Here is a number that should stop any retail leader mid-scroll: 35% of e-commerce revenue now comes from personalized recommendations. Roughly a third of every dollar you make online is already being shaped by an algorithm deciding what a shopper sees next. That is not a future retail is bracing for. It is the floor everyone is already standing on.
And yet most retailers are stuck. Not because they lack the ambition, and not because the models are weak. The gap sits somewhere less glamorous: the data. As of January 2025, over 80% of CPG and retail businesses are utilizing or testing generative AI technologies, and about 90% are expected to invest more in them in the coming years. But piloting is not shipping. The hard part was never the demo. It was building the version that runs on real customer data, under real permissions, with real consequences when it gets something wrong.
That is the specific problem Databricks Mosaic AI for Retail was built to close. This piece walks through what it does for retail teams, where it earns its keep, and why the winners over the next two years will be the ones who treat AI as a data problem first.
Why Retail AI Projects Stall Before Production
If you inquire about the fate of AI experiments among retail data experts, the answer will likely never be “the model.” Oftentimes, it is one of the following three: data butting heads with silos, inconsistent definitions of the terms “customer” and “return,” and lack of supervision in transcending notebook trials into operational applications.
The cost of that stall is measurable. There is a real perception gap running through the industry: 71% of retail leaders believe they are effective in personalizing their offerings, while only 34% of consumers share this belief. That distance is not a creativity problem. It is what happens when personalization runs on stale, disconnected data that never sees the full customer.
The payoff for closing it is just as concrete. According to McKinsey study, AI personalization may increase revenue by 10 to 15% and more among leading companies. Companies benefitting from it do not have necessarily the best solutions on the market. They are the ones whose data was ready when the model showed up.
What Is Mosaic AI for Retail
Table of Contents
Mosaic AI is essentially a built-in component of the Databricks Data Intelligence Platform that consists of generative as well as machine learning layers. To put it simply, what Mosaic AI does is eliminate the necessity of acquiring different components from different vendors, and creates an AI related environment wherein all the components work together.
That is all there is to it. Since all the components in question share the same background as far as permissions and lineage are concerned, there is no possibility whatsoever of the AI agent gaining access to information which is not authorized for use by a human being. This situation is particularly vivid in retail, where the recommendation engine uses not just purchase information, but also loyalty data as well as pricing data. It is what lets the thing go to production at all.
The scale tells you it is past the experiment stage. As per Databricks, they suggest that there are around 100,000 agents that have been created using these Agent Bricks tools and they process a massive amount of data exceeding a trillion tokens per year.
Mosaic AI Use Cases for Retail
The temptation with any AI platform is to list forty capabilities. It is more useful to name the handful where retailers see returns quickly.
Demand forecasting and inventory. This is the workhorse. Mosaic AI runs time-series forecasting at the store-SKU level, trained directly on governed data and blended with external signals like weather and local events. According to Databricks’ own reports, one can save 20 to 30% of excess inventory and see improvements in stock availability by 15 to 25% with artificial intelligence-enabled forecasting. The largest retailer in the Netherlands, Albert Heijn, has been carrying out over one billion forecasts daily for 1,200 stores and managed to reduce waste by 3.6 million kg within one year after implementing the solution.
AI-Powered Retail Personalization. Because recommendation models train on the same Gold-layer data feeding the rest of the business, the personalization a shopper sees reflects today’s behavior, not last week’s batch job. Given that recommendations already drive up to 31% of revenue in engaged sessions, the freshness of that data is not a technical footnote. It is the difference between a relevant suggestion and a wasted impression.
Databricks Mosaic AI Agents for Retail Operations. RAG-based assistants answer questions. Agents built with the Mosaic AI Agent Framework can plan, query governed tables, call APIs, and loop until a task is done, whether that is a forecast explainer for a planner or a product recommendation agent for a shopper. One retail analytics team discovered that the usage of MLflow tracing helped them reduce the debugging time with a faulty pipeline from almost three days to just four hours just because they could finally see which documents the agent had fetched and for what reason.
The teams that see these results fastest are the ones that fix the data foundation before chasing the use case. In one retail engagement, Polestar Analytics rebuilt the estate on Unity Catalogue with MLflow versioning before layering customer intelligence on top, cutting cloud cost on offer-model pipelines by 75% and runtime by 40%.
What Retailers Need to Build Before Deploying Mosaic AI
Here is the pattern worth internalizing: most of the work happens after the model works. A mid-size electronics retailer built 60% of their agent frontend before discovering their product catalog fields were not structured for their vector search to serve efficiently, which cost them two weeks of refactoring. The lesson is not subtle. Clean the data first, then build.
This is also where cost discipline lives. Databricks Mosaic AI runs on consumption-based pricing measured in Databricks Units, and those bills get unpredictable when workloads run unchecked. The platform lets you route cheap, high-volume steps to a small model and reserve a frontier model for hard reasoning, turning what used to be a rewrite into a config change. That routing is not a technical curiosity. On a catalog with millions of interactions a day, it is the line between a system that pays for itself and one that quietly leaks budget.
This is where an experienced Databricks partner earns its place. Polestar Analytics scopes retail engagements to start with one governed medallion pipeline and a single use case rather than an estate-wide rebuild, which is what keeps a project moving from pilot to production instead of stalling.
Why Mosaic AI Matters for Retail AI at Scale
Retailers are not short on AI ambition. Almost 90% of retailers are making use of AI and the overall market reached $18.4 billion mark as of 2026. What makes the difference between teams generating tangible income and teams that are stuck in pilot mode is not the model in use. What makes the difference is the quality of the data used by the company and whether it was clean, combined, and updated enough for the model to operate effectively.
Mosaic AI, however, is not a magic key and cannot fix any data quality problems on its own. What it does is remove the excuse. When retrieval, models, and governance all sit on the same data your business already runs on, the distance between a promising pilot and a production system that moves the numbers gets a lot shorter. The retailers who win the next two years will be the ones who close that distance first.
Frequently Asked Questions About Databricks Mosaic AI for Retail
Q: Do I need Mosaic AI to do retail personalization, or can I build it myself?
Building the model by yourself isn’t the issue, but integrating retrieval, models, and governance of the data is. Mosaic AI gets rid of that integration, resulting in personalization based on today’s activities rather than last week’s output.
Q: Why do so many retail AI projects stall after the pilot?
Rarely because of the model. They stall on messy data, inconsistent definitions of “customer” or “return,” and no clean path from a notebook experiment to production. The teams that win are the ones whose data was ready when the model showed up.
Q: How does Mosaic AI keep costs under control at retail scale?
It runs on consumption-based pricing, and lets you route cheap, high-volume steps to a small model while reserving a frontier model for hard reasoning. On a catalog with millions of daily interactions, that routing is the line between paying for itself and quietly leaking budget.
