Embracing AI to Elevate Foodservice

CFVG August 20, 2025 Meeting Transcript

Meeting Facilitator

• Richard Poye, COO, Food Trends Think Tank

CFVG Member Participants:

• Ryan Blevins, Director of Food and Beverage Innovation, Weigel’s
• Brandon Frampton, Director of Fresh Food, Loop & Poppy (Loop Neighborhood Market)
• Stephanie Gallentine, COO, Lassus Bros. Oil, Inc.
• Kris Klinger, Vice President Auxiliary Services, Boston Univ
• Bonnie Zaring, Executive Director, Food Programs and Offers, RaceTrac

Presenters

• Mike Weber, CMO, Upshop
• Steve Brask, Director, Business Intelligence, Upshop

Guest

• Steve Brask, Director, Business Intelligence, Upshop

Ally Supporters

• Jon Cox, Vice President Retail Foodservice, McLane
• Mike Weber, CMO, Upshop

Vision Group Network Founders

• Myra Kressner, Founder, Kressner Strategy Group
• Eva Strasburger, President, StrasGlobal
• Roy Strasburger, CEO, StrasGlobal

Meeting:

Richard Poye:

Good afternoon. I’m Richard Poye. Welcome to our third gathering of the Convenience Foodservice Vision Group. We have a relatively small group today which is probably going to allow us to be more conversational, which I think will be valuable for this topic. I also think it’s important and beneficial that we have retailers of varying sizes from different regions of the country.

You need to think about our ambition when you go back to our first meeting. I talked about wanting to elevate the quality and perception of convenience foodservice. And one of the things that we’ve talked about is using data to help us structure our menus, but AI and how we use it is a fascinating conversation and I’m really happy that we’re going to be able to dive into that today.

A bit of housekeeping rules as we move on. Keep your cameras on for the people who are participating in the conversation if you can. Let’s mute microphones when we’re not talking and use the raise hand icon if you have a question or comment. We really view the chat as an opportunity for you to share something like a link to a particular relevant site so the rest of the group can see it, but we don’t want it used as a sidebar conversation that splits the group’s attention. If something is important, let’s get it into the text. And I think this is a really great opportunity to engage, discuss, and share some insights whether it’s insights from your business or if it’s something that you’ve gleaned from visiting a conference or speaking to colleagues in the industry.

So, like I said I am super happy about engaging in this. Also, if anyone wants to say something off the record, please let us know so we can make sure that it’s redacted from the text. And so, if there’s something that you want to bring up or say that you’re unsure about, definitely feel free to tell us. And I want to thank our Ally Sponsor members working with us — it’s super important to have Upshop and McLane as we move forward. At this point I want to reach out to Roy or Myra to go over antitrust or any of the other things I might have missed in the conversation.

Roy Strasburger:

Well, I’ll be happy to do that. Good afternoon, everyone. I’m Roy Strasburger, one of the founders of VGN. As you know, we are recording this conversation. The recording and the transcript of this conversation will be used in the Vision Report and as Richard mentions, you do have the opportunity to take anything off the record. Just let us know before or after you say it and we’ll be happy to work with you on that. The second thing is that some of us are competitors and we want to be very careful that we don’t violate any antitrust laws or regulations [Publication and Antitrust Statement shown on screen].

So please do not discuss anything about market share, dividing up customers, setting prices or anything else that may lead someone to think that we are fixing a market or being involved in antitrust activities. If someone does mention something that one of the VGN staff thinks is inappropriate, that person will be warned. If that person repeats it a second time, they may be eliminated from the meeting, so we appreciate your help with that so that we can keep everything legal and above board. So, thank you Richard.

Richard Poye:

I think we’re at the point right now where we introduce anybody who is new to the conversation, and so Steve, since you’re joining the call today as Mike Weber’s guest, but you’re also going to be contributing to the conversation, would you give us a quick bio so we can get to know you better.

Steve Brask:

Absolutely. So, my name is Steve Brask. I’ve been with Upshop itself going on four years now. Prior to that I worked for a company that was acquired by Upshop. I’m the director of business intelligence, but I also serve as the product manager for forecasting and data within the organization. I joke often enough that if folks in our organization have a question about data or numbers, they oftentimes come to me, so I run the gamut between raw data all the way up through forecast generation and dashboarding.

Richard Poye:

Fantastic. Well, thank you Steve. A few things as I get ready to hand off the conversation to Mike. Mike spoke about this topic in June and as I listened to what he was having to say I really liked the fact that it was kind of agnostic from a branding point of view, it was really about how we begin to kind of all dive into this. I understand that we’re all at different levels of assessing AI, or engaging AI in our business. And so, I think the presentation today will allow anybody on the call to be able to understand where they’re at and where there’s opportunities. I don’t think anybody’s at the end all area where it’s just totally operating by itself. So, I think this allows us to have a structured conversation. I’m going to hand it off to Mike and thank you very much for presenting today and looking forward to hearing what you have to say and the conversation that follows.

Mike Weber:

Really appreciate it, Richard. A little-known fact about Steve is that he’s also a new dad, so a big thank you, Steve, for making time because I remember I didn’t have the most coherent head when I had our first, so thank you for that. And yes, it’s really great because I’m only really an enabler of the content.

[Slide 1: Winning with AI] A lot of what Steve and his team do is what you’re going to see covered. And to Richard’s point, we were really excited about the opportunity to do this because how you approach AI should not be a secret. It should not be kept inside of a black box of a partner, a technology partner, or a consulting firm. I think with the amount of transformation that AI can influence, that we’re all probably starting to experiment with and expect, it’s for the best of everybody that we understand what the capabilities are now. But also, how do I think about my journey so that you’re approaching it in a way that works for you and your business.

And that’s really what we want to talk about because just like you, we are there as well. I mean there is a lot of learning still out there and that’s probably what keeps a lot of our business excited in what are the realms of possibilities.

[SLIDE 2: AI is Rewriting the Rules] So I want to go through how we’re setting up the presentation for today is really the key point. We start with this concept that AI is rewriting the rules of pretty much everything and it certainly has and will continue to have massive impact on foodservice and the business, your businesses. That’s an easy statement to make. But the point, “what does it mean for me? What do I do with that beyond picking what my ‘go-to ChatGPT is?”

[SLIDE 3: TODAY] So going at a pace so it gives the group time to discuss and for Richard to facilitate, I just want to cover three things today: What should it do? As you think about the scope of your role today or where you’re taking your department or how you want to drive success; What does Upshop believe it should do for you? What does that really look like? So, going from the classic principle to practice; and then how do you get started? And all of this is open. I mean, this is not, like I said, we are not precious about what we’re showing here because we’re encouraging you to take it away. If this is something you’re still scoping a journey yourself and your team, if you’re already working on things this will help you sharpen that effort. If you are working with current partners, we would love to work with you. But equally, these are questions you should be asking all of your partners in technology and data. So, feel free to leverage any of this as this is a group that we really want to increase collaboration with.

[SLIDE 4: Double Edged Sword – Growing Your food service offer..] And this is just one more slide as a setup before I get into it. The growth in foodservice has all these great challenges. This is why we do it. If it were easy, it would be done and you wouldn’t need great minds like yourselves to figure out how to handle things like growing food safety challenges, disconnected technology, how we guide tasks or be more prescriptive about tasks. If you grow your foodservice part of your store, that means that there’s going to be more things for your team to do. Item changeups like LTOs, finding your signature items, playing around with them, experimenting with them, I think that’s what makes the industry really exciting right now is all the different sort of taste profiles that exist. It’s pretty cool. And then just the final data disconnects, trying to get to a single data language in a retail environment sounds easy, but we know that it’s not.

[SLIDE 5: Does This Workflow Look Familiar?] And so, there’s all these challenges and there are many different workflows that still happen — I use this one all the time. If you think of a prep, getting your prep ready, if you have coolers, if you have things that you’re just warming, maybe you actually have assembly lines based on the complexity of your food. There are these steps that go into place and you’ve got to have, your team has to get in there and they’ve got their initial guide, maybe they print it out, maybe it’s on a tablet, they have to do a quick spot check if there’s anything left from the day before, hopefully not. If a lot of it was timed out, they’ve got to make some updates and then they’ve got to get going. And that’s a lot of work. It’s a lot of mental math.

[SLIDE 6: Worker Smiling – 6 Questions] Workforce And what we’re asking our teams all the time is, especially in foodservice, how much do I make? How long should it take me? When do I do it? Should I just do it all in one go? Am I doing it on the hour? How do I make it if I’ve never done the same thing before. Somebody the other day showed me a recipe for a hoagie, and I was like, wow, that looks pretty straightforward. And then they showed me all the sub recipes, from the sauce to the way that they want the bread warmed. And that’s not easy for one person if you’ve got all these other things to do. How do I handle exceptions? I’m trying to find all the things I need to throw together to get the day started and I get a particular item in the back room, what do I do now? And then the one we would all love the answer to is how much more could I sell?

