The three ways AI unlocks transformation in Retail, Travel, and Consumer Goods

The three ways AI unlocks transformation in Retail, Travel, and Consumer Goods | Databricks Blog

Skip to main content

*Retail, CPG, and travel lose value to the same gap between signal and action, taxed three ways: trust, time, and cost. Forrester estimates 60% to 73% of enterprise data goes unused for analytics, and this piece argues AI deployed as a system, not a collection of pilots, is the first technology built to attack all three taxes at once.

*Three movements turn dead signal into action: dark data becomes signal, probabilistic reasoning expands the question space, and human capability decouples from headcount. Proof points already in market include Harmons cutting out-of-stocks by more than half through shelf-scanning, Airbus Skywise reading telemetry across 12,000+ connected aircraft, and Walmart's Ask Sam giving first-week associates a veteran's recall.

*Governance, not the model, is the real competitive moat. The Databricks Platform lets enterprises widen the aperture, prove trustworthy reasoning, and act at machine speed on one governed foundation, positioning them for a shift McKinsey projects could redirect $3 trillion to $5 trillion in global retail spending to AI agents by 2030.

It is 5:45 on a Friday morning, and a store manager is standing in the back office losing an argument with her own building. Fourteen dashboards are open on two screens. A task queue, assembled overnight by four systems that have never spoken to one another, wants her attention in three different orders of priority. Somewhere on aisle seven, the highest-velocity SKU in the store has been out of stock since Tuesday. The shelf camera saw it Tuesday. The replenishment system inferred it Wednesday. The task to fix it surfaced this morning, ranked eleventh, beneath a planogram audit and a training reminder. She will find it on her walk, the way managers have found things for a hundred years, and by then the weekend traffic the product was ordered for will have come and gone.

The signal existed for four days. What never existed was a way to act on it before the moment passed.

That morning repeats itself, in different uniforms, across every consumer industry. Forrester has estimated that between 60% and 73% of all data inside an enterprise goes unused for analytics. In consumer goods, roughly 85% of new product launches fail within two years, and Nielsen attributes the majority of failures not to product quality but to misread consumer needs and positioning, which is to say, to signals that existed and went unread. And when the NVIDIA 2026 State of AI in Retail and CPG survey asked companies what blocks them from scaling AI, the most-cited barrier was not the technology. It was data. These are not three problems. They are one problem, taxed three ways: you could not believe the signal enough to act on it (trust), you could not act before the moment passed (time), or you could not afford to read and act at the scale that mattered (cost). The claim of this piece is that AI, deployed as a system rather than a collection of pilots, is the first technology that attacks all three taxes at once, and that the companies who understand this are about to separate from the companies who bought the pilots.

Why Twenty Years of Dashboards Did Not Close the Gap

The industry’s answer to the unread signal was business intelligence, and it was a reasonable answer. BI professionalized reporting, standardized the metrics that boards run on, and gave a generation of merchants and operators a shared factual ground. It deserves its place. But it carried three structural limits that no amount of investment could engineer away.

First, BI could only see data that fit a schema. The shelf image, the maintenance photo, the call transcript, the review, the social post: the majority of what a consumer enterprise actually witnesses never made it into the model, because reading it required human attention, and human attention was the scarcest resource in the building. Second, BI answered questions someone had already thought to ask. It computed brilliantly across the joins it was given; it could not form a view across signals nobody predefined, which is precisely where the surprises live. Third, BI ended at a dashboard. An experienced human still had to notice, interpret, decide, and route, so insight traveled at the speed of escalation, and escalation traveled at the speed of the org chart.

What changed is that each of those limits has now been removed, separately and recently. Foundation models collapsed the cost of reading unstructured data, so the witness statements finally enter the record. Probabilistic reasoning made judgment computable, so the question space stopped being limited to what an analyst anticipated. And the agentic turn decoupled acting from staffing, so the distance between knowing and doing stopped being measured in meetings. The three taxes, for the first time, have a counterparty.

