Structure-First AI for Retail: How Albertsons Builds Merchant-Trusted Systems

Summary

Albertsons is a holding company with 22 grocery brands including Jewel, Shaw's, and Tom Thumb with over 50 million loyal shoppers across 2,200 stores, and built a structure-first AI architecture on Databricks Lakehouse that models promotion and product relationships as graph structures rather than passing raw problems directly to large language models.
The system learns from 200 billion rows of transaction data using graph neural networks for relationship discovery, deep learning for predictions, mixed integer quadratic programming for deterministic optimization, and LLMs for merchant-friendly explanations — delivering recommendations in under 30 seconds.
The structure-first approach prioritizes explainability and merchant trust: LLMs are used specifically to translate optimization outputs into language merchants understand, enabling retailers to act on AI recommendations rather than receiving opaque predictions from a black-box system.

Structure-First AI for Retail: How Albertsons Builds Merchant-Trusted Systems

Retail promotion decisions drive billions in revenue, yet most AI systems fail to explain their recommendations to merchants. Albertsons built a structure-first AI architecture on Databricks Lakehouse that models promotion and product relationships as graph structures, enabling merchants to understand and trust the system's predictions.
Using 200 billion rows of transaction data, the system combines graph neural networks for relationship discovery, deep learning for predictions, mixed integer quadratic programming for deterministic optimization, and large language models for merchant-friendly explanations. This architecture delivers accurate, explainable, and actionable recommendations in under 30 seconds at scale across 2,200 stores.
🤝

Chapters

FAQs

What does structure-first AI mean at Albertsons?

Structure-first AI means discovering the underlying relationships in data before applying predictive models, rather than presenting a raw problem directly to a large language model and asking it to predict an outcome. At Albertsons, this means modeling promotion and product relationships as graph structures first, which enables more accurate and explainable predictions than unstructured LLM approaches.

How does Albertsons use graph neural networks for retail promotion decisions?

Graph neural networks are used to model the complex relationships between products, promotions, and customer behavior extracted from 200 billion rows of transaction data. These learned graph structures capture how products interact during promotions — including cannibalization effects — in ways that traditional tabular models cannot represent.

Why is explainability important for AI-driven promotion recommendations in retail?

Promotion decisions drive significant revenue, and merchants need to understand the reasoning behind recommendations to trust and act on them. Albertsons uses LLMs specifically to translate complex model and optimization outputs into merchant-friendly language, making results actionable for non-technical retail buyers rather than presenting opaque scores.

How fast does the Albertsons structure-first AI system deliver recommendations at scale?

The system delivers accurate, explainable, and actionable promotion recommendations in under 30 seconds at scale across Albertsons' 2,200 stores. This combines graph neural network inference, mixed integer quadratic programming for optimization, and LLM-based explanation generation running on the Databricks Lakehouse.