[SLIDE 7: Ideally, AI Should] StrategyI’m going to start by saying, what should AI be influencing or supporting your business on? And these are some big things, but in foodservice, this is what you should expect. That AI first and foremost should be supporting a more accurate forecast for items on a daily basis by location, so it should be guiding what your plan looks like by store. It should be telling you how to best guide production of that forecast. It should be telling you when and how much you could be marking down the items based on when they were made and when they were potentially going to waste out. It can be managing or measuring what your waste is and identifying items that potentially could have less waste if they were better ordered for. So, looking at scenario planning around waste. And then finally, it should play a role in traceability in terms of flagging high risk ingredients and be able to trace them all the way in from receiving but also being able to serve as a way that you are updated on nutritional facts and allergens. All these things are what should be in your brief for AI embedding over time.

[SLIDE 8: AI – Haven’t We Been Talking..] And before I get into how do you get there, what does that look like? I just want to stop and talk about the elephant room, which is: haven’t we been talking about this for a long time? Why all of a sudden has this blown up? And I like these slides because they’re a helpful reminder of the journey that we’ve been on.

[SLIDE 9: The Advancement of AI] Data AnalysisThis is not a surprise, but I think the history is relevant in terms of what’s happened in business because you think of the 1950s and this is just purely AI as a computation play. We all remember the pictures of, I mean I don’t remember, I’ve just seen them on TV, but you see these pictures of huge machines, and it was just about process. It was like an equation process, and it was just crunch, crunch the numbers, and start to model compute.

In the mid-eighties, they introduced something called the era of expert systems. And that was really a logic base of if/then rules for decision support. So, the compute just continued to get faster, and it was a matter of figuring out what are the ways that we can put together rules in order to move through those computations faster. In the nineties, we were defined by statistics and algorithms looking at how we could encode relationships between variables. And that was really about uncovering patterns in data. The nineties were starting to look for patterns and when you can see patterns, you can start making better hypotheses or at least try to solve hypotheses in new ways. And then 2010’s was the focus on machine learning where you could really start to allow the machine or your computer to learn based on those patterns. Now the big aha has obviously come in the past couple years with GenAI. And instead of using rules or relationships or patterns, we can now simply say, give me an outcome, discover my inputs.

And that’s pretty crazy. It’s really your chance, and our chance as organizations, to uncover complex relationships across massive data sets, in a very short amount of time that we as humans could never see, we could never be able to find those things because frankly we can’t cover that much ground and we probably couldn’t even hypothesize some of these things that AI is able to find.

[SLIDE 10: Generative AI] And I just like this one because it’s sort of like how do we keep this simple? And if you remember the inception of AI, it all goes back to computers starting to learn and we were starting to teach them things and when you get massive data sets you’re trying to label, we’re looking for things that could be labeled consistently. We were saying, okay, computer, is this an animal? Is this a cat? And ultimately, we are able to train, right? Because they’re easy for us to label. As much as they vary, we can start to figure out if the computer can tell if it’s a cat.

The key now is once we could say, show me a cat, and it could start doing that and we could say, “Hey, show me a cat that looks like my cat” and I can provide an image, and it can find cats in my area. All of a sudden it looked exactly like the one I have. Well, this just goes [the same] for retail. It was easy to ask something based on data. Did I reduce shrink today? Now the key is how do I start to use technology to answer the big question, which is, show me how to reduce shrink and what are my options? And this is really the point.

[SLIDE 11: The Goal] Strategy The goal of any AI brief should be how are we using AI in operations to optimize thousands of these big and small decisions every day to drive performance and to be able to do this in a relatively small amount of time.

[SLIDE 12: The Journey] We see the journey as four steps. If you’re writing your brief or if you’re amending or if you’re sharpening or if you have an AI council, feel free to take this step away. Because we see there are four big steps to how you embed AI. And I’m just going to go through each one and I’ll leave with all the key takeaways. And Steve, this is where I’ll ask you to jump in for each of these steps to give it a little bit of color. I’ll just tee them off, then if you could provide some comment. What we’re going to do is show you some examples of how we’re doing it so you see how this really works as a baseline or a benchmark for how you’re thinking about it.

[SLIDE 13: Introduce AI to Your Data] Data Analysis The first big thing is really around the introduction of AI in your data. And this is a matter of how do I start applying AI in the first case to the data I have?

[SLIDE 14: A Demand Forecast] And we believe the most effective starting point to embedding data or introducing AI to your data is by building a demand forecast that uses AI. And Steve, if you could just talk a little bit about what that means, that’d be great.

Steve Brask:

Absolutely, so our businesses are generating all of this really powerful data. And so, the demand forecast is really the foundation for where we can start to leverage that to drive a measurable outcome. And so, we’re incorporating historical sales, cannibalization and halo events, promos that are occurring, price changes that are occurring in our businesses, and then also layering in third-party data. So, data like weather, holidays and other special events, a calendar dimension on top of the existing data that your business is already generating. And so, in combining all of that, we start to build a machine learning model, an AI model that’s generating a demand forecast.

And effectively what we’re doing is predicting how much demand we expect for tomorrow or some number of days on the horizon. And so, at Upshop, our horizon ranges anywhere from one to 120 days. But once we have a demand forecast at the most granular level, in this case in foodservice, it would be for production planning. That’s where we can start to scale out into other areas of the business because our demand forecast from an individual store’s production can then scale into ordering, it can scale into labor forecasting, it touches a lot of different areas of the business. And we can use that to help predict other roadblocks or other opportunities that are on the horizon as we start to drive out that horizon even further than 120 days.

Mike Weber:

[SLIDE 15: Forecast Accuracy Trend Chart] If there is any pushback on the ability to create a forecast like this, if somebody in the organization says, “Oh, this is going to take us way too long,” or “our data is too dirty”, that’s where AI can come into play. If you have any amount of data, of course more is better, but the ability for the machine to learn and to look for changes and to be able to create a forecast is more straightforward than ever. And Steve, if you could just talk a little bit about how we read this. My point is, is that a forecast like this and introducing AI to it is the first step because it is more accessible now than it’s ever been.

Steve Brask:

Absolutely. In this case, at Upshop, we generate demand forecasts for production planning, ordering, enterprise replenishment, for all of our customers. And so, we evaluate accuracy very heavily when we generate these forecasts on behalf of our partners. And so that’s part of the feedback loop that you would want to establish as you look to deploy AI. You need to ensure that there isn’t hallucination. And in this case, our front-facing metric is bias, but we evaluate a host of other metrics to essentially evaluate the accuracy of the demand forecast we’re generating. And in this case, all we’re looking at is what was the predicted demand for a given period compared to actual sales or actual demand in this case?

And that’s what this chart is evaluating. And so, at Upshop, we have a goal of 95% accuracy or better at an enterprise level. We exceed that in general across all of our partners easily. In this case, you look at 06-13, we’re looking at Father’s Day weekend. Holidays are always an opportunity for every single one of our partners. We don’t have a single partner who isn’t concerned about holiday forecasting. And so, in this case, we’re returning that holiday weekend with 99% accuracy in terms of demand forecasting. And so even with poor quality data, the nature of machine learning and AI in general can bridge the gaps that exist.

Mike Weber:

[SLIDE 16: Train the System] Strategy So that’s the first step. The second step is really starting to [train the system]. The key for AI is you’ve got to tell it to do something and it wants direction, it wants to solve a problem. Back to my whole point of the slide on the cat, it’s not “is it shrink?” It’s like “how do I more effectively manage shrink? How do I get rid of it?”

[SLIDE 17: Manage to the Right Metrics] And the next step is really being able to identify the metrics that you want to use AI to optimize. This is really critical. And I shared some of these metrics at a session recently, because I was like, everybody should have a point, should have their own “approach to Moneyball”. Your business has its own strategy.

Editor’s Note: The Moneyball approach refers to a strategy that emphasizes data-driven decision-making and organizational change.

Strategy You’re building a very differentiating foodservice strategy, so that will drive what metrics matter to you, but this is where you need to then train the AI to help you identify scenarios for it to optimize those measures. And Steve, before I go to the next slide, I want to move through this one quickly, but maybe you could talk about some of the measures that we recommend looking at in foodservice on a real-time basis. Some of the core metrics here would be great.

Steve Brask:

Workforce Absolutely. When we push out an AI tool to our store teams, the biggest thing we want to ensure stores are doing is using the tool, right? We’re putting all of this time and effort into the tool. We want to make sure that we’re getting the most value out of it. And so, in this case, we evaluate a couple of different core metrics to arrive at our overall performance score. And so, planning engagement rate is simply: Is the store team engaging with the tool? Are they using the tool in their production cycles? Forecast compliance is evaluating whether the team is following the forecasted amount that we’re showing them in the UI (User Interface). And so, in this case, we’re going through all this effort to generate a forecast. Is the store team adhering to that when they’re in their production cycles?

