Sammanfatta detta blogginlägg med:
The majority of 2025 saw shoppers visiting an online shop through ChatGPT or its equivalent and accomplishing the least purchases. They looked through products and left empty-handed. Retailers watching the numbers reasonably concluded that AI was a research channel, not a sales one.
After the holiday season everything turned upside down. As per Adobe Analytics, which relied on data from over a trillion visits to US online retail, AI referrals proved to be at least 31% more effective than any other channel of acquiring customers. Revenue per visit from AI traffic climbed 254%. On Thanksgiving the gap hit 54%.
The shopper who used to be window-shopping now arrives ready. They have already compared, already narrowed, already asked something specific. Sometimes an agent has done that work for them. Gartner at its 2025 IT Symposium predicted that AI channels would help transact over $15 trillion in B2B sales by 2028, reflecting activity across both channels rather than just consumer-level buying.
This changes the retailer's central question. It stops being "how do we rank in search" and becomes something harder to fix: when a shopper or their agent asks for something specific, can our data answer well, right now? This is precisely the gap Databricks Mosaic AI for Retail was built to close.
Retail runs on fragments. Point-of-sale logs sit apart from e-commerce clickstreams. Loyalty records live in their own system. Inventory updates arrive on a different clock than everything else.
A recommendation engine that cannot see all of it at once produces confident nonsense. It pushes an item that sold out an hour ago. It greets a ten-year customer like a stranger. The model is rarely the problem; the model was reasoning from a partial view.
What that costs is now well measured. IHL Group's September 2025 study puts global inventory distortion — the combined cost of out-of-stocks and overstocks — at $1.73 trillion a year, or 6.5% of global retail sales. Roughly the GDP of South Korea, lost to not knowing what is on the shelf.
The more pointed finding sits further down that report. Retailers deploying AI and machine learning against the problem are achieving sales growth 2.3 times higher and profit growth 2.5 times higher than competitors who are not. Yet fewer than a quarter of retailers have successfully rolled out AI in the areas inventory distortion hits hardest. IHL's president Greg Buzek frames it bluntly: evolve or get left behind.
So appetite is not the constraint. McKinsey's State of AI, published November 2025, found 88% of organizations regularly using AI in at least one function. Money is not the constraint either. The constraint is the distance between a model that works in a notebook and a model running in production against live, governed data. Nearly two-thirds of organizations in that same survey have not yet begun scaling AI across the enterprise at all.
Databricks Mosaic AI is part of the Databricks Data Intelligence Platform, which has just been branded as Databricks AI. It does not refer to a model. It is the platform which enables retailers to make use of their own data when creating a number of applications, including model serving, tuning, AI Search for retrieval, and constructing an agent framework, with all such processes being regulated via the Unity Catalog, ensuring that all models are correctly traced and monitored in production.
Source: Databricks
Two things matter most in retail.
The first is that the AI agents for retail reason over your unified data rather than generic web knowledge. An agent that knows your SKU hierarchy, your regional demand patterns, and your actual stock position answers a question differently than one working from the open internet.
The second is the agent framework, which lets teams build agents that retrieve live inventory, call pricing APIs, and act, rather than just talk.
The engineering underneath has moved fast. Databricks rebuilt AI Search on a storage-optimized architecture that scales to billions of vectors at up to 7x lower cost, with billion-vector indexes now building in under eight hours. The published pricing comparison is stark: roughly $7,000 a month for 1.3 billion vectors, against roughly $47,000 on the standard offering. Model serving now handles more than 250,000 queries per second.
Those numbers matter for one practical reason. Semantic search across a full product catalog used to be economically unreasonable at retail scale. During a Black Friday spike, throughput is the difference between real-time and too late.
Here are the core Databricks Mosaic AI retail use cases teams are deploying today:
- Personalization and product discovery. Agents read browsing history, purchase patterns, and context to handle messy human requests. Someone types "a gift for my sister who loves cooking and has a small kitchen" and gets space-saving, genuinely relevant results rather than a keyword dump.
