How Tesco's Clubcard Data Advantage Cracked the Agentic AI ROI Problem McKinsey Healthcare Hasn't

24 million Clubcard households. That is the behavioral database Tesco has been building since 1995, one shopping trip at a time, across three decades of purchase decisions, seasonal patterns, and brand substitutions. This week, MarketingWeek reported that Tesco is now deploying agentic AI across all of it -- and the results are already measurable: record customer satisfaction scores alongside growing profits, numbers that need no pilot-study caveats.

Meanwhile, McKinsey just documented that 50% of US healthcare organizations are now implementing generative AI, with the industry's attention shifting from deployment to "integration, ROI, and agentic AI." That phrasing deserves a close reading. Three years into the generative AI era, healthcare has moved from "what can we build" to "how do we make it worth what we spent." Tesco appears to have made that transition quietly, ahead of schedule, and without issuing a single press release about it.

The divergence is not primarily about AI capability. It is about how often the data underneath it generates a new signal.

The McKinsey Healthcare Reality: Pilots Without Feedback Loops

McKinsey's new healthcare AI report documents genuine progress. Half of US organizations are running real deployments, not just evaluations. The use cases have matured past novelty: clinical documentation automation, prior authorization support, radiology workflow assistance, patient triage. These are operationally significant applications.

But the report's framing of where the industry is going -- integration, ROI, agentic AI -- implicitly describes where it has not yet arrived. Healthcare AI has demonstrated capability. It is still struggling to demonstrate durable, measurable value at the system level.

The structural reasons are worth understanding clearly. Electronic health records contain extraordinarily rich longitudinal data: lab values, imaging, medication histories, clinical notes, genomics in some cases. The depth per patient dwarfs anything a grocery chain captures. But depth is not frequency, and frequency is what agentic AI needs most.

The average US adult makes approximately 3.5 primary care visits per year. Add specialist visits, urgent care, and lab draws, and the number climbs -- but the behavioral signal between those visits largely disappears. Patients fill prescriptions inconsistently. They follow discharge instructions at varying rates. They sleep, move, and eat in ways the EHR never sees. Healthcare AI models are personalizing interventions across a dataset that is structurally sparse in the time dimension, regardless of how deep it runs in the clinical dimension.

Why Agentic Systems Need Dense Observation Windows

Agentic AI is distinct from earlier AI deployments in a specific way: it acts, observes an outcome, and adjusts. The adjustment loop is only as fast as the observation window. Deploy a clinical recommendation. The patient sees it, possibly acts on it, and the outcome gets documented weeks or months later -- if it gets documented at all. The model learns from this at the pace of clinical encounters.

Retail closes this loop in days. Healthcare closes it, at best, across clinical visits measured in months. Agentic systems deployed in each sector are training on fundamentally different signal densities, and that gap compounds over time with every iteration cycle completed.

What Tesco's "Record Satisfaction" Numbers Actually Reveal

When Tesco reports "record" customer satisfaction alongside accelerating AI investment, the specific mechanism matters. The MarketingWeek report describes agentic personalization being deployed across 24 million Clubcard households -- not piloted with a test segment, but running at operational scale. The satisfaction improvement is a system-wide outcome, not a controlled experiment result.

This distinction is significant. Healthcare AI outcomes are typically measured in controlled trial conditions: a cohort using AI-assisted care compared to a cohort without it. Those trials are methodologically rigorous, but they measure capability under favorable conditions. Tesco is measuring outcomes in the least controlled environment possible -- the full complexity of 24 million real households making real decisions under real-world conditions including competitor promotions, economic pressure, and changing household composition.

The Clubcard mechanism provides something that no healthcare system currently has: weekly behavioral ground truth. A Clubcard member shopping twice a week generates 104 distinct behavioral observations per year -- each one a timestamped, itemized record of a purchase decision made under real conditions. The AI recommended a product substitution last Tuesday. By Thursday, Tesco knows whether it was accepted. By the following Tuesday, it knows whether the customer repeated the behavior.

Economically, this creates a compounding dynamic. Each feedback cycle makes the next recommendation marginally more accurate. Over 104 cycles per year, marginal improvements accumulate into measurable customer satisfaction uplift. Over 24 million households, marginal improvements aggregate into the revenue and margin numbers showing up in Tesco's financial results.

The Feedback Loop as Infrastructure

Tesco's agentic advantage is not architectural in the AI sense. It is architectural in the data sense. The Clubcard was designed as a loyalty mechanism, not an AI training pipeline -- but its 30-year accumulation of weekly purchase decisions has become, retrospectively, one of the most effective agentic AI training datasets in consumer-facing industries.

This was structural luck, not strategic foresight about AI. A business that happened to operate on high-frequency consumer interactions built the feedback infrastructure that agentic AI requires, before anyone knew agentic AI would matter.