And the last metric there is menu compliance. So, in this case, we do a lot of work around SKU rationalization and understanding the optimal menu for a given store. And we want to make sure when our store teams go into production they’re producing all of those menu items that they’re available for sale to our customers. Among the three of those, those are really the foundation of driving that change management component that comes along with implementing an AI solution.

Mike Weber:

[SLIDE 18: Chart – Sales by Day of Week..Optimized Schedule] Steve, can you talk about once we have AI in the system, it’s learning and being trained on how we produce and how our stores respond to those production or prep schedules in foodservice. Then you can start to optimize your actual prep rates or production rates to improve your sales. And Steve, I know this is a screenshot of the dashboard, but could you talk about how we are using or how you should expect to use AI to optimize for production?

Steve Brask:

Absolutely. So we just talked about the change management component. That’s one component of deploying an AI tool. The other component is optimizing the tool itself. And in that case, we want to make sure that the tool is giving us the best possible outcome from the forecast that we’re generating. And part of that is optimizing production schedules in our menus. And so in this case, we’re leveraging things like shelf life, average sales and other components of our menu structure in order to arrive at a point where we are defining how often we should produce, what days of week, what time of day, using AI to tell us this is the optimized production guidance for a given item at a given store.

Mike Weber:

I think this is scraping the surface of how AI can guide if I’ve decided to staff in a different way, how does the production change based on that staffing? It could do the same thing for that scenario as well.

Steve Brask:

Exactly. And that’s where we’re leveraging the AI component to expand beyond just the simple rudimentary dashboard analysis. We want to automate this process.

Mike Weber:

[SLIDE 19: Partner with AI] Workforce Training So just a couple more slides. The third step is partnering. And this is really where we have to ensure that our teams understand the power of AI and how they continue to embed it in their day-to-day work. And how it can, not just in our teams in the office, but more importantly, how is AI being partnered with our store teams and our associates so that they can ultimately be more digitally enabled to do their job.

[SLIDE 20: Guide Your Team] This is a system we’ve continued to finesse. Over the next three to six months, AI will truly transform task management. Today, in many cases, it serves to prioritize tasks. Tomorrow it will be prescribing the right task, at the right time, for the right team member and will dynamically change based on agreed system conditions. We see that as happening in a very short amount of time because of the amount of data that can be put into a system like a task management solution for your teams. So, this is definitely one of the biggest influences that AI is going to have is in dynamic task guidance in store.

Data Analysis The other area is going to be in your day-to-day data. And I just used a little video that I made a long time ago now for how our platform just has a simple overlay, that enables people to work with it and partner with it in the home office, in your corporate team. So, you can create dashboards and certainly run analysis in a matter of minutes instead of hours. And so Steve, I don’t know if you just want to talk about how this works? This is a pretty straightforward overlay that frankly, if you don’t have something like this, we would definitely recommend some ways to build this into your system.

Steve Brask:

Absolutely. So, one of the challenges that we encounter with every single one of our partners at Upshop is we have some folks who are power users of the data, and then we have folks who are not especially data savvy. And it takes both types of people to have a successful business, but we want to help give the folks that are less data savvy the tools to make the best possible-informed decision.

And so, layering in an LLM (Large Language Model) over the top of your data in this way, which is essentially what we’re doing here, is layering an LLM over our partner’s data to help give a natural language query tool to our partners. And say, “Ask questions in whatever language you feel comfortable with, and then get actionable insights out of your own data.”

Mike Weber:

I think the other big takeaway here is that people talk about metrics in different ways. Or when you ask: “How is my store performing?” There are many different ways to think about that or to look at and evaluate it objectively. And the LLM can continue to learn based on how your business is oriented. And so, it can get even faster and even more effective after just a small amount of use across your team.

[SLIDE 22: Autopilot] The final step is the biggest one: “when do you turn on Skynet?” And I say that jokingly, but at some point, there is the ability for AI to start drawing conclusions and to offer recommendations, or at least at minimum, to provide different scenarios based on what’s happening in your business.

And so, this video is similar to the last one where this is the world that we’re shaping right now with the data and the platform we’ve built. This is not that far away where you can consolidate all your data.

You can create all the different reports you want, but then more importantly, you can start to look and ask the system to proactively anticipate change and bring you recommendations. So, this is just a video where we started to show you can double-click into different performances, it is what we call a cockpit.

But most importantly then, you can start to ask an assistant, “Okay, what things should I consider changing?” So not just what data I should understand and be able to work with, but I’m going to start to build an agent on the platform that can be an expert in things such as supply, such as limited time offers.

You could build an agent around waste, and you could start asking those agents to learn based on what they’re seeing in your data, and to provide you with recommendations. Or to at least at a minimum, provide you with recommendations for stores that should be focused on, and what you think are the drivers of those issues.

So, this is where it gets really like the next level up. You can start turning those agents live on your data and looking for how they practically help you think differently. So, this is just one where we’re like, “Okay, let’s look at a holiday.” You’ve got the sales leading up to it. And we’re just starting to ask it questions.

But ultimately, what we want to do is to start looking for its recommendations. And these are topics that it would feed to you automatically, that you can then start interacting with your data. And this is where, yes, the sky’s the limit.

[SLIDE 23: The AI Journey] StrategyThere are four core steps back to “what should you do?” There are four core steps that we recommend for any brief or workaround embedding AI into your org and your specifics or remit. “Introduction” is the first one, and that’s not just data cleanliness, but how do you apply AI to your demand forecast?

Second is “train.” This is where if you’re able to identify those core metrics, both in execution metrics, but as well as in the key ones you’re watching for in terms of inventory, turnover, sales, that you can use AI to optimize. And so in our case, we showed you the production schedule, but you can optimize ultimately for anything based on what metrics are your benchmarks.

[Third is] getting your teams to “partner” with it. So, in terms of guided tasks, you can create content at rapid speeds now. If you create a recipe, I’m sure that in a matter of seconds or a minute, you can create a little video. We’re creating so much more content for our recipe management system when a retailer says, “I’ve got all this stuff in PDF and I’ve never really documented how to do it.” If you feed AI a recipe, it can create a video for how to make that recipe. And then that real-time scenario management.

Then finally, “autopilot.” There will be scenarios in your business when you want a machine to start to model and to give you not just insights but give you recommendations for how you should think about, or how you should prepare it differently for a launch.

Or what’s the likely turnover or sales potential of a new store we’re about to open? There’s incredible potential in what they call universal or global training. Once an AI can see so much data and it can see all these different variances, it can start taking some big jumps in how it estimates things that have never happened, like a store opening or a new product launch.

[SLIDE 24: Q&A] So I will call it there and leave time for questions and for discussion.

Richard Poye:

I like the framework. I think it’s simple. When I look at the retailers here, they’re all dealing with different types of issues and problems and complexities. And so, I think it’s going to allow us to break it into those quadrants and discuss it. And I think that there may be people on the call that really haven’t started down the path.

And so, if there are any insights that anybody wants to share with people, that might be interesting also for us to talk about jump in. I think about some of the smaller retailers that are out there that might read this report, and they’d be like, “I’m overwhelmed by this.” And so there might be things that you could help them with.

Because if you started cleaning up your data four years ago, five years ago, you’re in a much better place than somebody who is dealing with a more complex system and things like that, especially smaller retailers. I’m going to throw a question out to Jon. From a McLane point of view, are you engaging any retailers from an AI point of view yet? Is anybody feeding you information or can we not talk about that?

Jon Cox:

Inventory Management Not yet, but we’re getting a lot of questions. What we think AI can do is from just an order predictability standpoint if you think about it. I’ll pick on Bonnie because I know her. If you said, “Hey, this RaceTrac is at this latitude and this longitude. And every year at this time, there’s a bicycle convention and we see your sales go up 10%.”

At some point, AI is going to help us write the forecast for her so that her store team doesn’t have to remember that. And so, what we think over time is that AI is going to help us better predict what product we need and when we need it. Even to the point of being able to go back to some suppliers and say, “You can expect a lift eight weeks from now based off of all the data we have, so if you don’t have yellow paper to wrap the Peanut M&Ms, you should go get it.”

And so that’s what we’re really trying to use it for is from a predictability standpoint of how do we engage with the suppliers to keep our customers better in stock? We haven’t had anyone yet that said, “Hey, could you take and consume all this data for us and help us be better?” If that makes sense.

Richard Poye:

I think it makes sense. I think also, for instance, if Bonnie would be moving into opening a store in a new region, you might be able to push some things that would help her develop her forecast. So maybe that might be something to consider.

Richard Poye:

Roy?

Roy Strasburger:

Great presentation. Steve, thank you very much for the information, and Mike, as always, a lot of information to take on. Going back to what Richard was saying about the medium-size, smaller-type retailers of how to get into this.

Can you give me a one-paragraph, two-paragraph reassurance that if I got into this, I wasn’t going to get totally swamped? What do I need to do? If I were a medium-sized retailer, how do I approach this knowing that I can probably use it, but I’m concerned about taking that step going forward?