Full transcript

[00:09] Okay, so topic is unusual. Molecules, merchants, and models. Molecules are from chemistry, obviously. Models are from data science, and merchants are from retail. So, what do these things have in common? Phenomenally, they have a lot of things in common. I am working on my doctorate on uh
[00:25] molecular biology. I focus on predictive learning using world models on how toxicity or chemical toxicity can be predicted with more precision for drug discovery. So, I am not a chemist. So, how many of you are chemists in the room or
[00:41] pharmacists in the room? So, I'll shy away from saying anything that uh pulls my credibility down if that's the case. How many of you are from retail, by the way? Okay, fair few. How many are you from grocery, by the way?
[00:56] Okay, select few. Okay. All right. So, I know the composition of the room of uh how I should introduce merchandising and merchants in that case. Now, I say structure first. So, this is a conference about data and AI. So, I'm going to go a little deep on data and AI, if you don't mind. So, what
[01:13] do I mean by structure first? So, today's science of AI is mostly about throwing an LLM at a problem and saying predict an outcome. So, that's the science most of us are talking today about. You throw an agent
[01:28] at a problem and you hope that it does a better job at an outcome. When I say structure first, it's the inverse. We are not throwing LLM at a problem and saying, "Can you predict an outcome?" We need to discover the structure first
[01:43] before we talk about it. But, before I go into the topic of structure first, just a little bit about Albertsons. So, how many of you know of Albertsons, heard of Albertsons, what Albertsons is? Okay, only a few people in the room and
[01:59] feels like most of the most of the team is also from my my team. So, they obviously know it. But, we are a family of brands, by the way. Albertsons is not one brand. We are a holding company called Albertsons, but we have 22 brands. So, in Chicago, people know us as Jewel.
[02:16] In In Boston, people know that know of us as Shaw's. In Texas, people know of us as United or Tom Thumb. In the Midwest, they know us as Albertsons. So, so on and on. But, we are a family of brands. That's how our customers know us. So, we are a family of brands. We
[02:32] have 50-plus million loyal shoppers. So, we are a family of brands. We are not one brand. We are a family of brands. Now, I talk about this topic in the context of retail, but you wind it back into the basics of marketing. What is
[02:47] basis of marketing? You probably have heard the four P's. Price, promotion, placement, and product. You have the right product at the right price and promotion, in the right stores, in the right locations,
[03:03] you have better customers. That's the essence of retail. You know your customers best by appealing to their price segments so that you can attract them by having a differentiated product alongside it. But, what happens in merchandising world
[03:19] or retail world is these domains are not connected at all. Pricing is a separate system. Promotion is a separate system. Assortment and space, which is placement, is a separate system. Product management or item management, my peer on item management is sitting here, so
[03:35] I'll be careful what I say. Um They are different systems. But, if you think of a merchant who is managing a P&L of a category, like salty snacks, like soda, like chocolates, like beverages, they don't think in
[03:50] isolation. They don't think, "Today, I'm going to go to work and I'm going to work on pricing. Tomorrow I will work on promotions. Day after tomorrow I will figure out what items I put on the shelf or take out from the shelf. Merchants are always balancing this
[04:06] portfolio and I have a merchant right up here. John, he's from Jewel. He's a lead merchant there at Jewel. Um they trade in these four dimensions. What items should I carry at what price and what promotion and at what cost and
[04:22] allowances is it profitable enough to attract our customers enough? But the science behind it, you've heard of plenty of solutions, Blue Yonder, McKinsey, etc., etc. They've all built narrow systems. They are locally optimized but globally blind.
[04:40] So if you ask a question of what happens when I promote Sargento cheese to my own brands solution, Sargento cheese is a shredded cheese. Most of you probably know what Sargento cheese. So if I promote Sargento cheese, what happens to
[04:55] own brands? How much space need to allocate to Sargento cheese at $2.99 in retail and $1.99 in promo cost, what what space do I need to allocate to it? No system can answer it today.
[05:11] There is no system today that is able to take a combined economics of cost, price, promotion, vendor funding or allowances and customer behavior and say what does it mean? So that's what we building at we are trying to build this new playbook
