7-Eleven is using AI from Databricks to help their teams understand store performance
Summary
- Hyper-local personalization: 7-Eleven franchise stores use Genie AI agents to ask natural language questions about their specific local market performance, enabling unique store-level insights across 14,000 US locations.
- Data answering bots: 7-Eleven discovered that using Genie API with consolidated data sets outperforms complex personalized search pilots, exposing answers through Teams bots where staff naturally communicate.
- ROI focus: Rather than bleeding-edge AI experimentation, 7-Eleven prioritized simple data-answering use cases that drive immediate value, like helping franchise owners understand their regional sales trends and performance.
7-Eleven is using AI from Databricks to help their teams understand store performance

7-Eleven is using AI Agents from Databricks to empower each franchise store to uniquely understand how they are performing. Now, their users can ask questions about their data in natural language and get back results immediately. Come listen to this talk from 7-Eleven, as they discuss their use case, their architecture choices, and the benefits they've received from using GenAI.
Chapters
0:00Introduction to Generative AI0:30AI in retail operations1:43Hyper-localizing data2:50Ensuring AI quality3:53Compound AI architecture5:32Future AI roadmapping6:54Driving ROI with RAG
FAQs
Why does 7-Eleven need hyper-local insights?
7-Eleven operates 14,000 US locations using hyper-localized management. Store managers need to understand local performance and assortment relative to local competition for better decisions.
How does the Teams bot solve data access?
Staff ask questions like 'What did Southwest sales look like this time vs. last year?' and get instant answers. This replaces contacting analysts and eliminates delays in accessing information.
What will enterprise data platform architecture look like in 2030?
7-Eleven's architecture shows the future: compound AI systems using embedding-based routing and LLM nodes to answer business questions through Teams bots, eventually migrating to more sophisticated LLM-based routing. The company starts with embeddings against consolidated data sets and plans to evolve toward agentic flows for handling ambiguity in queries and routing complexity.
LLM routing vs using one frontier model for everything
This video reveals 7-Eleven's architectural evolution: currently using embedding-based routing against constrained data sources works well, but as datasets expand and overlap between teams grows, LLM-based routing offers better handling of query ambiguity and context. Frontier models like o1 improve routing automatically, potentially eliminating downstream prompt engineering complexity.
What does the future of business intelligence look like in the age of AI?
This video shows 7-Eleven moving from Power BI dashboards to AI agents answering hyper-local retail questions in natural language via Teams. Staff ask questions like 'What did Southwest sales look like vs. last year?' and get instant answers, enabling faster decisions about inventory, promotions, and staffing without analyst involvement.
Full transcript
0:00 Introduction to Generative AI
Hi everyone. I'm O'Lenic. I'm a product manager at Databricks. And I'm the senior director of data and machine learning and R&D. So I lead both the data platform and the innovation groups. And I work for 7-Eleven, which is the world's largest convenience retailer. So a lot of our customers at Databricks are trying to figure out, "Hey, what should they build with generative AI? And then how do they get ROI out of these use cases?" , can you kind of walk us through at 7-Eleven, what are kind of the use cases that you all are seeing the most excitement around in your org and how does that align with like what you all are doing on the bleeding edge versus hey, where is where is actually the most value being delivered?
0:30 AI in retail operations
Absolutely. Yeah, so we've done a lot of really cool things at the bleeding edge in terms of personalized search and all these other things that we're piloting and trying out. But really where we've seen the most excitement is simple data answering bots. So we've we've trained LLMs, but what we find is just using the out-of-the-box models, especially the existing Genie API with data rooms has been a super effective method with a consolidated data set for people to be able to ask questions and receive answers that are meaningful.
And so we're currently working on exposing that through the most commonly used platform for communication at our organization, which is Teams, because people aren't logging in to Databricks to do this. Typically, they're pinging an analyst on Teams to say, "Hey, what did sales look like for the Southwest region at this time compared with the previous year, right?" So being able to just ping a Teams bot rather than hunt through, find the analyst, and hope you get the right answer has been very exciting for the business.
And is there something about 7-Eleven in particular that like makes it so that your business owners and stuff need to like better understand what's going on in like their area? Or what's kind of like the value that people are really getting being able to like talk to their their business data or their like retail data. Yeah, the biggest thing really is that we have just about 14,000 locations in the US alone. And we do something called retailer initiative where hyper-localized and we really care about how our stores are performing based on the local markets, right?
1:43 Hyper-localizing data
If you go to 7-Eleven, uh we want you to know the cashier. We want you to know the assortment and it's unique to your region. So, we are focused on really understanding how each store is uniquely operating and how each region is uniquely operating. So, sometimes discovering those insights at a very micro-geographical level is very important to each of our area leaders uh who are looking at that data. So, we haven't gotten to the point yet where we've exposed these bots to those leaders yet.