Mike Weber:

Data Analysis POS I would say not to be [a] canary, but the first thing that you can do, even with multiple point of sale [systems], let’s say you’ve put together 20 stores and you’ve purchased them with different points of sale, different data. AI is made to help you shape what the likely forecast would be across your different stores, and they could be in very different locations.

And that starts to demonstrate to you how you can think differently and how your stores should be set up differently. And Steve, maybe you could jump in. I love presenting this stuff, Roy, because I’m learning all the time, and I mentioned these models being able to train themselves.

And this is perfect for the industry because you’re creating an LTO or you’re maybe looking at buying another store or buying a couple stores, and you’re asking yourself, “What’s the ceiling there or what could that likely do?” And these models can start to show you some different scenarios very fast on a very small amount of data of how that could work.

Steve, can you just define, this is more of a practice, what is global training or how does that work? Because that’s where you get a lot of skepticism, and this is real in a lot of industries today.

Steve Brask:

Absolutely. So global training is something that’s really only been possible at any scale really. But on the kind of scale we’re talking about with POS data, global training is something where we can give a machine learning model a small amount of data, incomplete data.

So, we may have items that go on and off menu multiple times a year, year-over-year, and those items may not always be carried at the same stores. What global training allows us to do is take that limited dataset and generate a demand forecast for that item across all of our locations, regardless of how much data is available.

And so, if I’m approaching this from a small team standpoint, I’ve seen at this point a couple dozen implementations of demand forecasting across all different size retailers. And I will be candid in that the smaller and more tightly-knit the team is, the better the outcome generally is, especially because you’re minimizing the noise in that process of implementation.

But even with limited data, even if you just decided to pilot a demand forecast with a handful of stores, global training can learn from that limited dataset and provide a demand forecast across the board for the entire organization.

Roy Strasburger:

Okay, so two follow-up questions. First of all, is global training a technical term that’s used in the industry or is that an Upshop term?

Steve Brask:

So, there are multiple ways to refer to it, but it’s how we choose to refer to it. But global training, it’s essentially applying the same model features even without data. And there’s a lot of different ways to refer to it. Global training is just the term we use because it’s the most approachable.

Roy Strasburger:

Okay. And then the second thing is, going back to what you’re talking about to the incomplete data., are you suggesting that if I were a medium-sized retailer and I don’t have confidence in my data as far as whether I have enough of it, you can take what I have and work with that to get to a workable AI strategy to allow it to be effective?

Steve Brask:

Absolutely.

Roy Strasburger:

Because I think that’s what a lot of retailers are really worried about is their data, and whether or not they can make that step into this game. Because they keep thinking they need to be Google, and they don’t have to be Google as far as what their data is.

Steve Brask:

You definitely don’t have to be Google. And that’s the great thing about AI, especially with global training layered on top. We just did an exercise for one of our partners where they took some items that were previously on their menu about four years ago, and they were considering reintroducing them.

And they wanted to build a demand forecast for those items based on data that’s four years outdated. They haven’t sold these items in four years, and they wanted to generate a demand forecast for today. We’re talking about data that was generated back in Covid, which is a drastically different timeframe to generate a demand forecast from.

And we took that data, we generated a forecast, and with global training, we arrived at 87% accuracy just with that limited dataset. And then the great thing is that as you continue to leverage the tool, your teams start to use it, you’re piloting it, you’re rolling it out. You’re generating more reliable data and the models will always rely on more recent data more heavily.

That’s the great thing about it. So, as you continue to improve your operations and generate more reliable data, your model also continues to improve.

Roy Strasburger:

Okay, thank you.

Richard Poye:

Hey, Kris, you’re first and then Bonnie’s going to be after that. So Kris, what question did you have or what comment?

Kris Klinger:

Inventory Management Sure, and I guess more of a comment. We’ve deployed AI in regard to waste tracking and management and then plugged it into our menu management and ordering system. I’m at a university and we piloted it in one of our dining halls to where they were tracking the waste at the end of the day, and then that was being captured and it was pretty simple.

They would sit it underneath the sensor, and it would tell them what was left over and put it in the system. And then the system would adjust and produce production or updated production documents and sheets. And so that was for one of our smaller dining halls. We’ve since employed it in all the dining halls using the same or similar data, despite the fact that the demographics were a little bit different.

But we’ve also seen it where, to your point earlier, Steve, we’ve brought back old menus and done similar things to where it’s given us useful and fairly reliable [data]. I would say in the mid-80s, which is pretty good, and then it did continue to learn. So we’re automating our production. And our goal really is to centralize it as much as we can and then distribute it.

In the AI, we’re trying to train it on how to help us with the distribution and the overall production. So it’s a pretty interesting technology that obviously has a long way to go in regard to there’s a lot it can do for us, but it’s done quite a bit already. I just wanted to share that.

Richard Poye:

Thank you, Kris. Hey Bonnie, you’re next up.

Mike Weber:

Excellent use cases. Thanks, Kris.

Bonnie Zaring:

I would just say, again, great presentation, really appreciate you sharing. Mike and Steve, you talked about limited data and having that opportunity to have a small sample, and how to make it larger and meaningful. But a lot of what happens in the c-store industry around data is lack of specificity.

We sell a lot of things as a one. And sometimes being able to take that and make it actionable when you don’t actually know what was in the cup, or you don’t actually know when it was all rung up the same item. How do you overcome some of those data challenges when you’re trying to learn from what’s happening in the store?

Steve Brask:

Data Analysis Great question. And I think Kris’ use case is one of the solutions that we’ve been leveraging pretty heavily in terms of backing into things like assortment items. So if you think about a roller grill item where you might sell every item on the roller grill as a single PLU. How do you back into the different flavors?

Or a doughnut case, how do you back into the different flavors and types that you have in the case when you’re selling it under one PLU? In Kris’ use case where you’re leveraging waste data to back into those types. So, we can provide a demand forecast value that’s based on that single PLU.

So, we know we’re going to sell 400 doughnuts at this store. We can then use the waste capture data to back into the different varieties and then generate a demand forecast for those varieties. And that’s one of the complex use cases that we get frequently from our c-store partners, especially where we have one PLU for roller grill, how do we back into our varieties?

The other use case would be steam tables where you have products in a steam table and say, “I made four pounds of mashed potatoes or whatever it may be. How do I evaluate the waste that’s coming out of that? How do I evaluate the different varieties in my steam table if I’m selling steam table items under one steam table PLU by weight?” And so waste capture factors in heavily to that.

Richard Poye:

So, I guess one of the things within that, you might actually be able to get your shrink as well. So if you capture your waste and you know your ring and you know what types, then you might be able to get that.

So, I think about pizza. And if you sell pizza or a roller grill, same kind of thing, you could be able to then say, “Okay. Well, now I’ve got a shrink issue because I’ve got this many units sold and this is what I’m wasting.” And identify your shrink then as well, so interesting. Roy?

Roy Strasburger:

So, Kris Klinger, follow-up question to you. As I think most people here understand that you work with institutional food, specifically at Boston University, the cafeteria system.

When you’re talking about waste, are you specifically talking about the production side of what’s happening in the kitchen? Or are you trying to do any work on waste from the consumer side, whether or not students are participating?

Kris Klinger:

Both pre- and post-consumer waste we’re tracking. So that helps us to, one, understand what’s not being utilized during the production process and what’s being wasted. And then also post, in regard to what’s not being sold or served, or whatever the case may be.

Even to the point where we’re tracking in the trash cans what the students are throwing away as well to track that waste, so that’s a little bit separate. With that, part of it’s shaming the students for taking too much and then throwing it away, we’re trying to improve.

It’s selfish in our case, because obviously, we want to produce less food and produce less waste. But it also saves us money in the long run, but it also creates a more sustainable system and environment. So, we’re doing all of those, Roy, to answer your question. And we have different mechanisms that we’re using for that.

And then the system, one of them we’re using is called Metafoodx, and like I said, it’s a camera that sits over the food bins. And then also it tracks the waste in the trash cans as well, and it picks out what’s being thrown away.

Roy Strasburger:

Well, I was going to say, I didn’t want to be the person who was figuring out what was in the trash can.

Kris Klinger:

No, it’s automated. We used to actually do that though, Roy. We used to call them dumpster dives where we would have the students, they would go in and they would figure out what was being wasted.

Again, and then they would share that with the students and their fellow students, and it actually was really beneficial and helpful. That’s the old way of doing it. Now, we don’t get dirty anymore, we use technology.

Roy Strasburger:

Well, this leads to where I was going. Mike and Steve, is there anything that you can use Upshop for in monitoring whether or not the serving sizes are correct? For example, are we serving people too much food? Are we giving them a plate of food, of which they’re only eating part, that we could actually downsize the portions, save us money, reduce waste, possibly reduce cost to the consumer? But anything on the backside of it to make sure that we’re right sizing our portions?

Steve Brask:

We’ve actually done some experiments on this with pizza production. We had a partner come to us and say they’re batching their pizzas, they’re cooking a whole pizza, selling it by the slice, which I think is a scenario most of us are familiar with. And so in their case, they wanted to understand [the outcome] if they went down to batching half pizzas.