Where the Signals Have Been Dying

In retail, they die on the store floor. Our work on AI-powered store operations keeps returning to the same finding: the store is the richest signal environment in the enterprise and the least able to metabolize it. Shelf conditions are known only by walking. Labor plans are built on last year’s curves. Waste is counted after it is in the bin. And the manager is the single, overloaded integration point for a dozen systems that were never designed to converge on one human at 5:45 in the morning. The loop is now arriving where it was needed first: autonomous shelf scanning at Harmons cut out-of-stocks by more than half and pricing errors by roughly 75%, and Walmart’s Ask Sam gives a first-week associate a veteran’s recall on the floor. What is emerging is the store that runs its own loop, sensing shelf, queue, and waste continuously and acting within the hour, so the manager’s walk becomes judgment applied to exceptions rather than discovery of the obvious.

In consumer goods, they die in the innovation engine. The traditional stage-gate model runs sequential human reviews on physical prototypes, and every gate is a place where signal expires: the reviews that named the flaw went unread, the early sell-through that predicted the miss arrived after the production commitment, the social trend that justified the concept surfaced after the shelf reset. My recent work on innovation in this industry points to an inversion now underway, from sequential human gates fed by stale research to a continuous computational read of the consumer, running before any physical commitment is made. The public evidence is arriving: one global health and hygiene company unified the consumer research, product reviews, and social listening signals its marketers once spent weeks combining, and now moves from raw signal to actionable insight in days, by its own account reclaiming the hours its teams had been losing to assembly rather than judgment, the hours innovation was starving for. The binding constraint on that transformation is not model capability. It is proprietary consumer data, current, governed, and dense enough to reason against.

In travel, they die between systems that each know half the story. The telemetry knew the component was degrading; the crew scheduler did not. The fare engine knew demand was shifting; the disruption desk found out from the queue. The industry is further along than most in wiring these together: predictive maintenance reads signals across more than 12,000 connected aircraft on Airbus Skywise, Delta prices its network with generative models, and AI concierges are in market at Expedia and Hilton. Run those forward and connect them, and you get the trip that repairs itself: the disruption detected in maintenance data, the itinerary rebooked, the hotel and ground transport adjusted, the traveler informed of the problem and its solution in the same message. IDC projects that by 2030, up to 30% of travel bookings may be executed by AI agents, and the operators whose data can support that autonomy will own a disproportionate share of it.

Across all three industries, the pattern is identical. The models are not the moat. The data is the moat.

The Three Things AI Unlocks

The pattern has a shape, and it is the lens I use with executives across these industries. AI’s assault on the three taxes proceeds in three movements, and each movement is the inversion of one of BI’s structural limits.

Dark data becomes signal. The attention-cost of reading unstructured data collapses, and the majority of what the enterprise witnesses, images, transcripts, telemetry, chatter, finally enters the record.

Probabilistic reasoning expands the question space. The calculator becomes judgment. A model forms a probabilistic view across signals it was never explicitly coded for, weather against promotions against a competitor’s fare change, and the surprises stop hiding between the joins.

Human capability decouples from headcount and experience. Augmentation gives the first-week associate a veteran’s recall; automation, the limit case, removes the execution delay entirely. The distance from knowing to doing stops being measured in meetings.

The three movements run as a loop against time; real time is the clock; governance is the substrate that earns trust and controls cost.

Movement

Deployed now

Emerging

The tax it attacks

Dark data becomes signal

Shelf-scanning robots cutting out-of-stocks by more than half at Harmons; Unilever fusing POS with weather and multilingual social listening; Airbus Skywise reading telemetry across 12,000+ aircraft

Agent-based demand sensing that reads raw social, reviews, and sensor streams before transactions confirm the trend

Time and cost: human attention stops being the price of reading your own signals

Probabilistic reasoning expands the question space

Walmart’s AI-assisted forecasting, on the record with its COO; Delta’s generative network pricing; a global health and hygiene company reasoning across unified consumer research, reviews, and social signals in days instead of weeks

Continuous willingness-to-pay pricing per shopping request; today only about a quarter of airline offers are dynamically created