[05:27] of retail. How do you connect the four B's holistically? So we can make merchant decisions in this in this typical merchant behavior. We are making the science behave like a merchant twin. So when you when you hear Gartner say
[05:43] digital twin, that's what we talk about. How can we take a merchant's encoded knowledge in learned representations and translate them into machine behavior? So, that's what we are trying to build. That's what I mean by structure first AI. And to give you an analogy on
[06:00] chemistry, why does it make sense when I talk about molecules? Now, how many people have eaten bananas? Almost everybody has eaten bananas. How many of you know that there is one atom in the banana that gives it its smell?
[06:16] You pull out one atom from the banana, the banana loses its smell. How many of you use mosquito repellents? How many of you use lice shampoo? There is only one molecule, the phenyl molecule in mosquito repellent. If you
[06:32] turn it off, that repellent is no longer toxic. You can consume it in large portions, I've heard, and you still won't become ill. For the mosquito, it's a different question, but for a human, it's a different question. So, that's what we are trying to predict. If I do not
[06:48] promote Pepsi on a certain day, what does it mean for Coke? How will it lift Coke up? Most of you know Super Bowl day. I came to know from experience that Pepsi dominates Super Bowl. So, Coke does not promote on Super Bowl.
[07:04] So, merchant's encoded knowledge is I do not want to promote Coke on a Super Bowl week because we know that Pepsi's going to Pepsi's, I think, here, so I got to be careful what I say about Pepsi. Um so, we know that promoting Pepsi on Super Bowl day will give us the maximum
[07:21] leverage. We also know we promote Lay's chips on a certain Memorial Day, we get a different answer. We also know promoting turkey on a Thanksgiving Day, natural intuition will give us a separate different answer. But, if you have 30,000 items in a store, how do you know the chemistry
[07:37] between all of them? Pulling one, what will it do to another? That is the knowledge that we are trying to build. I'll cover this in four parts. The four parts, the first one is learning this
[07:52] behavior that I spoke about. How do I learn this behavior of these four P's connected together? And today I'll focus mostly on promotions. I'll not touch on the pricing or placement side of things. I'll focus mostly on the promotion side. The second is once that learned behavior is understood, how does Coke and Pepsi
[08:10] pull each other apart? When you promote one, what will it do to sales for another? You have to predict You have to be able to use that knowledge to make predictions. You have to make those predictions, whether it's unit predictions or revenue predictions or sales predictions or household growth
[08:25] predictions, what will promoting or not promoting a certain item do to my lift in units, revenue, AGP? How much more customers I'll gain? What do I get out of it? Now, predictions are useful, but if it's not explainable,
[08:40] they're useless. So, in this field of AI, we have something called XAI. It is not the company XAI run by Musk. XAI is actually a discipline in AI called explainable AI. So, how can we use explainable AI to translate predictions to answers? For
[08:57] most of you who come from AI world, you know integrate integrated in gradients. If you have worked with deep learning models, you probably know of Shapley values and graphs if you work with Shapley values. No merchant can understand a Shapley graph. No merchant can understand the violin plot.
[09:13] So, you need to make a translatable merchant language-enabled explanation. So, that is a science by itself. When you talk to a merchant, it's almost like a radiologist talking to a doctor. They talk about ischemic stroke attacks measured in 1.8 cm in diameter. When we
[09:30] talk to the merchants here, we are talking about VLC and DNC, and what is the impact on DNC on allowances, and what happens to promo retail, etc. So, there is a language that we need the machine to learn and the LLM to learn on what does it learn What does it take to
[09:46] translate these predictions into merchant language? That's my part three. And the last part is if I make predictions without actionability, it's of no use. Insight without action is of no use. If I just told you that the world is going to end tomorrow, but I don't give you a plan of how you
[10:02] can escape it, that is of no value. So, the last bit is the actionability. And that's where we'll use the agentic stuff a lot more on how we use agents only where it makes sense. Not the learned representation problem, not the prediction problem, not the
[10:19] explainability problem, but only the ability problem. That's my part. Okay, so I talked about this no clinician adopts it, so I'll make this real by showing it to you. This is a real MRI report. This is a real MRI report going to a doctor. I won't name the patient for the
[10:35] sense of PHI purposes, but you see the the radiology report. It says what did I do, which is the Firstly, it talks about what Who is the patient? Next, it talks about what is the technique they applied in data science world, the model we applied.
