Global Toymaker ZURU Fuels AI-Driven Innovation with Crisp Retail Data in Databricks

Summary

Crisp delivers clean, retail-specific data from major US retailers into Databricks daily via Delta Sharing, enabling ZURU—maker of Mini Brands, Bunch-O-Balloons, and Snackles—to run agentic AI workflows across supply chain, sales, and marketing.
ZURU has achieved 55% distribution growth, 50% markdown reduction, and 88% holiday sales improvement by using Crisp's retail data foundation and Databricks-powered AI for fast daily decisions across inventory, product, and marketing teams.
Crisp's vertical AI platform includes a conversational retail intelligence layer built on MCP, A2A, and Genie integration, enabling CPG sales managers to query sales decomposition, weather event impact, and price sensitivity through natural language.

Global Toymaker ZURU Fuels AI-Driven Innovation with Crisp Retail Data in Databricks

ZURU (Mini Brands, Snackles, Bunch-O-Balloons, and the latest plush trends) operates at the cutting edge of modern retail by investing in strong data infrastructure and enabling organization-wide intelligence.
Teams leverage centralized AI-ready insights across major retailers, to inform fast daily decisions across inventory, marketing, product, and more. Retail-specific semantic modeling creates a trusted foundation for AI workflows
At the core of ZURU’s retail strategy is Databricks. Clean, real-time data is delivered into Databricks daily via Crisp, and informs Agentic AI workflows across supply chain, sales, and marketing.
AI-enabled retail intelligence has been the key for ZURU to scale their multi-category business globally.
this video explores how pairing clean, AI-ready data from Crisp with Databricks enables responsive decision-making and supply chain excellence at retail scale.
Talk By: Julienne Biglin, VP Partnerships, Crisp ;

Chapters

FAQs

What is Crisp and how does it work with Databricks?

Crisp is a retail data infrastructure platform that has spent eight years building data foundations and analytics for CPGs and retailers. It delivers highly granular POS, inventory, marketing, and shipment data from major retailers into Databricks daily via Delta Sharing, covering up to 80% of a CPG's total revenue footprint.

What results has ZURU achieved using Crisp and Databricks?

ZURU, the toy manufacturer behind Mini Brands, Bunch-O-Balloons, and Snackles, achieved 55% distribution growth, 50% markdown reduction, and 88% holiday sales improvement by using Crisp's retail data foundation and Databricks to power agentic AI workflows across supply chain, sales, and marketing.

How does Crisp's AI platform enable conversational retail intelligence?

Crisp's vertical AI platform is built on its Rosetta semantic layer and integrates with MCP, A2A, and Databricks Genie to enable natural language queries over retail data. CPG sales managers can ask questions about sales decomposition, weather event impact on store performance, and price sensitivity—and receive answers grounded in structured retail data.

Why does Crisp use a vertical AI platform rather than a general-purpose approach?

Crisp argues that purpose-built, vertical AI platforms outperform general-purpose horizontal platforms in retail because retail data requires domain-specific semantic modeling, retailer-specific data integrations, and knowledge of CPG business processes. This retail-specific context is what makes AI outputs reliable and actionable for sales and supply chain decisions.