Time and trust: the question no longer waits for someone to predefine it, and evaluation makes the answer believable

Human capability decouples from headcount and experience

Walmart’s Ask Sam on the store floor; L’Oréal’s Beauty Genius advisor, 1.1M+ U.S. conversations since launch; Sweetgreen’s Infinite Kitchen running ~50% above a standard make-line

Agentic execution of the transaction and the task itself, narrow today, widening fast

Time: augmentation removes the escalation delay; automation removes the execution delay itself

Each movement is an amplifier, and amplifiers are indifferent to what they amplify. Aperture alone produces more noise. Reasoning alone, on ungoverned data, produces false confidence. Acceleration alone produces faster mistakes. Signal without reasoning is noise; reasoning without signal is false confidence; acceleration without either is a faster mistake. Put all three on the same governed data, in real time, and you get something that did not exist before.

This is why governance, the least glamorous line item in the data budget, turns out to be the payload. Governing data once pays a dividend twice over: you can believe it, which is the trust tax repealed, and you never pay to derive it again, because every model, agent, and application draws on the same processed, certified foundation instead of re-extracting the same corpus, which is the cost tax repealed. Governance is chronically undervalued because it is booked as compliance while its role as the cost lever goes unmeasured.

The Architecture Underneath

The operating posture described above asks one architecture to do three things at once: run the sense-to-action loop in real time, prove the trustworthiness of every answer it produces, and do both without the enterprise paying to process the same data twice. That is an architectural specification, and it is the one the Databricks Data Intelligence Platform was built to meet.

Widening the aperture is an ingestion and openness problem. Lakeflow Connect brings more than 100 managed connectors across enterprise applications, databases, and streaming sources into Delta Lake and Apache Iceberg, open formats that keep the signal in storage you own rather than a proprietary engine you rent.

Trustworthy reasoning is a governance and evaluation problem. Unity Catalog governs data, models, and agents in a single layer, Unity Catalog Metrics makes business definitions first-class governed assets so every agent answers from the same meaning, and MLflow 3 supplies the evaluation and tracing that make a probabilistic answer auditable, the gap that governed data alone cannot close. Unity AI Gateway extends that same governance to model and agent traffic itself.

And acting at machine speed is an application problem: Agent Bricks builds evaluated agents from enterprise data, Genie puts governed answers directly in the hands of the store manager, the planner, and the ops center, and OpenSharing and Clean Rooms extend the same governed signal across the value chain, supplier to retailer to partner, without copying the data or surrendering control of it.

The consequential point is this: partial adoption of the loop is not partial value. An amplifier stack with one stage missing amplifies the wrong thing. The platform decision is where an enterprise chooses whether its three movements point the same direction, on the same governed substrate, or whether it assembles false confidence at increasing speed.

The Case Studies Are Being Written Now

And the deadline is no longer purely internal. The same discipline is about to be examined from the outside: consumers are beginning to delegate shopping, booking, and buying to AI agents running on protocol rails that shipped over the past year, and McKinsey projects that agent-mediated commerce could redirect $3 trillion to $5 trillion in global retail spending by 2030. Those buying machines will read only what governance has made legible. The enterprise that closed the gap between its data and its decisions will discover it also built the only storefront a machine can read; governance’s two dividends quietly acquire a third.

Thirty years ago, established retailers faced a question about the web: channel or threat, resist or embrace. The companies that treated e-commerce as a new physics rather than a new storefront spent the next two decades being studied by everyone else. The same divergence is forming now, and it will not run between the companies with AI and the companies without it. It will run between the companies that metabolize what they already know and the companies that keep paying the three taxes on it.

The preparation is knowable: data that is current, governed, and semantically coherent, reasoning that is evaluated, action that is instrumented. The signal on aisle seven has been waiting a hundred years for someone to read it in time. The discipline was never overhead. It is the difference between a company that knows and a company that acts.

Get the latest posts in your inbox

Subscribe to our blog and get the latest posts delivered to your inbox.

Sign up

View all blogs