[10:52] And what is the finding from the report? And then, what is the conclusion behind it compared to the past period compared to the past period? When a radiologist produced this report to a doctor, it tells you in underline very clearly 1.8 cm in the right
[11:09] cerebellum. It basically means there was a prior stroke the patient had on the brain. And then read the world, stable minimal, meaning it's not changed. So, he's not regressed, he's not
[11:25] progressed, it's latent. Latent, it is sitting dormant. Nothing has happened adverse to the patient. But can you imagine showing this to a patient? No patient will understand what this means. So, when I say explainable AI, there is
[11:41] different ways of explaining the AI. One, you need to say a story to the doctor, which is very different than saying a story to a patient. When we explain model predictions, we are talking about explaining that level to a lead merchant. When we are talking about
[11:58] the common mass, we have 1,000 merchants in our company, we are talking about simple English predictions. And I will show you an example of that, too. So, that's what I mean by auditability and trust behind it. Most of you have heard the word hallucinations over and
[12:13] over and over again. If you throw an LLM at a problem and you try to constrain it by using a few guardrails, the trust will not go away. Hallucination will remain. So, the unique thing that we are doing with Databricks is how you build this foundational layer all the way from
[12:30] learning structure to predictions to explainability to actionability in one ecosystem. We don't want to go off one ecosystem because developer productivity is critical for us. We don't want to take a best-of-breed one from here, one from here, one from here, one from here,
[12:46] and try to piece it all together and say to an operations team who is handling 200 billion rows of real data on a daily basis to say, "Go piece it together between a llama a llama solution or a portkey solution or a AI gateway somewhere
[13:04] else." We want the unified infrastructure to sit in one place. That's how we are using Databricks. Okay, so three pillars I'm going to talk about. So, this is the part two three two three four that I'm going to talk about. So, the the part two is about how do we make predictions.
[13:19] Learning from structure. The part three is about how do we do explainable AI. And the part four is about how do I use the agentic infrastructure to drive these conversations.
[13:36] So, what that means is what we built. It's a sensing layer. We have a sensing layer that is running on 200 billion rows. It's looking for data on a daily basis. So, we teach the model on a three-year history of all our data, which is roughly about 200 billion rows. I think it's about
[13:52] 10 terabytes of data, something like that, not the petabytes or zillions that Ali talked about yesterday. So, it's a contained sizeable model. It's a sizeable model if you're running a foundational model in data science. Then comes the prediction layer. And the
[14:07] prediction layer is nothing but a neural network that we have. It's a two-headed uh neural network model. It's a deep learning model, which we call the foundational model. It's because it's learned based on our data. Then it's the decide layer. What promotions or what
[14:23] recommendations do we want to produce to the merchants? Then we have the explain layer. Then we have an act layer. And it's a loop. Every day it's a loop. Something changed in sales yesterday. What is the signal? What do we need to predict? What do we need to explain? What do we need to
[14:39] decide? What do we need to act? And it's a loop. It's a continuous loop that happens on a daily basis. So, that's what we are building. Now, reliable predictive AI. So, uh as I told you before, you have to be able to explain it. If I pull Coke out of the
[14:56] promotion, what does it mean? What does it do to Pepsi? How do we learn these learned representations? Now, most of you have heard of graph neural networks. How many of you working with graph neural networks here?
[15:11] A few. But, what does it mean for visualizing a graph neural network? So, when you look at chemistry, I told you about the banana example or the mosquito example. You pull out one atom in a molecule, you can change the behavior of the molecule. But if you look at a distributed motive,
[15:27] it's a collective decision. You pulling out one does not change the impact of the other. You need to understand the dimension of it dimensions of it in totality. You know, when you have a 12 oz Coke and a 12 oz Pepsi, you need to understand all the
[15:43] conditions that's driving behind it. There are 104 variables that go behind it. So, if I visualize the graph, that's what our graph looks like. This is what graph neural network means for a merchant. This is not how you describe in edges and nodes
