the stores. We're taking care of our shoppers, our customers, our consumers. And so we want to bring what we can
[07:06] leverage really well on top of what they're actually bringing from a platform perspective and be able to move a lot faster. By doing this, we're actually supporting this with three different pillars. First, it comes down to you have to have a model. No matter
[07:18] what, we need driver-based modeling. You need to have something that's looking at weather and promo and calendar and all those things that go into a model to make it effective. That's important because that's the base table stakes accuracy that you need. Second, we
[07:30] needed parallelization. We couldn't take too long to be able to run these processes. We needed to figure out how do we actually run at 125 plus million different combinations every week. And we needed this to be done in the planners window. I can't run the model
[07:42] and get it done 5 minutes before it needs to be deployed where the planners don't have a time or a chance to go in to actually figure out if everything is accurate. if they need to adjust or how they give different messaging. So it need to be able to be done quickly in that response. And then the third is
[07:55] that governed reuse. New markets pick up where the other markets had built. They don't just start over. It's not just building and then lifting and shifting something over. It's building from where you have been to be able to get that multiplier effect. Now what we're going
[08:09] to talk about is we want to show you three different views of the same engine. We're going to show you a view from a planner experience. We're going to show you a whatif scenario. And then we're going to show you a dialogue with data or when Genie helping out on some of those pieces. Now, this is the same
[08:21] forecast going into it and it's the decision coming out in minutes that can be able to drive that change. Uh, obviously we've changed the data around a little bit from ours, but this is all based on our real production grade systems that are in in production right
[08:34] now. So, Vata, do you want to take us through? >> Yeah. >> Thanks, Michael. So, a forecast is like the carpet in my living room. On the outset, it looks great, but as you start putting your
[08:46] fingernails in it, you kind of feel all sorts of patches, slime, glue, and stuff like that. So, you have to be careful with these forecasts. The intent behind Pep Planner, that's the application that we've built, is to exactly identify
[08:59] those bad patches, those anomalies, trends, and then help the demand planners and the data data scientists to be able to rectify and adjust those forecasts using agentic AI. So before
[09:12] actually before I actually dig into pep planner the application I want to spend just a couple of minutes talking about what we are forecasting and how we are forecasting um like Michael alluded to before what we are forecasting is what
[09:24] we call a demand forecasting unit or a DFU. A DFU comprises of primarily three parts uh the product that is the product and the skew etc. It cons it also has a
[09:36] component of geography. So which area are we forecasting for? And then it obviously has the customer, the Walmarts, the Publix, the the the Targets and so on. So these are the three main components inside a DFU that
[09:48] make up a DFU. Why is that important? Because there are hierarchies in each of these categories. In in the product category, for example, you can go as high as the the the highle category of the product which can be
[10:00] snacks. And then you can dig deeper into different levels of granularity like you can go right till the the product packaging size for example. And why is that important? Because depending on where you are in the hierarchy the
[10:12] results of the forecast is intended to be used for a different purpose. If you're on a higher level category then it can be used by let's say the FPNA team or the finance team to plan budgets. If it's actually on a lower level uh granularity obviously that's
[10:25] going to be used for demand forecasting and um inventory management and all that. So that that is sort of a critical component. How do you manage hierarchies in a forecast? And then also the from a technical standpoint hierarchies are important because they add more
[10:38] complexity and scale to the overall project, right? Different hierarchies mean different forecast combinations and so different forecast combinations means that we need to scale accordingly. So like Michael said, we have like 1.5
[10:52] million DFUs that are being forecasted on a weekly basis. So the the the cadence of the forecast runs are weekly. And then let me just quickly talk a bit about the window of the forecast as well. The forecast window or what we