They were originally packaging pizzas in two-slice containers and single-slice containers. And so, by conducting that kind of experiment with package sizes, you can start to arrive at where the demand is. Does that drive sales on additional varieties, which is another thing they were looking for? Is there a halo effect to going down in package size?

Do customers buy other items in the assortment, rather than buying the two-slice box? And so that’s where you can start to really leverage the data in some of these experiments. Even if it’s just a simple pilot program, we’re going to push it out to 10 stores and see how this performs, you can start to leverage that demand forecast data to understand where the opportunities are.

Roy Strasburger:

Right. But I was trying to approach it more from just a position of what the consumer thinks they want and what they actually consume.

And that’s where we get so much food waste from a post-production side of it, if people’s eyes are bigger than their stomachs, as it were. And whether there was something that might help us try to figure part of that out if it’s consumed on-premises, and if you’re able to track those types of things?

Mike Weber:

Yes, I was going to say, we don’t have a direct answer. I think one of the things that Kris has alluded to that is going to be pretty available very soon, will be just the ability to access all imagery that you have and use that data. So even if you have security cams, I mean, I won’t speak to how old. But pretty soon there’s going to be a lot of layers that we can just add in from other suppliers that say, “Give us all your imagery you have in your store,” and we can try to make those leaps, Roy. Because the imagery from not only what’s getting tossed or what goes out the door, these are the use cases that are the next layer up.

That’s the next level up is that right now we’re solving in one line, just how much should I make to sell a certain amount? Now we want to get to the next layer, but it was like, “We’ll look at the total store in order to answer some of these much bigger questions.” That’s what we’re working with. I think that’s what comes next. We were just in a meeting and that was the example: all imagery and be able to ingest and relate to those business questions. That’s it. That’ll come next.

Roy Strasburger:

Thank you.

Richard Poye:

I’m going to pivot over to Robert, and then after that, I’m going to ask some questions for some people who haven’t had the opportunity to speak yet. So Robert?

Robert Hampton:

Thank you. Great presentation, guys. A question, I think it’s more for Steve. I did quite a bit of work at my previous role with data and cleaning data. What would you say, or what advice would you have for the small to medium retailers who are not jumping into this, into AI, use of AI, and with your product, that they can do to overcome the fear of, “Oh, my data’s not clean enough. I don’t have the right data,” things like that? Do you have any suggestions of, “Here’s the data you need, and here’s the way to ensure that it’s good and can be used in any system?”

Steve Brask:

Data Analysis Absolutely. When I work with our partners at Upshop, the first thing I focus on is the amount of data available, the amount of history. And so really, when you’re doing any kind of AI evaluation, you want to be able to do trend analysis, so you want at least two years of history. And that history can be incomplete. We talked about LTOs that are on and off menu, seasonal items, holiday items, things like that.

That history can be broken up. But really, what you’re looking for is at least two years of history, but the more, the better. And from that, that’s where you get into the trend analysis. That is the root of everything that we do in AI, is trend analysis. And so as long as you have a reasonable amount of historical data, even if that data is incomplete, AI can close a lot of the gaps.

So you’re looking for the history. The other thing that we get a lot of questions around is I have SKUs that change, items that we sold one flavor now that SKU is being reused, or we’ve now mapped that item to a different sellable item, right? And so in that case, that’s where AI can also help out, is leveraging that demand from previous items, where we may have changed that SKU historically, but if we can identify that cutoff date, which most of our partners can, they can say, “We went through our SKU rationalization in fall 2023,” that’s where AI can start to determine, “Here’s the shift in demand.” So, we’re not having to do a bunch of manual cleanup, the machine learning models can detect that change in demand, and generate a forecast based on that updated demand signal.

Robert Hampton:

Okay. And just to be clear, this would be presumably from the point-of-sale TLOGs (Transaction Logs), which pretty much every retailer has access to and can go back, or at least start capturing today and archiving, so that they can build two years of data.

Steve Brask:

Exactly. And we talk about some of these other inputs, like weather data, promotions, and pricing. We work with partners that have different levels of data availability across the board. We have partners that have no historical pricing data. They change their prices, and they don’t track it.

We don’t necessarily need that because we can track price changes and price elasticity from the TLOGs themselves. As long as you have the sales data, that’s really the foundation of a demand forecast. All of those other features, like weather, we gather weather data from a third-party source. We’re not reliant on our customers for that.

So as long as you can tell us your store zip codes, we can layer in the weather data on our partner’s behalf. So, as long as we can get the sales history, the more data we can layer in on top of that the better around promotions and pricing and things like that, but the sales data is really the foundation.

Robert Hampton:

Got it. Thank you.

Mike Weber:

I would just add there, Robert, that I think there is skepticism around your data, but there’s also skepticism around, “Don’t these other data sets cost me a lot of money? A weather data stream, I can’t get that. I can’t get a supplemental data feed that maybe looks at local events by my zip code.” And the thing is all that data now, those streams are much more accessible, and the cost of accessing them [is going down].

You should probably go to somebody that’s able to pull them together, but that’s what you can add to. It’s like, “Build the picture of my business without me having the data.” You can do that. And I think that’s been the eye-opener when I speak frankly about all this stuff, is you just don’t realize how much stuff is out there data-wise that’s been pulled together in a really ingestible way, and that’s for a medium-sized player. I mean, small, maybe, but medium-sized, definitely.

Robert Hampton:

Got it. Thank you.

Richard Poye:

Eva, you have your hand up.

Eva Strasburger:

Workforce From the point of view of someone, an employee working in the store, what kind of feedback are you getting? You’re talking about the stress of someone who comes in, has a ton of work, is overwhelmed, but have you found that to them, they’ve said, “This is such an easy system. It’s made my job so much easier”? Are the people who are using it saying, “It’s a way to make us more interesting for someone to come and work for us because they think it’s cool that we’re using AI?” The people on the ground, what are they saying?

Mike Weber:

Yes, I love it. And that’s what keeps me excited about the business. There are, I’d say three takes. One is, in many instances, it’s been easier to roll out than ever because if you think about the way we use technology, you open up an app and you don’t need any training, right? You just start working with it because it’s so familiar.

Eva Strasburger:

Intuitive?

Mike Weber:

Yes. And I think that intuitive nature is what AI makes that. I mean, talk about if you’re not using AI to even guide your own interfaces and things like Lovable (AI), there are so many options out there for how you can learn from best practices and intuitive interfaces, and be able to have that up and running and changed in a very small amount of time. The second thing is that in many instances…and that’s why I was talking about guided tasks…the team doesn’t even know that something has changed. It actually is easier just to update things in…I wouldn’t call it real time, but it updates things, and you don’t need to worry about it.

It’s just done and you can just keep going. Imagine where the stress comes from, where I have to totally stop and change up and think “I need to now do a bunch of things in muffins and all of a sudden, something derails me because I don’t have enough wrappers. I can’t wrap them all.” And “Where’s that box?”

I think the big thing is you talk to associates, and they’re like, “I can’t remember the time that I had an attack, that something caught me by surprise. I just come in, I do everything, and it’s all done.” It just feels easier and that’s a huge win. And then, I think the final one is more around the satisfaction teams do have, when you explain why you’re doing it, and I’ll use Kris’ story, where now they know what they’re doing and how they’re using technology that does lower waste, and you show them results.

And so some partners have the classic countdown screen in the back room, and they show every week how much less has been wasted. If someone doesn’t say it, I do know that there’s got to be a better feeling where you’re like, “Look, our location, I don’t want to know how much it used to throw away, but now it’s throwing away a lot less,” and I think that does leave a positive impact on your team, even if they’re not saying it to you.

So those are the three that we get a lot from operators in terms of the outcome.

Steve Brask:

I would just add one thing. Before Upshop, my background was in retail. I worked in stores, managing different production departments, and I used to coach my teams all the time around, the customer is the priority, but the team oftentimes feels overwhelmed by the number of tasks they have to do, and they feel interrupted by the customer. And that’s not what we want when we’re in day-to-day retail operations. But when we take away tasks through this kind of automation, like maintaining perpetual inventory, we’re not doing constant cycle counts of our production inventory and we don’t have to go and count every item every day before we produce. We’re just making minor adjustments in a UI.

Now, we’re taking away that time-consuming task, and it’s easier for the team to focus on customers. They don’t feel like they have all these competing priorities. As I’ve met with store teams, and we’ve gathered this feedback, I think that’s the thing that stands out to me, is we’re reducing the task overload through these processes, rather than having the team feel like they have all these competing priorities all the time.

Eva Strasburger:

Okay. Thank you. If I could ask Kris a quick question. Kris, how do you do your food forecasting?

Kris Klinger:

What do you mean by how?

Eva Strasburger:

As in, you’re saying that you use this AI model to do waste. How are you predicting what the demand’s going to be?

