Beyond Dashboards: Real-Time Sales Insights with Databricks Genie and AI

Summary

Grupo Bimbo implemented machine learning for contextual anomaly detection and route risk scoring, Databricks Genie for natural language Q&A over sales data, and AI functions for personalized executive reports across 54,000 sales routes in 39 countries.
The architecture is built on a medallion model using Azure Data Lake with Unity Catalog governance, enabling the company to detect missed sales opportunities and surface actionable alerts to field supervisors without requiring BI tool access.
The initiative achieved 100 percent adoption among managers and a 17 percent reduction in missed sales opportunities by delivering the right insight to the right user in the right format.

Beyond Dashboards: Real-Time Sales Insights with Databricks Genie and AI

Delivering timely insights to field teams in high-paced commercial operations is challenging. Grupo Bimbo operates 54,000 sales routes across 39 countries, generating thousands of transactions daily. Traditional dashboards failed to drive adoption: field supervisors spend most of their time on-site and lack access to BI interfaces, while sales leaders became overwhelmed by data volume without clear guidance on where to focus. The company needed a new way to deliver the right insight to the right user in the right format.
Explore Grupo Bimbo's implementation of three integrated Databricks solutions: contextual machine learning for anomaly detection and route risk scoring, Databricks Genie for natural language Q&A over sales data, and AI functions that generate personalized executive reports. All solutions are governed by Unity Catalog and built on a medallion architecture using Azure data lake. The results: 100 percent adoption among managers, expanded field team efficiency, and 17 percent reduction in missed sales opportunities.
🤝

Chapters

FAQs

How does Grupo Bimbo use Databricks Genie for sales operations?

Grupo Bimbo deployed Databricks Genie to enable field supervisors to ask natural language questions over sales data without needing traditional BI dashboards or SQL skills. This is especially important because supervisors spend most of their time on-site and lack easy access to BI interfaces during daily route operations.

What is contextual anomaly detection and how does Grupo Bimbo apply it to sales routes?

Grupo Bimbo uses machine learning to score each of their 54,000 sales routes for risk by detecting anomalies in sales patterns relative to historical context and peer routes. This surfaces missed sales opportunities—routes with lower-than-expected sales—so supervisors can investigate and take corrective action in the field.

How did Grupo Bimbo build personalized executive reports with Databricks AI functions?

Grupo Bimbo uses Databricks AI functions to generate reports tailored to each executive's area of responsibility, translating raw sales and route data into narrative insights. This replaced one-size-fits-all dashboards with targeted summaries that highlight the most relevant opportunities for each leader.

What results did Grupo Bimbo achieve from their Databricks analytics implementation?

Grupo Bimbo achieved 100 percent adoption among managers and a 17 percent reduction in missed sales opportunities after deploying the integrated suite of machine learning, Genie, and AI functions. The platform governs over 1,700 sales centers and 54,000 routes globally through Unity Catalog on Azure Data Lake.