We're in the process of letting the executives look through them and interact with them, but the next step naturally is to allow field-facing individuals to leverage this data more naturally. Today, they just rely entirely on Power BI reports. What we found at Databricks is many of our customers are really struggling to get to quality. Like getting quality GenAI apps into production is hard. And so, many of our customers are finding ways to sort of embrace data intelligence.
2:50 Ensuring AI quality
So, by using their enterprise data to improve uh general intelligence to achieve quality or they're using compound AI systems. So, rather than shipping like one monolithic model with like maybe like a huge prompt on it, you're actually breaking it down into like modularized components where you can like specialize each component. Yeah, I mean, for us, especially with user questions related to sales or other data, we wanted to give a real very valid answer.
And so, having an LLM perform non-deterministically is a risk that I'm often not willing to take when it's related to either um performance, store performance data, financial data, or any query anyone has. So, we've been using embedding models to route rather than agentic flows. Um we're starting to look into agentic flows as a way to do this effectively as well, but as of now, we're using embeddings. And I can share some of the architecture diagrams to show you where we're at today and where we're trying to go.
So maybe, can you kind of like walk us through what the like compound AI system is that you all have built at 7-Eleven? Yeah, I'll go through two things. I'll I'll go through the first version. I I have a document which may help follow along. Um So at a basic level, the first version of it that we built was user submits a query, we use a text embedding model in LangChain to do a loader basically with face. And essentially, we make sure it's not a query, then we choose LLM based on the text embedding model, and we do the general processing, right?
3:53 Compound AI architecture
Where we're headed is more towards using an LLM node to do some of the selection of where to route things, whether vector store, Genie API, Confluence. We have a number of different tools we use. So you are going to use an LLM to do routing? Eventually is what we're attempting cuz right now the embedding-based model is working and we're happy with it. But long-term what we've seen is that as the complexity of our use cases grows, right?
As we expand the number of data sets that we expose as data products, how do you ensure the routing's consistent in a way that there may be overlap between certain data sets, right? Like merchandising may look at sales a certain way versus how marketing looks at sales. So how do you make sure you route both sales queries correctly depending on the context? Um And so here we're doing a couple different things with a document grader, um then based on relevance, an output generator, we validate there aren't any hallucinations using an LLM node, and if there aren't, then we move on to the answer grader, and finally whether an answer is the question.
So this is still a rag-based approach with multiple agents throughout that process. Um but this is still early stages for us cuz as vanilla as this is, it works, and um we've seen great success, and users are happy with this. The reason we're looking at the future facing is because in many ways LLMs, as you're seeing with 01 and others that are coming out, they're becoming more and more commoditized where you can have huge gains in in performance simply by swapping out a new LLM at some point.
5:32 Future AI roadmapping
Um and so if they're able to process some of the ambiguity in a query and the routing portions of the query um as a single interface routing, it'll solve a lot of the downstream issues we see with prompt chaining and setting up the right set of prompts. I see. And the reason that you're going to like eventually move architectures is cuz you're expecting an improvement in quality or what sort of motivating uh Yeah. change in architecture?
So embeddings work really well when you have a very constrained search space. So if you know your users are always asking a very specific set of questions within a given data room, very powerful. As you expose more generalizable uh queries to the user set and then have it routed accordingly, that becomes more and more complex. So because we're exposing greater data products to the rest of our business with varying varieties of what the context may be, we're looking to make sure that we can handle the different types of users and queries that they may put in, right?
An expert in finance will ask a question very differently than someone in the field. And so we want to ensure we don't drop in performance as the language being used changes. Time to value is actually surprisingly quick for a lot of low-hanging fruit. So, you know, we we've done a lot of simple things with troubleshooting manuals um for our field technicians where when they would have an issue with a particular system, um they would have to go hunt through a bunch of documentation to find the answer.
6:54 Driving ROI with RAG
Now they just ask a question and they get the answer. So, you know, it's not groundbreaking work or anything that we're doing. We're just taking the existing manual, passing it through an LLM, generating embeddings, and then returning answers, but it's yielded enormous lifts in terms of productivity and resolution of tickets. Um so for those who are hesitant to try stuff out, I think RAG in particular with embedding-based models are very simple ways to get into things to drive productivity.
Don't hesitate there. Everyone will talk about the bleeding edge stuff, but the highest return on investment and value is the boring stuff with RAG. I really appreciate you taking the time to be here and we're excited for what you build next on data bricks.
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.