Kris Klinger:

Data Analysis So we have a certain number of students who subscribe to the meal plans on a regular basis, or on a semester-by-semester basis. And so we plug in that information, and then we also plug in historical data. And this is the old way we did it. Now, obviously, AI is bringing a whole litany of additional knowledge and information to the process and the system.

We used to dump in the amount of meal plans, and we would put some of the demographic information in regard to the students, and then based on historical information, it would tell us how many students we would expect for lunch, dinner, breakfast, and then the flow, and typically, obviously lunchtime. There was an hour or so where they all try to come in, and dinner was sporadic or drawn out. Breakfast was pretty light.

It’s interesting Eva, because we have multiple stations in the dining hall, typically six or seven stations, like a main entree station, and some specialty stations, and a vegan station, a pizza station, and a grill station. And so again, all of that would be based off of historical records in regard to what the students were eating, where they were eating, and depending on what was being offered, and some of it was a bit of a crapshoot with AI.

Now, we even can tell if you offer enchiladas in the Mexican station and chicken sandwiches in the grill station, how many of each of those you’ll sell based on how the data has been pulled from the past. Does that answer?

Eva Strasburger:

Okay. So you’re using AI?

Kris Klinger:

Oh, yes. Absolutely.

Eva Strasburger:

You just tied it to waste. So I was wondering if you were doing something…

Kris Klinger:

Yes. Initially it wasn’t AI. We used couple different systems, menu management systems throughout the time I’ve been in higher ed, but now, we’ve latched AI onto it, and it’s integrated. Aramark has an AI tool that also is integrated.

So I think most of the larger feeders in our space are looking at, or either if they haven’t already added AI to their menu management systems, and then it’s just adding the other elements like the waste tracking and so on to help you with that. And the waste tracking brings it full circle, if you think about it, then the loop is closed, so you know what you’ve produced. It’s like a theoretical, and then you can compare your theoretical to your actual usage. And then, our goal is to be able to order and use it to order, so we don’t even have to place our orders, and even use it for inventories too, and we’ve had some limited success on the inventories.

Eva Strasburger:

Thank you.

Richard Poye:

Thank you. Stephanie. So it’s been almost a year or something since I was in your store, and looking at your offer, and you have your touch screens, but you also have grab-and-go offer. You have a pretty complex menu as far as lots of cooking. How are you managing that, and what’s some of your thinking as far as utilizing some of this technology to optimize your team’s work and things like that?

Stephanie Galentine:

Well, we’re not using AI. This is actually my second presentation today, so I’m really starting to understand the opportunities, and they’re inspiring, I will say. I still feel like I’m 10 years behind. But to answer your question, we have streamlined our warmer. So we do not just throw in whatever is on the menu, we’ve streamlined it, and we’re promoting with deeper discounts to push that product so that we can manage the back of the kitchen with less made-to-order because we’re making what is being sold.

We’ve been able to take the data that we do have, which is not beautiful. It’s not clean. Only we know what it means. So it’s difficult to have conversations with third-party support, but we’ve done enough where we can manage some of the work in the back end so that we’re serving the food faster and we’re serving what they want, and we’re able to manage that front end and give our employees a breath. I think it was Eva mentioning, “What are the teams saying about this?” And that’s really where our focus has been because we do have an extremely large menu.

And I tried to work back there, and I’m pretty sure I would’ve gotten fired if it was a real-life scenario. It was just very apparent to me that we do a lot of great things, and we have great food, and that’s great news, but the way we do it is antiquated. We’re not using the tools available to us. And so I listened to this presentation, and I’m like, “Well, maybe I just leapfrog my current plan and go start living in the AI space,” because we’re just trying to find a program that we can put on top of our price book and start to manage things online, the waste and the production. I struggle with our label maker and all of the intricacies to our label maker.

And so I think I was about to walk into something, and I need to walk around it, and go to the next opportunity, which seems to be the speed at which we move today. So it’s been enlightening, for sure.

Richard Poye:

How amazing. I liked your candid response, and I think your food tastes really good.

Stephanie Galentine:

One good thing we’ve got going.

Richard Poye:

Yes, customers really enjoyed it. So, it’s like some people might have great AI capabilities, but maybe their food’s not that good, so you’re winning there. That’s awesome. And, Myra, you have a question?

Myra Kressner:

Workforce Yes, for Stephanie. So I love, again, the honesty that you just talked about. What kind of pushback…or how do you then address this with the rest of your team? Do you need to bring them along with you? How challenging would that be, or would they say, “Yay, let’s do it,” “We agree with you,” “Let’s get going”?

Stephanie Galentine:

I’m there now. Myra, we had to redesign how we’re building the products, going through the oven to make sure that we’re getting to temp, and that the quality is held in the breads and the proteins and things of that nature. And I was all, “Yay,” and they’re like, “That is impossible. That is not something that’s going to happen here.” So it has been a slow evolution.

Thankfully, we had two in the room that were some of the busiest, and they’re like, “I think we can do that. We’ll come back in 30 days.” I’ve had a couple trailblazers that have been willing to be uncomfortable and actually go do it the way we’ve designed, and so we’ve got some leaders that are giving us support. But I would argue that’s the most difficult part is getting everybody on board and using the process as designed, being willing to adjust. And I was very frank, I’m like, “If you guys can fix it or make it better than this, I’m all in. Tell us how to do it.”

But some of my people have been doing it for 15 years, the exact same way. So for me to say, “That doesn’t work, that’s never worked, it’s always been wrong,” it’s offensive. And that’s a good thing, because that means my people care, but it’s been difficult. Lots of conversations, lots of follow up, lots of questions, but we’re in the people business. We’ve slowed down the process, I will say that. We had a schedule, and we’ve had to really slow it down to make sure everybody’s on board.

Myra Kressner:

Thanks.

Richard Poye:

Hey, Brandon.

Brandon Frampton:

Hello. First, great presentation. Really enjoyed it. And I think I’m somewhere between early adopter and whatever.

We get along pretty good with it, but we haven’t got high-tech yet. So, I’ve got three questions. One is: Is there a difference between OpenAI, Grok, or Claude? Do You feel one’s better for long-term in this space? The second is, I’ve got lots of raw data, really good. And to populate this, to make it work great…and I’ve got 150 locations and we’ve got food in about 100 of them…How long would it take to do this? We talked about, “Well, it’s easy, it’s not that hard, you can learn it,” but how many hours would that take, or weeks? And then, the last one is: Can I go hire these people? Is there somebody coming out of college that already has this knowledge?

So those are my three questions. Thank you.

Steve Brask:

Data Analysis So I think in terms of, “Can you go hire these people?” There are lots of data scientists who are more than capable of building a demand forecast model around your data. And candidly, the larger the data set you have, the better off you’ll be in terms of a forecasting standpoint. In terms of AI models, from my perspective, they continue to leapfrog one another, and I think that will continue to happen. Right now, I think ChatGPT…OpenAI is playing catch up with GPT-5…but I think the different models will continue to leapfrog with one another for the foreseeable future.

And I think that’s also the net benefit. In a lot of the AI tools that we leverage, we leverage the ability to interchange the models.

Brandon Frampton:

Great. And then, the last part was, “How long do you think it would take to build this out?”

Steve Brask:

So, from an Upshop perspective, from receiving data into Upshop, and granted, we have the infrastructure, we can go live on forecasting within about 14 days of receiving the data. So typically, what we do is ingest the data into our model. We run a test essentially, where we generate forecasts and evaluate against sales for a period of about 14 days, and that’s really to ensure that there isn’t any model hallucination. Typically, there isn’t. Typically, we’re running just fine, but we do encounter opportunities sometimes where we have to go back and say, “Oh, this payload failed in the data,” and so we just need a quick reload, and we close the gap and we’re fine. But we take that two-week window to evaluate model accuracy, and then we turn it on live, and we’re running.

Brandon Frampton:

Excellent.

Steve Brask:

But in terms of actually building the model out, I think it’s fairly easy to stand up an ML (Machine Learning) model for demand forecasting. I think you can build it in the space of weeks. It’s refining the model that takes probably about nine months to get to a point where you will have all of the inputs that you would expect: promotions, pricing, the sales, all of the other components that I would recommend from a demand planning standpoint. Refining all of those takes six to nine months once you have the initial model in place.

Brandon Frampton:

Great. Very helpful, Steve. Thank you so much.

Mike Weber:

And Brandon, one build, because we’ve been talking about this. I mean, I do think there’s something. You can go get a data scientist, there’s probably a cost to that. Knowing what the market looks like right now, however you could also get people that are really getting good about prompting the different AI tools.

And frankly, we find it very valuable when you bring somebody in that doesn’t do the data science, but they really understand the way that you can work and prompt AI to give them that something they’ve never done like, “I want to look at this workflow differently. How would we change the UI?” In AI, a person who has a lot of familiarity with prompting can come in and do something in a very short amount of time with the said tool on that task in a probably more effective way than somebody that’s been designing that UI.

Brandon Frampton:

Right.

