The bottleneck was never data or algorithms: building a learning utility for AI-enabled learning health systems Authors: Jiang Bian, Majid Afshar, Christina M. Scifres, Emily C. Webber, David Burton, David Vawdrey, Fei Wang, Genevieve B. Melton, Nigam Shah, Rachel E. Patzer, Yiye Zhang Published: 2026 ABSTRACT Although envisioned nearly two decades ago, the learning health system (LHS) remains largely unrealized at scale. The limiting factor is not data or algorithms but rather the failure to build, the technical, human, and cultural infrastructure needed to operationalize continuous learning. We propose the learning utility: the full institutional stack that turns data and AI into continuous, bidirectional, measurable learning at the speed of care. If data is the fuel and AI is now a new generation technology, learning is the electricity and the learning utility is the grid. The utility’s core mechanism is bidirectional learning flow between evidence and practice: evidence informs practice, while practice continuously generates data that refines evidence. Today both directions run at project speed, not utility speed. Building the utility therefore requires a three-layer infrastructure stack: Foundation (data/technology, heavily invested), Machinery (feedback loops and monitoring, the largest gap), and Enabling Environment (governance, trust, and alignment, largely absent). We argue for a measurable “GDP (Gross Domestic Product) of learning” as the alignment mechanism across stakeholders, operationalized through a learning compact that distributes obligations and supports across the ecosystem. The recommendation: invest differently, not necessarily more, starting with the translator workforce that bridges the three layers. Introduction: the promise, the progress, and the gap Nearly two decades ago, the Institute of Medicine (now National Academy of Medicine) articulated a vision for a learning health system (LHS), one that converts routine clinical data into best evidence, translates evidence into practice, and feeds outcomes back into the next cycle of learning at the point of care, with innovation emerging as a natural by-product of care1. The aspiration was clear: a healthcare system that learns as fast as it delivers care. Progress has come on multiple fronts: conceptual scaffolding has matured (e.g., Friedman et al. 20102; Foley et al. 20232; Welch et al. 20243); implementation science has advanced (e.g., the Consolidated Framework for Implementation Research 2.0 [CFIR 2.0]4, Non-adoption, Abandonment, Scale-up, Spread, Sustainability [NASSS]5, Reach, Effectiveness, Adoption, Implementation, Maintenance [RE-AIM]6,7,); national programs now support LHS training and operations (e.g., the Agency for Healthcare Research and Quality [AHRQ] / Patient-Centered Outcomes Research Institute [PCORI]8 LHS K12 and more recent LHS Embedded Scientist Training and Research [E-StaR] P30 Centers; the National Institutes of Health [NIH] Health Care Systems Research Collaboratory; the 25-year record of the Department of Veterans Affairs Quality Enhancement Research Initiative [VA QUERI]); and LHS-specific maturity models track organizational readiness (e.g., Lannon et al.‘s Network Maturity Grid9). Among these fronts, data infrastructure was the right first bet, and drew significant focus and concentrated investment. It was the first-order bottleneck: without shared, harmonized data, no learning cycle can begin. It was also the most tractable: amenable to standard software engineering, fundable through federal meaningful-use incentives (the Health Information Technology for Economic and Clinical Health [HITECH] Act, 2009), and aligned with the informatics and IT workforce already in place. The result is a more mature data layer, relative to other aspects of the LHS, where electronic health records (EHRs) are now near-universal; federated research networks like the National Patient-Centered Clinical Research Network (PCORnet) span 47 million patients across 13,000 clinical sites (https://pcornet.org); common data models like the Observational Medical Outcomes Partnership (OMOP) model have been widely adopted for observational research including synergies through the Observational Health Data Sciences and Informatics (OHDSI) collaborative communities; and interoperability standards like Fast Healthcare Interoperability Resources (FHIR) enable real-time data exchange. Yet the data and associated technical systems layer itself remains uneven, with quality, access, and governance gaps persisting. Haendel et al. recently proposed closing these gaps by governing health data as a public utility, infrastructure that serves the public good through distributed stewardship and public oversight10. Despite this multi-front progress, particularly with data and technical infrastructure, the LHS remains largely unrealized at scale. Translation from evidence to practice with effective uptake is still long, estimated an average of 17 years11,12. The LHS cycle also rarely unfolds cleanly: competing interventions, evolving workflows, and shifting organizational priorities accumulate along the way, making it difficult to attribute improvement to any single learning cycle. “Pilot-itis” persists. As such, promising technologies or clinical programs cycle through grant-funded pilots without becoming sustainable13. And, while pockets of durable success exist like Intermountain