Full transcript

[00:10] Hello everybody. Good afternoon. Thank you for joining the session today. My name is Mauricio Jimenez. I'm the director of data analytics in Grupo Bimbo. Here with me Um hi everyone. Esteban Ramirez. Data science manager for internal audit department
[00:26] in Grupo Bimbo. Well, today we're going to present how to develop new ways to deliver insights to our users. How we switch from the traditional way of delivering information through our stand dashboards to develop developing uh new ways to
[00:43] deliver insights using Databricks capabilities. The main idea of this session is how to deliver the right insight for the right user in the right way.
[00:59] But first of all, let's introduce our company. Grupo Bimbo is a Mexican company. It's world leading in the baking segment. Besides we're known about the breads, also we operate on sweet baked goods and also salty snacks. We have operation in 39 countries that
[01:15] brings a lot of complexity our operation. And also we work with more than 100 brands. In the here in United States, I believe uh the most common most known brands is Sara Lee, Little Bites, and Takis.
[01:35] Well, also let me introduce you the process that we this project is based on is in the sales process. So the sales process in Grupo Bimbo is start by the sales order they receive from the clients. This the plants they produce the products and they ship to the sales centers. In
[01:50] the sales centers, we have the distribution of the products by clients. And then we allocated the products into the trucks. And each truck has some There is a plan for delivering the the products for the clients and we call the route. A route can attend like an
[02:06] average 10 clients. And when we say clients here, we're saying about the retail stores, right? So, big retail stores like Walmart to small ones like convenience stores and gas stations. So, as you can see here, this process is
[02:22] highly distributed, very operational, and dynamic. Also, as you can imagine, this generates a lot of data, right? So, we have worldwide more than 1,700 sales centers and we operate with 54,000
[02:39] routes. Each route again, on average, attend 10 clients. So, do the math and you can see how much generated data we have every day. So, in any process any company, we see opportunities, right? So, the main
[02:54] objective of the project is to highlight the opportunities for the sales department so they can act on the missing opportunities. Missing opportunities is like dollars left on the table, uh missing sales, and be more profitable, right? So, we developed a a lot of indicators.
[03:11] Some of them, for example, unattended clients. So, we have a schedule to attending the clients. Some clients we attend every day, others 3 days a week, other 2 days a week. So, we missing opportunity for, you know, visit a client is missing opportunity of sales.
[03:29] Other is the incorrect price. Incorrect price generate a lot of consequences on our on our process. For example, cancellations and returns, and also make some routes unprofitable. And also we have some indicators related
[03:44] to GPS. So, for example, when the salesperson is doing the sales, we record the the of the sales and we we got this the GPS geolocation of the transaction and we match with the client to see if the sales is doing the correct
[03:59] place. So I have other indicators here just three of them are example. And we did the traditional way, right? We developed indicators, developed the the dashboard.
[04:15] However, the results didn't expect what we are expecting for for this. Um the the usage of this dashboard and adoption of the dashboard was not what we expecting and people are not acting on the opportunities that we are highlighting. Then we start to start to to dive into
[04:31] the root cause of what is happening and we saw some some things here. So for example, sales leaders told us "Hey, this is too much data. I don't know where to start. I don't know where I can focus on first. I don't know where to attack first. So you are generating
[04:47] too much data that I cannot even know where to start." Secondly, is the context. The sales supervisor, they are the role to be on the field. So they are the role that they have is to be with the clients helping issue helping solving issues
[05:03] helping also the management with the clients and prospecting new clients as well. So they spend only one day on the office and they spend like most of the days on the field. So for them to be using dashboards on the field is very complicated and also for the last one
[05:19] this type of role, they are not used to be you know doing analysis on dashboards. So this is some the reasons that the usage of the dashboard was not expecting what we we wanted. In one point we have the challenge that we could on our dashboard bring the
[05:37] opportunities that we want. The technical side was solved but even though the challenge is how we make people to act on this opportunity and really solve this issue that we are highlighting, right?
[05:53] Then we come across these solutions from Databricks, some features that we implemented here. Uh first of all is a machine learning with algorithm that detects anomalies and we develop like a score. So we highlight and sorry, we can rank these the routes by more
[06:10] opportunities so we can the sales leaders can focus on the routes that have more opportunities to be solved. Also, we implemented Genie. I believe everybody knows about Genie, right? So is the model from Databricks that
[06:25] convert the natural language in a query in a database. So we connected Genie to our database of the opportunities. So what is more easy, right, for the sales supervisor there on the field. So just take the phone and do questions to Genie. So for example, which clients
[06:42] should attend this week? What what is the routes that need more supervision of my side? What are the products that are generating errors and cancellations? So simple questions that they can do and deliver insights right away instead of then be looking a