Mike Weber:

And that is what the industry is finding, that AI prompting, just in the ability to build custom agents, or to shape prompts in different ways, and explore how you could prompt with different tools and connect them, there’s a lot of value there, and that’s probably not going to have the same cost as going out and getting its said data scientists. So it’s something to consider as you look at the team structure.

Richard Poye:

Yes. It makes me think about Stephanie’s comment. You have employees that have been around a decade plus who understand the customer. If you could train them how to do the prompts and ask the questions, engagement with the technology would be better. They’re

more insightful because they understand and they could evolve together. I could totally understand Stephanie’s point of view of: “I’m trying to implement this program.” People feel like they’re pushed in there versus becoming a partner with a technology and being able to identify ways in which the technology could help them understand because of a weather event, a school event, those types of things.

One of the things we haven’t talked about is competitive intrusion. When a competitor comes in and builds a store in your area, kind of the impact of that. It’s so hard for the business that’s being intruded on to manage and change their forecasting, and AI would be really helpful within that environment because it would be able to make the adjustments, and also you’d be able to see where the competitor is winning against your product, and then you could shift your pricing, or your promotion, or planning, whatever it might be, to adjust to that. So there’s a lot of things, I think, within it, but yes, to the point, prompting is probably one of the better investments because you’re getting people to engage in it, and also teaching people about prompting is not complex because they’re able to ask baseline questions that are relevant to their business and they can contextualize it, and you can get somebody to really dive in and kind of engage with it.

Yes, Eva?

Eva Strasburger:

Training And to back Stephanie again, if you have someone who’s switched on in your team, and interested in such things, there are some great courses, which are not very long on how to become prompters. That’s becoming a career in its own right now.

Stephanie Galentine:

Yes. This morning, the AI presentation I was watching was really around marketing. So not totally related data-wise, but it was amazing. He did all of it live, so he was training us live. It was amazing to watch what words he used to ensure that he got what he wanted out of it.

I mean, I feel like it could take me a decade to do that, but the training is fundamental. There’s no way you can get good at this without someone walking alongside you and getting you where you need to be. So, I will be looking for training opportunities, for sure.

Steve Brask:

On the flip side of that, the benefit to some of these models is the model learns from your input too. And so we talk about the change management component, these folks that have been in the business 25, 30 years, in some cases. They’ve done it the same way their entire careers. The model learns from their questions, and it learns how to respond back to them in a way that’s meaningful to them. And I think that’s one of the most powerful aspects; it helps to win over some of those folks a little bit more readily because it’s learning from their own line of questioning. I encounter a lot of folks who are data skeptics. We’re implementing a demand forecasting tool to retailers, and we have these folks that have been with the business 30 years, and they’ve done it this way their whole lives. And so I think the models do a very good job of learning how to respond to the skeptics, just the way people do. I think that’s a skill set that I’ve developed, and I think that models do a good job of that too, learning how to respond to data skeptics and how to work with them.

Stephanie Galentine:

I agree. It was very impressive. They were very encouraging. Each response was like, “Great job. Let’s keep going.” So I literally saw it for the first time ever this morning and I was stunned. Very interesting.

Richard Poye:

So, I’m interested, without people really coming out and saying exactly how, where people are at or where do they see the challenges within their organizations. And Stephanie, you’ve come up and said, “Hey, look, we’re starting here and we’re kind of at a baseline and we’ve been doing it manually.” How are other people seeing where they’re at? And what’s their challenge? Because I think you’ve demonstrated like, “Hey, if you’re getting into it, there’s a learning curve.” But some people may have been coming down this whole process, and where does everybody see themselves within it? Maybe on a personal level and then on a professional level and a retail level. That might be a way to talk about it. Anybody? Kris.

Kris Klinger:

Workforce Yes. I may dance around your question, but I think it’s interesting because the adoption or the acceptance of AI by the managers, some of the things that Stephanie was saying and Steve and Mike were talking about, there’s this adage of “having on blinders.” You get these blinders and then you know what you know, but you don’t know what you don’t know. You all were talking earlier about asking good questions and sometimes the people that have been doing it so long don’t know those or don’t remember what those questions were anymore because they’re on auto- pilot.

I’m just pointing out two things to overcome. That’s one of them because folks are so used to doing the things that they’re doing, the way that they’re doing, using the systems that they were using for 10, 20 or 30 years, it’s difficult sometimes to get the adoption or the support that you need, particularly from the operations folks. And then level that down to the hourly folks who are supporting it. And they are afraid for a couple of reasons. One is they don’t understand the technology and what it does and how it works, but two is that it’s going to take their jobs away.

I’m just pointing those two out as major concerns. When we are rolling out the technologies and educating folks on the benefits of the technologies and trying to get them to adopt and adapt accordingly, there has been resistance that we’ve experienced in a number of different ways, in a number of different areas on that, and it takes a bit longer sometimes. I mean, the group here on the screen, we’re ready to go, well, as fast as we can within reason, but that’s not really the norm yet from what I’ve seen, and I work in higher ed.

Richard Poye:

And Bonnie, what kind of challenges are you facing?

Bonnie Zaring:

Yes. I guess I would say a challenge that we experience is there are business experts who understand the complexity of the business they’re in and then there are data and analytic experts who are very well-versed in synthesizing and understanding data and understanding even how these models work…what to put into it to get something else out of it. And I find the two sometimes are not in sync or we’re working more in silos and the individuals that are building and trying to develop tools and technology come back with something that then the business experts either don’t understand, how to apply it, or they feel that they’ve missed some natural complexity that’s in the business and say, “Oh, you didn’t deliver what I could use.” I think bringing those two together in a way where they’re really understanding important data points, “These are non-negotiables. For it to work, I have to be able to have this.”

It’s not the buy-in. I do believe everyone wants to be there, wants to get there, but they also want to open the box and it’s a tool that automatically anticipates what they need in order to manage their business. And it’s not always that easy to bring the two together. So there’s inventory management. There’s financial information. There’s production and how to bring maybe an ingredient into builds and, yes, a one-to-many relationship. There are all of these different parts that matter and pulling it all together to be able to synthesize what’s important to get a usable output, I think that is a complexity in the change management of where we stand and where we want to be.

Richard Poye:

Yes. I think that’s great input because each one of the businesses is going to have their own unique silos. And how do you create the framework to build that more harmonious stage building? Sometimes one is so far ahead of the other that something gets built and you’re like, “I didn’t need that,” or you feel out of the loop when something gets built or it creates tension. So it might be interesting to go and identify best practices at each one of these stages when you build out those quadrants, what’s needed before you move to the next one so that everybody moves along and understands some of the challenges.

Brandon, you’re out on the West Coast. You’re seeing different things. And you’ve worked on the East Coast. You’ve worked in Texas. You’ve worked across the country. Are you seeing that people are implementing technology more out there? Because that’s the assumption, especially the part of California you’re in. And how do you think that might be impacting convenience?

Brandon Frampton:

For everybody that doesn’t know, I’m right in the Bay Area, right in the heart of it. We’ve got stores all around it and all through California up in the Bay Area and down in LA. But I would tell you, it’s not like maybe I thought it was going to be here, but it’s far advanced from the rest of the country. And we have a lot of people who are very interested in AI. We’ve probably had, I don’t know, seven, eight big firms come to try to help us. My take on it is it’s going to change this industry more than anything ever has. We’ve had a lot of different things come and go, but we feel that it’s going to change every aspect, not just food, but how we train, how we track things and everything.

StrategySo we’re trying to lead with that. And again, it’s $20 just to hire one person for one hour and it goes up from there, so reducing labor, et cetera. We think it’s going to be huge. We’re going to pick a partner to go all in with. And then personally, I’m going all in as well. I think it’s going to change everybody’s life. So that’s just my thought.

Richard Poye:

Okay. I mean, are people using it day to day?

Brandon Frampton:

Oh, yes. Everybody here is using AI throughout the day for different reasons, from creating POP, of course there’s a lot of writing, different writing tools, and cleaning that up. But certainly, everyone on my team uses it. People throughout the industry, or not the industry, but throughout the area, are already way early adopters. Yes, sir.

Richard Poye:

And then Kris, you’re in an area…Boston is known for having lots of universities, lots of technology, lots of innovation. When you step out of Boston, are you seeing how the technologies might be implemented differently? Because when you go to western Massachusetts, I think it might be different than in the Boston area.

Kris Klinger:

I mean, I’m looking across the river. MIT is right across the river from me and Harvard’s right behind it. And I mean, they’re obviously leading. MIT in particular, one of the institutions leading the charge. We set up an AI task force a bit ago, one faculty AI task force and an administrative AI task force. And then they focused on their areas and then once they finished up what they were tasked to do, they merged. And we’re actually implementing quite a few [initiatives]. We have a number of AI initiatives on the academic side in particular because obviously it’s not going anywhere and for instructors or professors to think that students are not going to use AI and so on [is naive].