Full transcript

[00:08] All right, thank you for coming. Can everyone hear me? Okay, good. I'm uh Dagly Odin. I am um a co-founder and um head of products at Crisp. And today I'm going to show you how u we help retailers and CPGs u become more
[00:22] effective uh by using AI. First, a little bit about Crisp. What is it that we do? Well, we've been around for about eight years, and in those eight years, we've been building data
[00:34] foundations and analytics for retailers and CPGs to help them use that data more efficiently. Basically turning data such as POS data, so it's being sold in stores, inventory data, whether it's
[00:46] distribution centers or in stores, marketing data, shipment data, all of those things, turning those into something that is is actually actionable. We work with
[00:58] most of the largest CPGs uh and food manufacturers um in the US in particular. We help them in their Walmart ecosystem and Target ecosystem but also across the longer tale of
[01:10] retailers they might be working with. So if you're a retailer you want to understand your full book of business where you're selling not only to the big ones but also the smaller ones. Chances are that we can give you at least 80% of your total revenue footprint um with the
[01:22] integrations that we built um over the past eight years. So we source highly granular data whatever grain the retailers share whether that's on day store skew or some slightly higher um aggregations
[01:35] been a database partner for uh 3 four years now um and a very close relationship with them on building out the stack. So what does it look like in in practice? Well, we have our own AI
[01:49] platform. I'll talk more about that in a bit. It's a vertical uh AI platform, but if we start at the very bottom, it starts with data sources, of course. So, we help sourcing that data from retailers, which typically can be all
[02:01] the way on one end of the spectrum. You have companies like Walmart which has a very sophisticated data sharing capability directly on data bricks all the way through FTP servers and web portal scraping and Excel downloads and
[02:14] piv transposing those putting it into warehouse. So typically the setup is that we are running managed ELTL pipelines for all of these very brittle data integrations that the retailers and distributors have. Then we combine that
[02:27] with the LTL pipelines that our customers have using delta sharing. And then on top of that we then have what we call Rosetta which is our semantic layer. And upon that we build various applications. So we have of course sales
[02:41] analytics um supply chain ops space planning and many other things um as well. And on top of that we have our AI stack. That AI stack is integrated um with A2A or MCP uh with Genie. So we can
[02:55] speak both ways there as well. And I'll show a demo of this in a little bit by the way. Okay. So what what does this actually mean uh for a a CPG? So our case study partner here today is a
[03:07] company called Zuru. They're a toy manufacturer. Um they sell to almost all supermarkets um in the US uh and across the globe. Uh and they started working with us about four years ago. Um and it
[03:21] started with uh Delta sharing. So we were powering their um their data bricks and data lake um with all of these retailer data and then more recently they also started leveraging our AI platform and over the past few years
[03:33] they've had some pretty staggering numbers that they they they give us credit for. Uh I haven't checked all of these myself but it's it's really impressive numbers. So first they've grown their their distribution
[03:45] footprint. So that's the number of product times the number of stores they're available in 55% yearonear for the past four years. They reduced their markdowns by 50%. Meaning that they have the right
[03:57] products in the right place. If you overstock your products, you're going to go going left on shelf, you have to mark it down, reduce markdowns and therefore waste by 50%. And then they have had um toys is a very very seasonal uh uh sales
[04:12] cycle as you can imagine also why they have a lot of markdowns. So their holiday sales are up 88% in the same period. So what makes Crisp AI different? What
[04:24] is it? Well, we are not a horizontal AI platform player. We're not we're not developing Genie. We're not developing C-Pilot or anything like that. We are developing a forp purpose-built agentic harness for retail. Um and that makes it
[04:38] very very different. It's a lot more token efficient. Um it's a lot faster and I'll show that later. So what is the difference between a vertical and a horizontal play? Well, when we work with our customers, they're typically sales
[04:51] managers that sell into the retailers. They can be category managers that work on behalf of the retailers. It could be supply chain managers and so on. They have a lot of knowledge, a lot of context about their the their business.
[05:04] What we provide is a complete turnkey product that has a very rich semantic layer that describes all of the business metrics that are re relevant to a specific retailer. And then we also
[05:16] provide all of the skills and the documentation and knowledge base that is searchable by the agents so they understand how they're actually going to work with this. And uh I just learned from that because I missed the the the keynote that uh that data bricks now has
[05:28] an ontology layer. It's very similar to that and something we'll look into building on top of as well. But I cannot underestimate or understate how important that ontology is. And I'll show that in a second why that's so important as well.
[05:40] Um yeah. So our not only do do we build the agentic harness, but we also have people in the field with our CPG customers that are updating constantly improving the way that the tools work