[11:04] call the horizon of the forecast. So for each DFU that we forecast, we forecast it for 104 weeks in the future. So that's our forecast horizon. And also uh
[11:16] the actuals are lagging back by four weeks. So the actuals come in after four weeks. So as soon as the actuals come in, we analyze the data, we analyze our forecast to to sort of understand what's the accuracy, what's the bias, so on so
[11:29] forth. And then that's where uh the demand planners and the data scientists sort of uh sort of do a handshake and try to understand what are the anomalies, what happens uh what happens
[11:41] if there is a bad forecast and then they go about fine-tuning the process, fine-tuning the model and then adjusting the forecast. So that's the overall life cycle in a nutshell. So what is the challenge that we're trying to solve here? The interaction that happens
[11:53] between uh the demand planners, the business and the data scientist is today a little reactive and manual in nature. Um the data scientists are are severely outnumbered with the with the with the
[12:07] with the data sort of the the demand planners. So there are almost 6x demand planners compared to the data scientist. So it's hard for that team to scale up. So that's where we are using agentic AI and data bricks to be able to give more
[12:19] proactive insights and proactive recommendations by triaging those an anomalies to help the data science team. So in a nutshell the scale 1.5 million DFUS 140 150 million plus forecasts that
[12:33] are generated weekly. There are certain gaps that we also want to address. For example, let's say this is the FIFA World Cup season. So let's say Pepsi or a retailer um announced a shock promo like a quick promo that was introduced
[12:45] just the past day, past couple of days and so on. So we need to account for that as well uh in order to adjust the forecast. So how can we use AI to to um to to adjust for those shock incidents,
[12:57] shock promos, any supply chain incidents, any mishaps and so on. So so that's another thing that we want to solve. And then finally, like I said before, make the interaction between the demand planners and the data science team a little more fluid. Enable more uh
[13:11] more recommendations by the AI that that that can help this team to surface those anomalies and and those bad forecasts. All right. So like I said, being able to dig into proactive insights, being able
[13:24] to adjust, verify, and then retrain the models and then um incorporate those into their production cycle and then accelerate with trust using agentic. That's the three that's the three key goals that we want to accomplish as part
[13:37] of this this application. So with that said, let me actually pivot to a sort of live demo and uh let me switch my screens. Okay, here I can take the Thank
[13:50] you. All right, so this is the Pep Planner application that we've built for the data science and the demand planning teams inside of PepsiCo IBP. Um, so I'm
[14:04] going to be wearing two ha two hats here as I go about this demo. One is that of a demand planner uh and the other one is going to be a data scientist. as I um so as a demand planner the first thing that I see when I actually look at the pep
[14:17] planner application is a quick summary a quick attention card the attention card actually gives me the critical anomalies and flags flags can be incidents like I said before any shock promo or any any
[14:30] supply chain mishap any social event that might change my future forecast so it's going to flag all those anomalies and and and risks um in my forecast so that I can take action on
[14:42] Um so as a demand planner I can look at those anomalies. I can filter those anomalies incidents and so on. You can see that in this case I have also we have some rule based engines that are also working behind the scenes with the agents to to classify those anomalies
[14:55] like is it a chronic failure? Is it a volume spike? Is it a deviation from the actuals? Is it is it a sustained bias issues? So there are different types of anomalies that we try to surface as part of this. And then you
[15:08] can also see here the DFUs that are affected. Obviously there is a priority associated with the anomaly. But there's also the set of DFUs or the demand forecasting units. The demand forecasting units are important because like I said before they they comprise of
[15:20] the customer or the retailer. They comprise of the the product itself. So we kind of understand which product is getting impacted by the forecast. Which customers are likely to be impacted by this anomaly. And then we can also look
[15:32] at the different categories and some descriptions. And then you can see an action as well. Now, this is an action that has been uh sort of recommended by the the agent that's running behind the scenes. I have a few agents that are