[15:58] to a data scientist. This is how we explaining this to a merchant. The blue circles are essentially complements. The orange circles are essentially subs. So, the size of the circle is saying how strong the relationship is. And we
[16:15] defend this by saying this many transactions are linked to this bubble. So, if you buy Coke Classic 12 oz, there is a strong likelihood you will buy a Sprite 12 oz pack. That's what it's
[16:30] saying. A 12 pack 12 oz is strongly linked. If you promote them together, they will sell together. If you promote Coke Classic 20 oz at the same time, you will lose volume. So, you will lose 2.8% in volume. That's what
[16:48] the graph is saying. So, this is the graph that we try to decipher with the merchants on a given category to say, "If you're looking at carbonated soda drinks, how do you turn a graph into a merchant explainable outcome so that they can decide which items work
[17:03] together, which items don't work together?" So, that's essentially the learned representations that I'm talking about because you have to learn these relationship between items. And when I talk about the scale, and I don't want
[17:19] to give away a full data set, this is only a subset of our data set. This is just on one division for us. This is just on one division. This is the quadratic computations that we need to run. So, think about the GPU infrastructure to you you need to run 450 million quadratic computations on a
[17:36] daily basis. So, that's where we're running the model infrastructure on Databricks, so we can measure not just the performance of the model, but we can also measure the accuracy of what it's doing on a daily basis. If I change the result on a daily
[17:52] basis to a merchant, they will get confused. What are you talking about? These things are learned over time, and they rarely shift over time. So, we predict drift and accuracy on a daily basis to be able to say, "Has my molecule structure changed a lot?" If
[18:08] it's changed a lot, we give it a negative bias. If it's changing positively, that's also bad, we give a positive bias, and then we pull it back to say it needs to predict the learned representations over time. Don't drift. You can drift during season. You can drift during
[18:24] Thanksgiving Day. You can drift during Super Bowl Day, but you can't drift. So, that's what I'm talking about when I say learned representations. How do you take a graph of a equation, and you do this? Now, in a deep learning network, we have sub networks. For most of you
[18:40] who are running deep learning models, you know sub networks. We have a price sub network. We have a promotion sub network. We have a store sub network. We have an item sub network. And a multi-layer perceptron brings all of these networks together to say, "What
[18:55] would be my units if I did this with Coke, and if I did this with Pepsi on the same week?" So, that's what our prediction layer does in being able to predict what would be the outcome of promoting something or not promoting something.
[19:10] Now, for those of you who are working on LLMs, how many are working on attention, by the way? Anybody working on LLMs heard the word attention? Anybody read the paper attention? So, in retail, we can't apply attention
[19:26] as it is. You can't take a language model. A language model is fluent in language. For those of you who know, just because you're a linguist and you present on TV, it does not make you a weather specialist, does not make you a medical doctor, does not make you a
[19:41] chemist. They're good in fluency. Large language models are good in fluency. They need to be taught the mechanics of retail, the QKV of attention. Most of For those of you working on attention, you probably know
[19:56] what QKV means. How do we teach it the QKV for retail? That's what we are doing with our LLMs. We are teaching the QKV of retail because we can't train a large large language model in the way we want to train it. We are working with Open
[20:12] AI. Open AI is not ready to fine-tune their models with the size of our data set. We are working with Google. They're also on the same boat. Whether it's Gemini, whether it's XAI, whether it's Open AI, they're all proprietary model turned on world knowledge. They are not learning enterprise knowledge.
[20:29] So, that is why it required us to build our deep learning network, teach the QKV of retail in attention blocks, so they can attend to each other. When how is item linked to price, linked to promotion, linked to space? That's the attention block that we are teaching
[20:45] it. The QKV mechanics of attention is what we are building in the deep learning network. I said this is about data and AI. It will be deep on AI, so bear with me. Okay, explainable AI. So, how do we turn these predictions that I spoke about, the predictive part in the deterministic
[21:02] part? How many of you would have mixed integer programming? Many of you do mixed integer programming? 1 2 3 Few Few here. But, we can't turn these deep learning models LLMs by their very nature use deep language models and models in the back.