[06:58] dashboard, filtering and do some analysis. And lastly, one function that we implemented here we call um intelligent report. And this report is for the top management. So you use AI
[07:13] function that calls artificial intelligence that read this data of the opportunities and generate a report with the summary of these opportunities. This is the top management so it can also highlight for by regions, for trends. And this report we automate to
[07:32] be delivered every week to the management. So then they can use this information for their analysis. So those three um, uh, that we implemented through Databricks that we use to diversify the way that we bring information, bring
[07:48] insights to our users. So, now uh, Esteve is going to show you show us how we develop this more in a technical thing. Thank you, Mo. Yeah. So, well, we are going to review these three solutions very quickly. Uh, here is the
[08:05] the architecture that help us to make everything happen. We have the route to market market system. Uh, we have Databricks on Azure, so that's why we are using an Azure data lake to extract all the information. And then to the ingestion, we have two types of
[08:22] ingestions. We have batch ingestions and also CDC ingestions. It depends on the information that we are working with. And after that, well, we have the medallion architecture. Uh, just to review quickly this, the bronze layer
[08:38] the bronze layer keeps all the raw data. Then we have the silver one that has all the clean data and the duplicated data. And then the gold, uh, keeps the business ready, uh, information or the business ready reports.
[08:55] Uh, all this with Unity Catalog and it's orchestrated with Lake, uh, Lake Flow jobs. And well, these three solutions that Mo presents, uh, we develop with three different, uh, ways. We have the traditional BI,
[09:11] but the important part of this presentation is the AI BI Genie and also the the AI functions. So, we're going to start reviewing the machine learning. Uh, this solution we
[09:26] present the results on a dashboard, but this the dashboard is not the important part, but the machine learning model that help us to focus on those critical routes uh that has more risk on them or the worst
[09:43] score. So, now I'll talk about these scores that we calculate for each route um based on the ticket uh results. So, we want to focus on those one that have most critical uh risks, yeah? And when we start analyzing
[10:00] this, we find some correlation-ships between some supervisors or some sales centers also where we can get some our insights. Uh well, so we have some uh big challenges here um because we detect
[10:19] an anomaly is contextual, so this is a phrase that maybe makes some sense to you. So, what is anomalous in San Francisco maybe isn't anomalous in Mexico City, you know? And going deeper on this, if we have a route uh the dense urban area that maybe it makes 80 stops
[10:37] a day, and maybe each client is like 200 m away, so maybe if we find out a client or a visit uh 600 m away, maybe we will say that is normal, right? But on the other side, if we have a
[10:53] rural area route uh that has only 25 stops per day, and maybe the average distance between each client is 3 km, then a 60-m visit will be maybe anomalous, you know? We need to focus on those um anomalous visits.
[11:09] So, that's why we we need to our model to be contextual. And also, we face three main problems that we need to solve. So, the first one is, as we mentioned before, we have the information by ticket, and
[11:25] if evaluating each route is difficult, imagine evaluating each ticket, you know? So, we need to find a way to summarize all this information into route information. So, that's why we aggregate that at good at route level.
[11:44] Uh another The second problem is sometimes an average can hide a risk, you know? There are placed two examples. For example, if we have the route A with uh 50 tickets that have middle risk and then 50 tickets with no risk, but in
[12:01] the other side we have another route that have five ticket with extreme risk and just 95 uh tickets with normal risk, then maybe the average can be pretty similar and maybe we could skip this route B uh that
[12:16] is more risk for us. So, that's why we split the anomalous versus the normal with a weighted average. And this weighted average so obviously needs to be very intuitive cuz the sales centers people could ask us why we are
[12:34] measuring the risk like this. So, we make this uh formula pretty intuitive for them and explain them how we calculate this so they feel comfortable using these metrics on their analysis also.
[12:51] Well, once we solve these three problems, um so, it's pretty difficult to give context to just one model because we need to get a lot of features. So,
[13:06] that's why we don't uh create just one model. So, we create hundreds of models in parallel. So, I saw is the isolation forest model, but we divided by sales centers per sales type and channel. So,
[13:21] this give us some context for each one of the routes and no one route or one context for all the routes, no? So, the example that we review before that the urban and the rural route, they don't are in
[13:38] the in the same threshold. They have each one their own. Yeah? Um and well, each model learns what is normal to that model, no? They they don't compete between them. So, this is the the first solution. We
[13:54] create all this risk evaluation. We can focus on the more risky routes. But, this is just the the half of the of the project, no? We need to also develop some other things to deliver this information to to the right people.
[14:10] And there is where Genie gets in, right? So, with Genie, I'm going to show you um how we how our Genie works. So, it's pretty intuitive and also helps the people