[15:45] running behind the scenes. I'll go into that in a bit. But that that is an an agent that is recommending an action depending on all the feed all the feedback that is given to the agent by PepsiCo. Like PepsiCo like Michael said
[15:57] has a lot of IP internal IP as to how they want to uh set up the forecast, what can be an anomaly, what can't be an anomaly. So the agent uses all that rich PepsiCo information to be able to make that decision and act accordingly. Okay,
[16:11] I'll go there. But the key here is that a demand planner just by looking at it every morning can understand what are the what are the anomalies, what are the risks that need to be escalated to the data science team without fishing for
[16:23] this information in in five different areas and five different screens and and running complex jobs and whatnot. So that's number one. Obviously there are some other KPIs that I also expose here like for example what are the chronic failure views like what are the DFUs
[16:36] that have failed in the last 12 13 weeks. So these are these are chronic failures in in different areas that I'm I'm I'm flagging here. And obviously the the the usual suspects my actuals versus
[16:48] forecast um you see the DNA of the forecast. So PepsiCo has like Michael said promo data, pricing data, um holiday season data, all these occasions that also Mages spoke about this morning. So all that data actually feeds
[17:01] into the forecast and and we can give uh the demand planners a clear sort of a picture of what is the DNA of the forecast. What attributes and features have affected the forecast? If there is a spike, what attributes have affected
[17:13] the spike and so on. So all all those things are giving more context to the demand planner. Obviously this is all integrated with genies and and the supervisor. So agent bicks is talking to uh the data bricks genie instance here.
[17:26] So I can like what are the brands with the lowest accuracy. So the demand planner can easily have a dialogue with the data here in order to understand like what's going on, what's working, what's not working and and be notified and and get more rich context around uh
[17:40] the data. Um you can see here as you can see here it's giving me a quick overview of what are the brands that are not not uh performing well in terms of accuracy of the forecast. Okay.
[17:53] So that's that's what the demand planner does. But let's let's move a step forward. So once the demand planner actually look at looks at this attention card, one of the things that he or she could do is actually look at the event itself. So let's say I want to pick one.
[18:06] I want to pick uh let's say yeah let's let's pick this one a volume spike in like lays dip. So if I click on that it this will take me to the attention card. Uh this is the actions tab. So the
[18:18] actions tabs is listing out a few things. It's listing out agentic insights. So insights that have been surfaced by my agent. It's listing out action skew. So what are all the actions that have been escalated by the demand
[18:30] planners or an agent to the data science team to look at. So this is the proactive insights that I'm talking about here. And then obviously it has some demand sensing incidents. So demand sensing incidents are like I said incidents that might affect the
[18:42] forecast. It might affect the l forecast that's in the lag or in the future. Um so think of any supply chain mishaps and a new promo that that has been introduced and so on. So let's actually look at agents inside. Let me just
[18:54] quickly refresh the screen. All right. So, it it is it it is pointing me to the the the actual chronic failure that is is leading to this issue. And then what as a demand
[19:06] planner I could do here is I can see the actual DFU that has been affected. I can see the type of the DFU again, but also it gives me a status. The status tells me whether it has been worked by the
[19:20] data science team already or if it's something that's pending. And then I have a causal agent that's that's running behind the scenes that's actually giving me a causal reasoning agent that's running behind the scenes that tell tells me whether the anomaly
[19:34] or a risk has occurred because of a weather event or whether it has occurred because of any geopolitical event or whether it has occurred because of data quality. So this is critical and this is uh this this there there's a lot of work
[19:46] that has went behind it because think of it as it is trying to get and match patterns do demand sensing and understand patterns based on PepsiCo's pricing information promo information uh
[19:58] the the supply chain information. So all those incidents are captured and the agent is thinking around all that context and then trying to make a decision as to what might be causing that anomaly based on the time window that the anomaly has been flagged. So