So we dove in both feet first. And we have a data science building actually we just built, it looks like a Jenga building, it was built a couple of years ago, that houses our data science folks that are doing a lot of that technology and a lot of that research and they’re trying to find ways to apply it. The interesting thing though is in higher ed, you would think that that would come to us first. It does and it goes out to the real world where they can make a lot more money than doing it with or for us. But we’re always looking at how we can implement it. And my scope also includes housing and convenience stores and bookstores and transportation and parking and some different things.

We’re looking at how we can apply it to those different areas and then also just to help track the students and support the students better. But yes, there’s a lot of it going on and there’s a number of folks in downtown Boston as well that are specializing in it. And it’s exciting to be part of it. And I agree with Brandon. I’m all in. At times, it can get a little overwhelming because there is just so much to learn. I’ve taken some classes and online classes and so on. I mean, we could spend your whole day learning that stuff and you would still barely make a dent. But yes, we see a lot of it, Richard.

Richard Poye:

I took a course at MIT a couple of years ago and one of the things talked about was writing prompts and they said treat the ChatGPT or whatever format you’re going to be using like an intern. Be able to give them solid advice. Give them the information that they need to perform their job. Give them very specific details, context and what the objective is and so that it can think. And so I think as we try to intertwine the partner that’s going to lay on top of some of these solutions here, I think that might be a way to think about it. They said treat the thing well too. Don’t be mean. Okay. So I thought that was funny.

Kris Klinger:

Well, it’s interesting, Richard. Your point is we grew up, most of the folks on this screen, using sentences and speaking clearly. Now this younger generation, I mean, that’s a challenge for them to do the proper prompts because they’re so used to communicating in a certain way. And it’s funny because I’ve seen some tests on it where AI, they’ve trained it, but then it’s trained to one user… a different generational user. Thank you, I could go on.

Richard Poye:

I think it’s interesting because if you think about Stephanie’s stores or stores that Bonnie has… Bonnie has stores across a number of states. And so even the way that people would talk or ask a question could vary from a regional perspective. And so there’s lots of stuff there. All right. Does anybody have any additional questions? Mike?

Mike Weber:

I was just going to make the comment to build on what Bonnie said and what Brandon said about AI playing a role across multiple functions. The differentiator will come in when you have a single data lake and then you have AI connected across all the actual workflows. And I think it’s easy when you can pick one thing to do with AI, but that one thing could influence so many other things. And there’s still a lot disconnected. I think Stephanie mentioned labeling, right? We’ve got labeling over here and you have recipe things. You have price book. I think one of the mechanics just standing in the way here is that effort, which is harder for the medium, small to medium. That’s going to be their challenge is how do you get all this stuff with one data language? Because when it’s we can all talk to each other, then yes, then you can turn on the Skynet and start to see what you could really do with the company.

Richard Poye:

And just for the people who would read this document later on, we’ll give a little description on data lake. We’ll put that into the piece here because it is important.

(Editor’s Note: A data lake is a centralized repository that stores vast amounts of raw data in its native format. Unlike a data warehouse, which typically stores structured and formatted data, a data lake can handle structured, semi-structured and unstructured data. This allows organizations to store data as it is, without needing to structure it first, making it flexible for various analytical and machine learning purposes.)

Alright, I’m going to move on to the next piece. Where we left the last call, we talked about doing something potentially around GLP-1. And I think it’s a very hot topic. I’m getting ready to speak at two different conferences in September. And for both conferences people have asked about GLP-1 beforehand and being able to speak to it, even though they’re not foodservice conferences. They’re more centrist or more leadership conferences. So I’m interested in making sure we’re still on to have that conversation. And also within that, doing a little bit of learning from what some of the other Vision Groups are doing, and looking at having people come in that are part of the group and actually maybe have a 10-minute presentation or talk about what they’re doing, what they’re experiencing, what they’re seeing in the marketplace. And so if anybody would be interested in engaging in that, whether it’s a talk, a presentation or whatever, and be one of the people that would do that in the next meeting let me know.

That’s something that we’re looking at and just wanted to make sure we’re all still aligned on that piece. And I think it’ll be fascinating to look at it regionally and understand how people are being impacted by it. And if you haven’t been impacted by it, this would be a great learning session to understand how that’s going. Might be interesting things, like McLane may even be noticing that certain items are dropping off. It’s just not about just the drug and the treatment. It might be just the ramifications of so many people engaging in this and how it’s impacting purchase habits and eating habits. So we’ll go forward with that.

And then I want to introduce Robert, I didn’t really give you a chance to introduce yourself. Maybe do a quick introduction and then he’s going to talk about the Summit that’s coming up and give you a little bit more insight on what’s been done and what the Summit will be looking like. So, Robert.

Robert Hampton:

Great. Thank you, Richard. Good afternoon, everyone. As you might remember from our last meeting, my name is Robert Hampton. I’m a consultant within the retail and convenience space. I’ve been working with my colleagues here at Vision Group Network to develop some exciting new programs. By now, you should have received the invite for our very first Vision Group Network Virtual Global Summithttps://vgnsharing.com/vgn-summit/. And at this summit, we’ll bring together the members and supporters from all seven of the Vision Groups on November 19th at 11:00 AM Eastern Time.

You heard Richard talk a little bit about some of the other groups. Well, this is the opportunity to hear what they’ve been working on and hear the report outs from them. We expect to have about 100 participants in this virtual gathering, and we’ve already had about 80 members accept the invitation. It’s been very well received. We do have an exciting agenda, including Gerd Leonhard (founder/CEO of The Future’s Agency) as our keynote speaker, as well as reports, as I mentioned, from all the Vision Groups from all over the world. You’ll be able to hear findings and insights from your peers and colleagues from the EV, Global Convenience, and other groups. And this will be a three-hour discussion. We also have a section on the VGN website that will be updated with more details, and we’ll be sending out an update by the end of the month as well. And again, that’s November 19th, starting at 11:00 AM Eastern Time, going to about 2:00 PM Eastern Time. So, thank you, Richard, and back to you.

Richard Poye:

And to that point, I know that these meetings are a time commitment and if for some reason you may not be able to make a meeting let’s talk about how a guest from your group or from your company could join. Or if you look at the agenda, and you think someone else from your team might contribute and benefit from joining let’s have that conversation also. Myra, you have something?

Myra Kressner:

I just wanted to remind everyone that we will have the CFVG Vision Report for the industry in a few weeks. We started a few months ago delivering the Vision Reports in FlippingBooks, which is interactive, so it’s very searchable by keyword or names and easy for folks to really hone in on either a subject or an individual that you just want to pay close attention to, as well as having the video recording of Mike’s and Steve’s presentation. So we think that we’ve made this much more user-friendly. And again, I want to remind everyone, please do share the Vision Report with your colleagues or with other industry folks that you think would be interested, either your state associations or others that are going to get value from the Vision Report.

And of course, I want to say, Mike and Steve, thank you. Excellent presentation. Thank you for dedicating the time you put into this, as well as being our Ally Supporter and John Cox with McLane as well.

And lastly, we’re not necessarily planning on having a formal get together at NACS, but let us know if you will be attending and we may be able to bring folks together in an informal way. Over the years VGN members have said, “We like seeing everyone’s face on the screen, but if we can actually be together in person, we’d love that as well.” So do be sure to let us know if you’ll be at NACS. Sorry, Richard. I didn’t mean to interrupt your comments but wanted to make sure that I got that in before folks took off.

Richard Poye:

Okay. Totally appreciate that. I think everybody likes to see each other human to human sometimes versus just screen to screen, especially since we’re across such a wide geographic area. Anyway, I think we’ve covered a lot, and I think if there’s additional follow-ups that anybody has, feel free to share. And I think one of the things we go back is that we’re trying to really help the industry overall and I think for somebody who’s just coming into this it’ll be helpful. So, I think we did our jobs in that sense. And I think there will be people who will find value in this and will probably find the topic less intimidating since they’ve heard peers speak about it. So, thank you very much.

Myra Kressner:

Thanks, Mike.

Mike Weber:

Yes. Thanks for having us. And it was a great discussion. I just dropped in a fun one that I’m just personally playing around with. It’s pretty crazy. You can turn your agents that you create in GPT into a person. And if you talk enough to it, it actually just takes on you. I sort of shut it down. I got freaked out about it, but now I’m getting sucked back in because I want to see how far it can go. So hopefully, let’s see. Hopefully, if a Mike Weber reaches out to you, don’t answer the call. Don’t do it. Because AI has taken over. Anyway. It’s great discussion.

(Editor’s Note: this is what Mike put in chat and referenced above: if you want to get “freaked” out by AI – The OS for Human-AI Interaction – this is one. You can translate an agent into a virtual person or use your likeness to act as the agent. Pretty nuts. Tavus is a human computing research lab building a new kind of intelligence: interactive AI Humans that see, listen, adapt, and act with emotional awareness.)

Roy Strasburger:

And Mike, thanks. Thank you to Upshop for being an Ally Supporter as well. So thank you very much for all the help. And also, thanks to McLane for being an Ally Supporter.