[21:19] They're not explainable. An output from a large language model is not explainable. There is no confirmation bias that it can give the scoring around to say what is the confirmation bias in your hypothesis. You've You've heard the joke, you know,
[21:35] predict a number between 1 and 100, it's almost always 43 or 73, you know, whether it's Hitchhiker's Galaxy or Sheldon's Sheldon's number that it come up. So, it is only as good as it's learning. So, we cannot make the recommendations
[21:51] coming out of a deep neural network explainable unless you apply the guardrails on top of it and you try to explain what the model is trying to say. So, we use MIQP, mixed integer quadratic programming because we need to learn the
[22:06] relationship between all of them. And I told you the equation of X over Y, what would that relationship looks like? So, that's what we do in our guardrail layer. It is a huge MIQP layer. It's an optimization that says which promotions
[22:22] that I need to run for my merchant. For Coke Zero, what promotions do I need to run? If I have five promotions, what is the score for each one of the promotions? If I do promotion A, how much units will I gain? How much loss will I make? If I do promotion B, will
[22:37] it have the opposite effect? Will I attract more customers by doing promotion three, but lose on units and sales? It's about household growth. So, what is the right mix of the equation? That's our deterministic AI layer. And the LLM is simply a translation
[22:53] layer. It is our linguist. It is our linguist to translate a graph into a model. It is our linguist to turn an MIQP recommendation into a simple interpretation in English. What does that look like? That's what it looks like. So, you see on the left, that's the
[23:09] explanation that we have for a merchant. We don't talk in quadratic pairs. We don't talk in graph neural networks. We talk in simple English. So, this is the English that I was talking you about. For a super user, we provide a different report, but for a
[23:25] simple merchant, we explain why should we run a save 75 cents promotion. In simple English. That is explainable AI. You can't do this with an LLM and expect the same answer 800 times you run it a day.
[23:41] If you have 1,000 merchants in a day and they're all running the same thing and you run the same 75 cents promotion on a given item and you expect the same answer with double-digit accuracy, there is no LLM that is guaranteed to do that today. There is no agent that can
[23:58] do that today. And by the way, the requirement for us is 30 seconds. No more than 30 seconds to produce a recommendation. So, you can't have a deep research agent think for 3 and 1/2 minutes for every question to say, "Well, let me I'm
[24:14] exploring. I'm thinking. I'm doing. I'm checking. I'm validating." Merchants get bored. So, we need to do this in 30 seconds at scale and that's the machine that's running the number behind it.
[24:32] So, now agents are still useful. It's not that agents are not useful, but as I said to you, predictions are good, but if they're not actionable. If they're not actionable, I already talked to you about this problem space of merchandising where things are fragmented. If I want to enter a cost change, I need to do it in price system
[24:49] A. If I want to do a price change, I need to do it in system B. If I want to do a regular promotion, I I to do it in system C. If I need to do a digital promotion or a personalized promotion, I need to do it in system D. Our merchants are juggling between these systems. They have multiple desktops open. They go and
[25:05] key in something here. They go and key in something here. They go and key in something here. Sometimes they will send out a mail to somebody offshore to say, "Can you update the cost on a nightly basis, etc., etc.?" But the actionability layer for us is bringing to that in one loop. So, the actionability layer, the agents
[25:23] discover the systems that's sitting underneath it, take a recommendation for cost, price, promotion, allowance, and backfill these systems. So, we don't use the standard reason-act framework of an agent. We use the act
[25:40] framework of the agent lot more than the reason framework of the agent. So, that's our role for how we use agents. It's not a regular usage. It is not throwing an LLM and an agent at the problem and saying, "Can you do your best to predict what the promotional
[25:56] retail is?" It cannot do it on 200 billion rows in 30 seconds. You have to train a model. You have to train the QKV on the model. You have to learn the representations that come out of it. Then make some predictions out of it. Then explain it. Then decide. Then
[26:13] act. That's the loop. That's the loop that we keep doing on a daily basis. How are we going for time? Are we going for time all right? 10 minutes? Okay. So, where does Databricks come into the equation? So, we we run uh data
[26:30] pipelines on lakehouse. So, we run roughly about 17 pipelines that pull data on a daily basis on the 100 billion rows that I talked about. Delta feed, obviously, we're not pushing that many data in a on a daily basis. So, that's