[20:12] this sort of tells you like the the context that Ali was talking about this morning like data bricks is critical here because it gives you that context of mixing different types of uh data sets and then giving you that rich
[20:24] context with unity catalog. Um and then there is an action that al has also been recommended by the agent. So action is so let us look at an action. So these are all retraining complete. So I I'll
[20:36] go into an event where retraining is not done yet. So I I'll probably pick on this weather event. So if I pick on this weather event, you can see that it has highlighted this particular event. It is saying that there is a heat dome in
[20:48] Pacific Northwest because of which there might be a sort of an adjustment of forecast that is required. So what I can do here as a demand planner is okay, I've I've done some context uh sort of investigation and I feel like this is
[21:01] something that might affect my forecast. So I can say okay, I'm going to approve this and I'm going to send it to a retrain. Okay, so I've approved and you can see that I'm getting notified as to what's happening uh with that particular anomaly. So as a data scientist now what
[21:14] I can do is I can go to the uh adjustments tab. So in the adjustments tab the data scientist is going to look at what are the things that have been escalated or flagged by the demand planner and as a data as a data
[21:28] scientist I'm sort of getting this. So I'm getting so as you can see here this one I I think we flagged this one. So this particular event it's saying okay preposition for the winter storm northeast and then it's saying uh like you need to use light GBM. So again the
[21:41] agent is running behind the scenes to gather the right context. The agent also gets all the feedback from the previous retrains. So there is a way that once you're done with the retrain you can give the agent feedback and it also
[21:54] looks at all the ML flow logs in our machine learning cycle and then it gets all that context and makes an informed decision as to how it can actually go about the process of retraining this this particular issue and then as a data
[22:06] scientist you can see that there is a recipe here as well and I can update the recipe to update the weights of the model the hyperparameters the the feature weights and so on and so forth and then I can save this config and run
[22:18] the retrain once I run the I can also see the the number of DFUs that are affected by this retrain here. So you can excuse me you can see that almost 50 DFUs are are affected by this retrain. I'm not making the adjustment yet. I'm
[22:31] just retraining the job. So once I retrain the job it should start. Let me refresh the screen.
[22:46] Yeah. So once I retrain the job, there is a retraining uh job that has been completed here. So you can see all the DFUs that were affected. You can see the ajusted weights and what's what what is the delta of the adjustment. So it gives you a preview of if you run that
[22:59] retraining job and and rerun the model, what is the adjusted rate and what's my accuracy bump after I retrain the the model based on that uh the the variables, the recipe updates that I
[23:11] have made. It also gives me the weight of the recipe and you can also uh it you can also give it feedback. Now this feedback is again very essential because this is captured and then fed back to the agents to to gain more insights as
[23:23] to what's working and what what's not. Now this is where once we get the updated forecast this is where the data science team can start talking with the demand planners to understand if this adjustment works for them doesn't work
[23:36] for them and then uh they can either do two of things. E either they can update just the forecast numbers or they can go back and update the entire model itself. If I update the model then obviously
[23:48] these DFUs will start leveraging that model from that point onwards. And so there are there is there's obviously a healthy discussion that goes on between business the data scientist and uh and the demand planners over there or I can
[24:01] choose to ignore it and I can say okay I'm I'm not going to do it through this. I'm going to just do a manual override of my adjustment and that's absolutely possible as well. Um, so that's sort of a quick overview of what uh PEP planner
[24:14] does. Like you've seen here, it's uh it reduces a lot of friction. And it it makes it very seamless for the demand planner to to understand what are those anomalies to understand what might impact his or her forecasts and then uh
[24:29] and then escalate those to uh to the data scientist where the agentic AI is working behind the scenes to do all the triage recommend action. So it becomes very easier for the data scientists and the demand planners to to operate on
[24:42] that level. So that's a quick demo. Let me um go back to the slide deck.