[05:52] together, looking at token efficiency, looking at speed, uh looking at developing new sub agents that take care of certain types of analysis and and so on and so forth. So with that, I'm actually going to get into um a demo.
[06:04] And I'll there's a lot to say about the the platform and I can't cover everything but I want to I want to first start with the conversational part of of the platform. So showing some of those
[06:16] uh uh what what makes it different feel different as as a platform and also get into um what is the really the ultimate goal is to have fully autonomous agentic runs agents that run or constellations
[06:29] of agents that run entire business processes. So this is the Crisp AI platform's web UI. Uh it can operate as a headless platform as well. Like I said before, it can be um the agents can be
[06:41] exposed to other platforms such as Genie. We also support to A2A so you can do streaming uh integrations with other AI platforms. Um and here we are going to look at some uh synthetic data but it
[06:54] is based on the Walmart u semantic model. So um we'll get into this. So um I I'll do some live things here as well but a simple question such as what was
[07:07] my revenue by brand last week um that is a trivial question um but already there you can start unpacking some things. So if you give it to a general purpose uh uh agent first thing okay what do you mean by last week is it the current week
[07:20] or is it the last week and it's so is it the fiscal week um or is it a a Gregorian calendar week these things are pretty messy especially when you're around like week 53 if some of you are in retail you know about the the retail calendar it can get pretty messy and the
[07:34] just just modeling time intelligence on top of uh a a retailer data set is is is quite complicated um but this is also what we do we kind We tune these the the prompts and the skills and what we
[07:46] inject into the prompts at various uh places so that you can get the right answer very quickly. So here broke it down by brand of course. So it's a little bit of a wide screen uh here. So um but here's a another example. If I
[07:59] say I want to take a look at this this percent here. So what um or explain this number? So what you would want in a general purpose AI when you ask this
[08:12] question you'll probably say well 77% is uh previous year minus current year over previous year plus just do the simple math like that right um because this agent understands the domain not only
[08:26] does it well in the first place we ask for revenue it understands we want year on year because that's typically how uh retailers look at this data but it also understands that when I'm asking the question about what is this number I don't want to just show the math. What I
[08:39] actually want to want to do, I want to understand what is driving this growth of 77%. So, um, as it's working through here, we can click into it. It understands what
[08:51] Oh, okay. I lost my model axis here. Looks like I mean just I'm going to pull up one I ran for 10 minutes ago. Uh, here we go. Sorry, I asked for
[09:04] There we go. So instead of just telling me how this number is defined, it actually does a sales growth decomposition and looks at okay the 77% growth. What is it that drove it? Some
[09:17] of it is volume. That's the least interesting answer to typically. Um but a selling price, your traded stores, traded stores is the distribution footprint metric in the Walmart world. Those are the things that matter. It
[09:29] also drives a little bit deeper because instead of just looking at the brand total, it's gone into what are the items that have been driving this. It looks at are there any any draggers on it. So understands that okay, I have digital
[09:41] sales, I have physical sales. What are things holding it back? Of course, we could do now do follow-ups and go much deeper into this. But it gives you an idea that this agent understands the domain of the analyst and the the person
[09:53] that is working with it. Instead of just kind of giving the obvious data answer, it goes a lot further. Um
[10:07] an interesting example um one of the first clients we had uh that was onboarded to this we we started rolling out this platform uh in early March um and uh we had some winter storms in the US that have been rolling across uh especially the southern part um and a
[10:21] question that they have is did those how did those winter storms actually impact us um and posed a simple question to the to the agent. Uh I rerun this now. You know what what um can I I I shown a lot
[10:35] of things here but how how did it impact us? And intuitively the agent is able to understand okay here is the impact. I'm looking at weekon-week uh drops in sales during the storm. I'm um sorry and then
[10:48] oh actually before the storm you have an increase after the storm you have a surge because you had a lot of pantry loading that happened after. And these kinds of insights into how big events impact your sales, your
[11:01] supply chain and so on. These are questions that typically a retailer or an analyst would never even bother to ask because they are so uh they know they would take so much time to to to um to get an answer to them.
[11:14] All right. Some other um capabilities, price sensitivity. Um a lot of our clients in the past when we have delivered you know standard dashboards
[11:27] uh to them they ask okay can you help us run a price sensitivity analysis for X. If you ask an agent any kind of model do that they will come up with a solution to it but it's probably not something that is tied to the data that is
[11:40] available for this particular retailer. So when we ask an agent like this, help us do a price sensitivity analysis, it helps us guides guides the user through this process and it already knows all
[11:52] the metrics that are available, all the ways that it should look at the business for for example in this case for Walmart and very quickly ends up um delivering the results for that. So uh we have where we used to have data science we