[26:45] our Spark layer that ingests all this data, build the lineage graph around it because we run a lot of metrics and the features that associated with the data so that the explainability is there. So, that's a bottom layer. The next layer is the governance around it. Who can see what data? We don't want
[27:01] a merchant A in Northern California pricing something in Jewel. So, we need to have a strong governance around it. How did we govern and protect our data for right usage? As well as how do we understand the lineage that
[27:17] exists between all this data set? For people who know of graphs and semantic layers and all that kind of stuff, that's what that's what our layer does. It understands the knowledge behind our data set and we stitch together a lineage that says, "A price is computed off of 16 different
[27:34] tables coming off system X." Our data team is sitting right here with me. They can tell you all about the complexity of the data that drives behind it. Without data, the science is meaningless. Anybody who is obsessed about science is obsessed about data. That's why it's data science. We spend a
[27:50] lot more effort on 01 than any other layer on the top. It took us 9 months to get 01 right. 01 was the absolute foundation for us. When data is coming from various systems on the back and they're not all
[28:06] connected, how do you connect the dots at the data layer first? What does the feature engineering look like? You heard the talk about row versus column. We endlessly debate row versus column. When does it make sense to use a row? When does it make sense to use a column?
[28:22] Because the layer on the top that's learning as the deep learning model needs to scale efficiently. We are running predictions for 2,200 stores and it needs to run daily and we only have a 4-hour window between East Coast and West Coast. So, we need to run all these
[28:38] computations on a daily basis. That's why 01 is the most critical piece of this whole infrastructure that we've been building towards. Then comes the compute. I talked about model Then comes the AI gateway where we register IML flow, we register
[28:55] our models on the AI gateway so that we can trace back our solutions. Then comes the partnership that we have with OpenAI and Genie together to make this work. So, unlike what you heard in many of the conference, we are the probably the only customer, the first customer use Genie
[29:10] as an API. So, we use Genie as an API because our enterprise chat is based on OpenAI's enterprise chat SDK, not what the chat GPT you see as a consumer. So, we are one of the very few Sam Altman announced us in July last year that we are
[29:27] to actually adopt chat kit, which is the enterprise version of the GPT because we can customize it for our data needs and our enterprise needs and our security needs and our training needs. So, we don't use chat GPT, we use a different flavor of it. And then we use Genie API
[29:44] under the covers to bring insights on that data. So, that ice cream stuff that I talked about, the Sargento cheese that I talked about, that's what Genie is doing for us. It's going and looking at this large lineage graph, whether it's predicted outcomes from a model or the data that we pipe on a daily basis
[30:02] from there. Okay, so what does it mean for us? So, it's a significant lift for us from the business. We are able to more pre- accurately predict what units revenue and AGP we derive given a promotion, given a price change, given a cost
[30:17] change, given an allowance change. So, that's the business lift. The merchant time compression, thousand users feeding data into many different system, hand keying that stuff is liberated completely because of the action ability. Third is the customer value. Obviously,
[30:33] we are doing this to attract our customers, to retain our customers, to service our customers better. So, we get lift on our customer value, too. That's what we are doing. Five take to take aways for me, you know, don't rely on the chatbot as your strategy. Invest on structure first. If you're
[30:51] building a clinical grade mission-critical system, it's the rounding errors will throw your business off completely. For us, if we make a price error on point one cent, that's millions lost or gained if you're a customer. So, it's
[31:08] a it's a big change for us. Accuracy is not the only answer. They need to be explainable. If they're not explainable, accuracy is as much it'll give you at the front door, but it'll never close the sale. And actionability trumps accuracy for us because if you can are not able to act
[31:25] on it, that's what it means. So, that's the closing slide I have. You know, the next world of is not going to be successful with the model that talks. It's going to be successful with a structure that's openly revealing itself.
[31:40] That's all I had.

Learn more about the Databricks Data and AI platform.

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