- Agentic commerce. Discovery, recommendation, and transaction handled on the shopper's behalf, with real-time stock and price signals pulled at the moment of decision. This is where agentic AI for retail stops suggesting and starts doing.
- Dynamic pricing and inventory. Agents weigh demand, competitor pricing, and margin across thousands of SKUs continuously, rather than waiting for a weekly review cycle that was always a compromise with reality.
- Customer service. Service agents resolve "where's my order" against live order data, offer real alternatives when something is out of stock, and escalate genuinely complex cases with full context attached.
- Supply chain and demand forecasting. Multi-agent systems coordinate across warehouses, distribution centers, and stores, reading demand signals and supplier risk to move product before a stockout forms rather than after.
The compounding is what matters more than any single use case. Because these agents share one governed foundation, the pricing agent and the personalization agent read the same truth. That is the difference between agentic AI and a scatter of disconnected bots that each know a little and quietly disagree with each other.
The same governed foundation powers autonomous pricing and promotion decisions, not just recommendations. See how to orchestrate revenue growth agents on Databricks Mosaic AI.
These are Databricks AI solutions running in production:
Adidas built a RAG chatbot to analyze more than 2 million product reviews. Latency dropped 60%, cutting average response time from 15.5 seconds to six. Moving to more efficient models cut compute costs 91.67%. Analyst efficiency on review-based decisions improved 30 to 40%, and more than 50 decision-makers across product, design, and marketing now pull insight directly instead of queuing for it.
Trek Bicycle migrated to a warehouse that could scale economically. Retail applications that could take 48 hours now take only six to eight hours — 80–90% faster — along with refreshing the Lakehouse three times daily instead of once. Before that, Trek performed analytics only once a day in North America, which meant other regions received reports late.
FOX Sports created an AI-powered search feature using the Databricks AI Search platform for Super Bowl LIX. The feature handles thousands of requests at once, delivering two times better search results than before. As of June 2026, those who used the Sports AI platform spent roughly twice the time inside the mobile app.
Reckitt built a GenAI insight engine unifying consumer research, reviews, and social signals across more than 60 countries, cutting campaign analytics that once took weeks by 40%.
A note on aggregate ROI figures: treat them carefully. Most circulating numbers come from vendor-commissioned surveys, and the spread between them is wide enough to suggest the methodology is doing a lot of work. The named case studies above are auditable. The survey averages are not.
The gap compounds, which is the uncomfortable part. McKinsey's 2025 survey found 23% of organizations already scaling an agentic AI system in at least one business function, with another 39% experimenting. Every month a competitor spends learning on live traffic is a month of iteration a late entrant cannot buy back later at any price. Models improve on production feedback, and production feedback only comes from being in production.
McKinsey also found high performers are roughly three times more likely to have scaled AI agents across the enterprise. That is not a technology gap. It is an operating-model gap, and those close more slowly.
None of this demands a moonshot. There is a need for managing data in a systematic way and formulating a predefined use case — normally personalization or customer service — that yields results early and obviously enough to invest in a new case. The process is more about creating reliable grounds for trusting the information used for decisions than about chasing the most impressive model.
Mosaic AI isn't a model, it's the layer that turns your own data into working applications, with model serving, retrieval, and an agent framework all governed through Unity Catalog. The difference is that agents reason over your unified data (your SKU hierarchy, real stock positions, regional demand) rather than generic web knowledge, so answers reflect what's happening in your business right now.
A recommendation engine suggests; an agent acts. In agentic commerce, discovery, recommendation, and the transaction itself are handled on the shopper's behalf, with real-time stock and price signals pulled in at the moment of decision. It's the shift from AI that points you toward a product to AI that completes the purchase.
The hardest gap isn't appetite or budget, it's the distance between a model that works in a notebook and one running in production against live, governed data. Nearly two-thirds of organizations haven't scaled AI across the enterprise at all, and the ones pulling ahead treat it as an operating-model challenge.