[12:04] still have data scientists on our team but they are now working on building a corpus of blueprints that agents can use that are tied to our semantic models and helps our users now customize these analysis in a way that they previously
[12:17] they would have to pull up a Jupyter notebook and do that. Now they can do this in a matter of of minutes. Same thing for store clustering. Um same thing for diff and diff an analysis and stuff like that. So uh it just it just
[12:29] brings a completely new level of of capability to uh team members that otherwise would have to depend on data science team and they have to you know how that how that that works. Things take time. Um so a lot more to say about that. The
[12:43] platform has a lot of other capabilities like persona extensions to agents which they can use to get adversarial reviews of of some of their strategies. We have a pretty sophisticated virtual file system which allows for agents to have
[12:57] their own uh file systems and they can interoperate with one drive and Google drive and even the local file system on the laptop because a lot of our customers actually use desktop applications that don't run in the cloud
[13:09] and they need to share files with their agents u directly with through their through their browser basically. Um but um I wanted to also talk about missions because this is really where things are are going. So the first phase of
[13:22] deployment of of analytics will be conversational. It's the things that you've seen. A lot of our users use it to also schedule reports of course, but what you really want is something that is fully autonomous. Um and that's what
[13:34] we have missions for. So missions are basically a team of agents that have a forpurpose prompt uh and a for purpose goal. The the mission we're looking at here is anyone here work with Walmart or
[13:48] know anything about Walmart's business? Uh yeah, a little bit. Okay. So Walmart is very openly stated that they're competing with with Amazon and that they have asked all of their vendors to increase their digital sales over the
[14:00] next year to 25%. as a vendor to Walmart, if you're going to be able to comply with that, you have to figure out what are the things that I can do to either grow my digital sales
[14:12] or at least cut the draggers, the things that hold me back from from gaining digital uh digital sales. Um, so what this mission is doing is that it has sets the context, all the KPIs,
[14:25] references to the semantic model, what's going on, what's going on here. We feed it with what we call a a base query. Instead of having an agent just get a prompt, hey, go figure out how I can grow my digital sales, we say, here are
[14:38] all the metrics that matter that are relevant for this KPI of driving digital sales. So, it's things like, of course, your total sales, your in stock percentages, your replenishment in stock. These are a lot of the business metrics. This is the kind of the
[14:51] starting query that will be run towards the semantic model whenever an agentic run is initiated. Then we have um KPIs. We have a primary KPI which in this case it's a digital transa the the um uh
[15:09] I was confused. I must have changed this by accident here at some point. Uh so it's the it's the u the KPI of um your digital sales and then we have what we call finding profile. So this is used when triaging problems. So the agent will detect changes in the data set and
[15:21] it will figure out what kind of problem is this. Is it a an instock problem? Is it a a content problem online where we don't have the right uh uh images available? Uh is it that the the digital
[15:33] shelf has allocated too little of the physical inventory in the store because Walmart uses stores as physical uh fulfillment centers for example. So this gets routed to um oh sorry I was on the wrong mission that's why. There we go.
[15:46] Um this gets routed to the um to the right playbooks. So instead of have giving aids, okay, you can do whatever you want after here. You have to give them actually ideas of what is it I can do
[15:58] about this. So if it's a a problem with with online content, then there might be a content fix packet that should be sent out. If the problem is that uh we're out of stock, maybe you need to send a push
[16:10] order uh to the distribution center. If it's a problem with the allocation between the digital shelf and the physical uh shelf, you have to email the buyer or the replenishment manager at Walmart. So currently, unfortunately, in
[16:24] the retail space, most of these playbooks end up with sending some sort of email to someone because that's just how this industry works currently. But we know that Walmart especially, they're working on making all of these types of
[16:37] interactions between the the merchant side and the CPG side automated probably through aic interfaces as well. So right now we're just kind of teeing up identifying all of the issues seeing all the opportunity and helping structure
[16:49] the old school way of communicating. But as the next 12 months roll forward, we'll start to see that these these uh playbooks will end up with actual uh API calls or agent to agent calls coming
[17:01] through. Um the this is mission studio which is basically our back office designer for designing these missions. Our goal is to deliver as many of these missions basically out of the box that
[17:14] are related to Walmart has 12 different dimensions. So 12 different things that uh are like out of the box missions that a a CPG can just deploy and get value from out of the box. But then it can also of course but as they get more
[17:26] advanced and get more AI literate they can start leveling up and building their own missions as well. And then the same thing we do in the target ecosystem the same thing we'll do for the for the Amazon ecosystem.

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.