[14:25] to go through through it. Sometimes it's difficult that the people on operation side school ask right what they need. So, this Genie helps you and drives you through your region, uh which health center you are exploring also, and it uh lists all the indicators
[14:43] where you can ask ask it. So, we have all these indicators. So, here uh in this example, we are asking for all the clients that have sales uh 1 km more than 1 km away from its
[14:59] geolocation that we have stored on on our RTM system. And something interesting that we can pick one, we can see the information as a table or as a chart. But, something uh
[15:14] we like a lot is that also we can see on the map on that point that mark is that there's was a sale that is not on the on the client geolocation and we can also investigate that if it is true or not
[15:30] that there's a client or not. So, that's pretty useful for the supervisors also. Uh well, as everybody knows, Genie uh help us to our team to converse with their data. So, it's like a self-service
[15:45] Q&A service. Um The idea is that we can ask a natural language and Genie just go place an SQL query and then give us back the answer, right? Uh So, but it's not that easy. I
[16:03] mean, it sounds pretty easy just connect the the info and that's it. But you need to to to take care of three ingredients, right? So, we think that the three ingredients behind accurate answers are the three. First, uh curated data data asset
[16:20] because I know if you as you remember this phrase that is garbage in, garbage out. So, you need to clean pretty well uh what we explained before, uh root level audit table and not just the rate
[16:35] the raw the raw data of the raw ticket. So, you need to clean it pretty well. And also uh say Genie and and explain Genie everything that you want to add, no? All the joints and all that. It's be very clear with with the data set.
[16:52] The second one is uh full metadata. Databricks have an option that where you can add metadata to your tables. It's pretty useful. So, I suggest you to use it. Um also, you can use AI to fill up all these information and then you take a look and
[17:09] and complete or modify what the the agent uh fills. But it's very important for Genie because sometimes uh you For example, here in Grupo Bimbo we have a lot of different terms
[17:25] to refer to different to the same thing. Yeah, so all these meta uh metadata help us to to describe these two Genie. And the last one is a custom prompt. Yeah, so we gave a prompt to the Genie on how we
[17:44] make our audit or the rules that are more important or how he how it needs to consider some um some of the rules that that we that we implement or the indicators. Uh sometimes auditors refer to the same
[18:00] indicators with other words. So all those rules is important to be on on our prompt. So this is how we build our Genie. Um
[18:17] and then the AI functions, right? So we found that AI functions are pretty useful for us. Uh I place here an example of on how we develop and how we deliver to each manager for each region these um
[18:37] executive summaries. And we use the machine learning model because this all this is uh summary is related to the context of each uh executive. So it's not just one single
[18:52] template or just the same indicators for everyone. No. It reads all the context and gives the executive the most relevant indicators and some suggestions also on on what he needs to work. Yeah?
[19:10] Uh how we make this? Well, first of all, I want you to introduce to you AI functions. Um, there are many of them. We use one specifically that is an AI query. Uh, this AI query well, is native from
[19:27] Databricks. You can run it as SQL either the SQL editor or your notebooks. That's pretty simple. Uh, they are governed by Unity Catalog. Um, as I mentioned before, you can run it with your serverless clusters, and you
[19:43] can also use any LLM uh, that that you like the most, Lama, Cloud, GPT, uh, the one that you prefer. Um, and you don't need to add any other infrastructure or anything else. You
[19:59] just place uh, which uh, serving endpoint you want to to to use, and you just run this this function. Um, and it's pretty friendly with your pipelines. So, as I mentioned, it's like an SQL function, so uh,
[20:15] it doesn't need much more. So, uh, this is like the prompt that we use. Uh, it's just just a part of the prompt that we use. So, here you need to place which one is the LLM you're going to use, maybe the fields of the table
[20:33] you uh, using to generate all these analysis. Everything you use uh, come the data from Unity Catalog. It takes it from them, analyzes all the all the columns that you place on the very top, and then you can write also a
[20:48] prompt, right? So, so in his pro- in this prompt, uh, we ask um, the function to develop uh, personalized report based on the indicators that we present before uh, for each region. Uh, take just the
[21:05] most um, important, um, tell some suggestion pretty easy suggestions because we know that sometimes AI could make some suggestions that is not possible for the company in this moment
[21:21] but we ask the LLM to give some easy suggestion or actionable that could be quickly to implement this this suggestions and then this is a story again on the Unity catalog. The last part of this
[21:38] is this analysis and this report that the AI function develops. We take it is on the markdown format and we use report lab to convert it to a a PDF because it's easier to everybody
[21:54] just open a PDF and read it. And automatically with the use of data bricks we open this PDF and send it by Outlook. So all this process is fully automated and nobody needs to move anything is on
[22:12] on a job it runs extract information the markdown converted and then send it by email. So for each region we can send one personalized email with all the indicators that must matter
[22:27] for them. Well, now let's talk about the results, right? So I think this before saying about the the hard numbers indicators I believe the great achievement of this project the great value is the change