[24:58] All right. Okay. So just let's talk about the architecture for a few minutes. So on the left you see uh data from different source systems getting ingested. So we use LakeFlow connect to connect to external source systems. We have data in
[25:11] SAP as well uh that that uh the the critical PepsiCo data that's in SAP as well. So we ingest all that pricing promo data and then use LakeFlow connect and lakeflow jobs to process that data
[25:23] and then obviously the machine learning models are running behind the scenes to to uh to to run those models on on generating the forecast. Once the forecasts are generated, they're all stored in Delta inside of Lakehouse. So
[25:35] that's the ingestion and the operation uh sorry ingestion part. On the consumption side, you have uh the users, the demand planners and the data scientists on top there talking to applications. The applications um are
[25:48] where they would go to see the anomalies, the the the trends and then also uh interact with agents. So I have three agents at this point. There there is an anomaly triage agent that looks at
[26:00] the anomalies and uh recommends the next best action. We have a casual uh sorry a causal not a casual a cause causal reasoning agent that's actually running as well to pattern match and give
[26:13] insights as to what might be happening with the forecast that might be affecting it. So basically uh giving an insights to the users and then obviously we also have an incident triage agent. So incident triage agent like I said is
[26:26] looking at any shock events and then uh and sort of recommending actions um and and sort of understanding the patterns there. So those are the three agents but also very importantly like Michael said dialogue with data is critical to this
[26:39] entire application and for that we have a supervisor with agent bricks and then genie running behind the scenes multiple genies running behind the scenes to to answer to all the queries. The key part here like I said is all this operational
[26:52] data is also stored in uh in lakebase. So this is very critical because there are a lot of actions that are going actions by users actions by agents. Uh the users are clicking on escalations
[27:04] the user the demand planners are are are having multiple actions and responsibilities as part of this application. So all those actions the session states the agents uh states all those critical informations are stored
[27:16] in uh lakebase. Also the adjustments to the forecast any adjustments that are made to the forecast are also stored inside of lakebase. And then where necessary we are actually using a sync to to uh to to merge that data with uh
[27:31] lakehouse um with an effective forecast. The couple of key differentiators here that I want to point out before I hand it over back to Michael is that because this is a unified system wherein you
[27:43] don't have to build separate isolated systems because this is a unified system with Unity catalog it's very uh efficient to go from a a failure to fix like from a failure of a forecast
[27:56] understanding why the failure occurred and then creating fix using the agentic AI recommendations it it's actually a very seamless transition from that and we are able to move faster, we are able to um uh also have high level of
[28:11] security. So think about a demand planner. When a demand planner is interacting with the data, he or she has to make sure that they only look at their category and not look at other demand planners category. So all that out of the box security is is built with
[28:24] Unity catalog and data bricks application. So that's a quick overview on the architecture side. Um I'll hand it over to Michael for the rest of it. Thank you. >> Great. Thank you. So with all of that fancy apps, great models, it means
[28:36] nothing at the end of the day if you're not actually doing something for your business. So we want to show you what we've actually been able to do. So through this, we've been able to increase our accuracy above 80%. That's when we're comparing it to an existing
[28:48] system that had been on premise for over 10 years with a lot of knowledge, a lot of buildup, great people who were able to think about problems in new ways. But we brought new technology and this new platform to help us decision on that. And that was important because it's
[29:00] driven millions of dollars in revenue. cut us down 50% of the cost reduction compared to the previous s solution that we had. But more importantly, our time to value increased 2x. What that means is that we were able to actually roll
[29:13] out projects 6 months earlier than planned. We're able to make these differences quicker and we're able to then extend expand and extend into new uh teams and new markets a lot faster. And we also have to make sure that we're looking at bias. It's not just about
[29:25] accuracy. If you're into data science, you know that you need the accuracy. You also need to make sure that you're within bias that you're getting the right amounts to the right places. And so we had the business targets that we're able to meet and and continue to exceed on. So what actually changed