The 25x Frequency Advantage Nobody Is Counting

Here is the original inference that does not appear in either source report. If we model agentic AI ROI as a function of feedback loop velocity -- how many closed observation cycles a system generates per customer per year -- Tesco's structural advantage over the median healthcare AI deployment is approximately 25-30x.

At two Clubcard shopping trips per week, Tesco generates roughly 104 behavioral observations per customer per year. A primary-care-anchored healthcare AI system generates roughly 3-4. That ratio -- 104 to 3.5 -- is not a technology gap. It is a domain architecture gap, a function of how often the underlying business operationally touches the customer.

This changes how enterprise AI investment claims should be evaluated. Healthcare has deeper per-event data than grocery. A hospital visit contains vastly more clinical information than a grocery receipt. But for agentic AI, depth without frequency produces systems that are sophisticated at the moment of contact and effectively blind between contacts. Tesco's model is shallower per event, but it is continuously updating. Over a calendar year, a Tesco customer's behavioral profile has been revised 104 times. A healthcare patient's AI profile has been revised perhaps 3 times.

The implication for capital allocation is counterintuitive. A less capable AI model running on a weekly feedback cycle will outperform a more capable model running on a quarterly one -- for the same reason that weekly compounding outperforms annual compounding at the same nominal rate. The frequency of the compounding period matters more than the sophistication of the model, past a certain capability threshold.

Where the Next Agentic ROI Winners Will Emerge

Sectors that will crack agentic AI ROI next are those that can compress their observation window. Pharmacy chains with daily pickup patterns already have frequency advantages over hospitals. Dental practices integrating remote monitoring tools are moving toward continuous signal capture. SaaS platforms tracking real-time usage telemetry can iterate on personalization daily. Financial institutions observing daily transaction behavior have frequency profiles that rival grocery retail.

The pattern is consistent: wherever a business generates behavioral observations more than weekly, the structural conditions for agentic AI ROI exist. Where those observations come monthly or slower, the ROI case requires closing the frequency gap before the capability gap.

What Healthcare AI Actually Needs Before Agentic Works at Scale

The path forward for healthcare AI is not more sophisticated models deployed on the same sparse signal. It is investment in signal density -- the infrastructure that generates more frequent behavioral observations between clinical touchpoints.

Wearable devices already generate the frequency healthcare needs. Continuous glucose monitors, heart rate variability tracking, sleep pattern data, and activity monitors create behavioral observation rates that approach or exceed Tesco's grocery frequency. The challenge is connecting this data to clinical AI systems in ways that are standardized, interpretable, and actionable within existing clinical workflows.

Remote monitoring programs and patient-reported outcome platforms represent another dimension of the same strategy. If a healthcare system can close a feedback loop in days -- prescription refill behavior, symptom check-ins, activity data -- it changes what agentic AI can realistically accomplish within the system. The AI becomes continuously updating rather than visit-triggered.

The pattern holds across sectors: enterprises that approached agentic AI most cautiously have also been slowest to generate the dense feedback loops that demonstrate ROI. Caution in healthcare is appropriate given clinical stakes. But caution also compounds the frequency disadvantage by delaying the iteration cycles that produce learning.

Retailers, operating with lower consequence per recommendation error and higher tolerance for rapid iteration, accumulated more feedback cycles faster. Their early lead reflects more iteration cycles completed, on shorter timescales, from a more favorable starting position in terms of touchpoint frequency.

The Strategic Question Every Enterprise AI Buyer Should Be Asking

The premium positioning problem we've analyzed before -- where organizations confuse AI capability with AI infrastructure -- applies directly to enterprise AI ROI cases. The organizations making the loudest claims about sophisticated AI are not always the ones generating the clearest ROI signals. The ones with the clearest ROI are the ones completing the most feedback cycles.

Tesco's record satisfaction numbers are evidence of this mechanism. The British grocer is not doing something uniquely innovative with AI. It is doing something structurally advantaged: pointing capable systems at one of the densest behavioral feedback datasets in consumer commerce and letting iteration cycles compound over time.

Healthcare will close this gap. The organizations that arrive first will treat wearables data, remote monitoring, and patient-reported outcomes not as supplementary features but as the primary feedback infrastructure that makes agentic AI worth deploying at system scale.

For any enterprise making AI investment decisions now, the diagnostic question is simpler than most vendor conversations suggest: how fast does your behavioral signal refresh? Where it refreshes weekly or faster, the structural conditions for agentic AI ROI exist. Where it refreshes monthly or slower, the investment case requires closing the frequency gap before the capability gap.

The same logic applies to any high-stakes decision with multiple lagging signals -- the moment to act is determined by when the signal density becomes sufficient, not by conviction alone. Tesco reached that density 30 years ago. Healthcare is building it now.

The first enterprise AI question should not be "what model should we use?" It should be "how short is our feedback loop?"

If your organization is building decision systems and wants to understand how signal frequency shapes AI ROI across your sector, our research platform models these feedback dynamics across more than 20 industries.