[22:45] behave, right? Change behaving on our team of developing so we understand that we have other features that we can implement to deliver insights for users and also for the users, right? The users always expecting to have the dashboard to be using for their you know the
[23:02] information to consume. Now they have other types of uh um um features that they can use from passive dashboards to more active ones like a genie you can interact, you can have questions. The report that receive you every Monday on the morning with the
[23:18] information that you really need. So this I think is the best of uh this project on terms of the achievement that we have. But also for example, efficiency uh for our sales supervisors, they have opportunity to look for the dashboards just once a week
[23:34] when they have one day on the week on the on the off-season. But the genie they have the sixth day they will work on available for them to be working in this and available for them to receive the the insights and act on the on the opportunities that they need.
[23:49] Also the sales impact um since we implemented this, we are monitoring the results and we could uh see that we decrease in 17% on the missing opportunities on the sales um process the region that we implemented.
[24:05] And also the adoption, right? The we can have more controls of who is using these tools with data bricks. Um also we can see that for example, 100% of managers the region that we implemented receiving.
[24:20] And one indicator that is difficult to put in numbers but is a feeling that you guys that sometimes they you develop a a dashboard and no one ask you to do any enhancement, any change or correction, the probability that no one is using, right? So
[24:36] we could feel here in this project they are asking for, you know, enhancements, some change, some some other features. It's a good feedback that are really using the the uh our uh features. And what we learn on the business side.
[24:53] First of all is to change it mindset. So change the way we work. I know that we also we deliver digital products. It's good for us that we have standardized process, you know, to be more efficient, to deliver more digital products. But also sometimes we
[25:11] need to change the way we do things to adapt to you know the the way the people use your your products. Secondly is understand your user, right? So go in in in detail with them, so us spend time with them on the field like
[25:27] we spend time with them looking how they receive information, how they use the the data, and then we can understand better the use behavior and then adapt to the tool for their own use. And also for linking the the third one
[25:45] is try to personalize your your your your deliverables. So again, when we deliver product digital products is is good to be standardized, but we need to understand that no one not everybody
[26:01] from the company use the same information the same way. So adapt to the users the way they they consume the data, it's very important. And for the technical part, the prompt prompt it really matters,
[26:18] right? Because we know that the artificial intelligence itself we don't have so much adjustment that we can do on the artificial intelligence. It's like a black box, but we can do is on the prompts. So the prompts try you know different formats,
[26:36] different ways, so as good that you have your prompt better is your result you can have from the artificial intelligence and try also some frameworks, try new ways, try different context and roles and you can have more from the artificial intelligence.
[26:53] The second is try to use also the the building features. You know, Databricks has a lot of them and we can use them um already built it for example Genie, right? They have already the artificial intelligence connected there. So, you have like a front user for the users.
[27:10] This more for us is to adapt and and customize this tool for for the ends end user. So, not always doing from scratch is the solution. Try also to look what you have in place already to take advantage of
[27:26] the tools already built it. And Stephen also talked about the data, right? So, data is very important um to prepare to clean the data before loading in the artificial intelligence, right? So, I sometimes see people
[27:42] loading all their raw data on the artificial intelligence and giving a lot of prompts to say, "Hey, clean this data, prepare this data and then do this analysis, right?" So, I know this sometimes is tempting, but uh my recommendation is go with analytics that
[27:57] you can clean and and prepare the data. Do the most that you can to prepare the data and load this information to the artificial intelligence to let the artificial intelligence do a specific analysis that you want. You know, as much things that you asking
[28:14] for them is more results that you need to be managing on the on artificial intelligence. And lastly is the machine learning. Um nowadays that we always see about artificial intelligence, generative AI, um don't forget about machine learning
[28:31] that is a very powerful tool for working with data. Uh for us here, we have a very good combination with machine learning to bring some sites and also generative AI to how consume this data. And today also in in in the summit we
[28:46] saw new features from from Databricks from machine learning. I'm very excited to use them. And this is what we like to present to you today. Sorry. So, how we could manage to change the the
[29:03] behave for you know for you as to develop their tools to have other ways to deliver insights to our users and also for our users, right? So, how they can have more information, more insights to act on the way they need to do. So, delivering the right insight, the right
[29:19] user in the right way is what we are implementing here. Hope you all have a good time have a good time here. If you have any questions, Thank you everyone for joining.

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.