[29:38] through all this? We talked about the the platform, the we talked about the EA structure. What changed? What changed is we went from a legacy view of tooling that we' carried for years, that technical debt that's built up, and we rearchitected. We didn't just do a lift
[29:51] and shift. We actually redid it all from the ground up and did it faster. We moved from forecast as a as a deliverable to decisions as the deliverable. That's why so much focus is on the demand planners. How do you get decision-m to the people who need to
[30:04] make it? Moved away from brill handoffs to govern in one place. Uh moved away from permarket builds and moved to reuse across the markets. And then we had our models get closer. Our models and apps are next to the data. So there's not
[30:16] this this lag and this difference between them. So it's really about rearchitecture, retiring tooling, and picking the right metrics to success, which is getting those insights faster on out there. So what did this look like from a year perspective? Well, in Q1, we
[30:29] started with the legacy and we retired that. We didn't do the lift and shift. We just completely looked for the new solutions we could step into. That's a big ask from PepsiCo when we're talking 125 million different weekly forecasts that we have to change around. That's a
[30:42] lot of data we're putting into it. Then we migrated our models over using that driver-based rebuild that we wanted to do to make it effective and efficient. Then we started doing our first market life test and endtoend decisioning that we're able to do in one geography. We
[30:54] saw success there. And so then we moved into that next step in that quarter 4 which was reusing to secondary market not just rebuilding but how do you take that and you start over because that's where that multiplier starts to kick in and makes it more beneficial for the
[31:07] other work that you're doing to get to that value. And then finally, what we're doing now is we have PF plus at scale uh with the planners. We're doing what if with Genie. You saw the root cause analysis that we're talking about. And it's all based on the same forecast.
[31:19] It's not multiple forecasts. It's not multiple data sets. It's all in the same place. So people get consistent answers. Now the thing that I think is the most impressive from a data science side on this is that our cost for compute went
[31:31] down by 60%. Our accuracy went up above 80%. and our scope increased to 125 million weekly combinations. It's not often when you were able to do all of those things together. And the
[31:44] reason why is because we looked at it from a different solution. We weren't trying to say how do I reduce my cost. We were saying how do I actually get the decisions made in the right place and through that process and through those different columns that we built up were
[31:56] able to realize that and we're still looking for how do we reduce that and how do we make it faster because that compute reduction while actually going with the scope is going to be important as we move forward. So what does this all really mean mean or what did we learn from this? We
[32:10] learned that decision is the product not the forecast. Number two is that reuse is the hidden multiplier. This is where you're going to have that interest that compounds if you're
[32:22] thinking about how to build with the reusability and you're going to get value that you don't even realize is going to come as you continue down that path. And then finally, we learned that agents earn trust the way people do. Uh you've probably all seen everyone trusted agents blindly at first, LM's
[32:35] blindly and now there's a lot of hesitation in different areas. You have to have clear scope, show the work, make sure that you grow autonomy, autonomously with the evidence. And so what are we doing now? This next step right now we're on agentic root cause
[32:48] with Genie. We're looking at how do you actually have the system tell you why and what you need to fix so you can move that forward. We're also scaling towards uh compressing the decision latency. How do you move faster? So we're doing the weekly forecasts. If we have a major
[33:00] disruption that we could not plan for if you have a World Cup game that's coming up and you have no hes no history of what that could be in your data, how do you plan for that? But then how do you adjust for something that's very abstract like weather that you had no
[33:13] idea was coming up in a specific day and how do you get the routes readjusted for that? And then finally, we're looking at how do you push decisioning to partners? We're focusing a lot on our demand planners right now, but how do we get this to the people who are out in the field, who are the drivers, who are on
[33:26] the routes. How do we help them with decisions they still have to make that are still difficult to come across? And so this clock is still ticking. The game is still happening. It's not going to stop. And it's because we're still
[33:40] delivering that forecast. Our routes are still going. We're still rerunning processes. We're still building forecasts. And we're doing it all at the speed of the game. And so I want to thank you all for your time today. Uh enjoy the game tonight. And we're very
[33:53] happy you're able to come and spend some time with us. >>

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.