Healthcare’s clinical-program model14, the VA’s 25 years of QUERI-anchored learning15, and registry-based learning networks like the Vermont Oxford Network for neonatal care16, each remains confined to a single institution or delivery system or to a particular speciality or use case, and none has generalized into field-wide sustainability with sharable and repeatable infrastructure. In parallel, the AI landscape is scaling rapidly and shifting what knowledge generation means. The first decade of AI in healthcare was almost entirely predictive (e.g., risk scores, early warning alerts, image classification). Generative AI introduces new cognitive functions extending not only to prediction but also synthesis (ambient documentation), reasoning (structured trial matching, differential generation), and generation (patient-facing communications, discharge summaries). These solutions represent a transformative and qualitative expansion of the clinical workflow in contrast to the incremental gains from traditional AI systems. The deployment scale of these solutions is already significant: 71% of US hospitals run at least one EHR-integrated predictive AI tool17; 62.6% of Epic-using hospitals have adopted ambient AI18 (e.g., Kaiser Permanente’s Abridge deployment covered 2.5 million encounters across 7,260 physicians in its first year19); and autonomous agentic systems capable of multi-step clinical reasoning are emerging without clearly established regulatory and deployment pathways20,21. Generative AI also democratizes participation: clinicians without heavy quantitative training can now design and evaluate tools that expedite care. But scale is not the same as sustained impact. Only 26 of 1247 AI studies met criteria for hospital implementation22 - and the few that deploy often go unmonitored—the Epic sepsis model missed 67% of cases in external validation, a failure surfaced through investigative journalism rather than systematic monitoring23. The pattern echoes LHS’s own history: scale without feedback, deployment without monitoring, and capability without the infrastructure to sustain or improve it. What has stalled LHS for two decades is shaping AI deployment right now. Why have both LHS and AI streams fallen short in broad healthcare impact? The answer lies in a structural misunderstanding of what they require. Data is the fuel, AI is the generation technology, but neither produces power on its own. Generating electricity requires more than fuel and generators: a grid, generating capacity, and the regulatory structures that make the system run. The analogous infrastructure for healthcare transformation, which includes feedback loops, governance, workforce, and trust, has not been built at scale. The distinction is critical because putting AI and LHS together is not the same as building an AI-enabled LHS. An AI-enabled LHS requires two feedback loops operating in parallel: practice-level feedback which refines workflows, clinician response, and clinical protocols as outcomes accumulate, and model-level feedback that retrains and recalibrates the AI solution itself on real-world data. Without both loops, AI may remain a predictive tool, but it does not become part of a learning system. An AI-enabled LHS is nonetheless beginning to take shape in isolated exemplars.Kaiser Permanente’s Advance Alert Monitor. This solution demonstrates the practice-level loop: prediction → clinical rescue → outcome, with surveillance cadence refined from 6-hour to hourly and workflow standardization across 21 hospitals, achieving a 16% relative mortality reduction across 548,838 hospitalizations24,25. The model itself was not, however, retrained on deployment outcomes.HERACLES, a sepsis AI-LHS at Lausanne University Hospital. This LHS closes both loops with the model continuously retraining, a real-time dashboard surfaces performance and outcome indicators, and human-in-the-loop validation refines predictions, producing measurable before-and-after improvements in sepsis detection and time-to-antibiotics across 97,559 patient stays26. Kaiser Permanente’s Advance Alert Monitor. This solution demonstrates the practice-level loop: prediction → clinical rescue → outcome, with surveillance cadence refined from 6-hour to hourly and workflow standardization across 21 hospitals, achieving a 16% relative mortality reduction across 548,838 hospitalizations24,25. The model itself was not, however, retrained on deployment outcomes. Kaiser Permanente’s Advance Alert Monitor. This solution demonstrates the practice-level loop: prediction → clinical rescue → outcome, with surveillance cadence refined from 6-hour to hourly and workflow standardization across 21 hospitals, achieving a 16% relative mortality reduction across 548,838 hospitalizations24,25. The model itself was not, however, retrained on deployment outcomes. HERACLES, a sepsis AI-LHS at Lausanne University Hospital. This LHS closes both loops with the model continuously retraining, a real-time dashboard surfaces performance and outcome indicators, and human-in-the-loop validation refines predictions, producing measurable before-and-after improvements in sepsis detection and time-to-antibiotics across 97,559 patient stays26. HERACLES, a sepsis AI-LHS at Lausanne University Hospital. This LHS closes both loops with the model continuously retraining, a real-time dashboard surfaces performance and outcome indicators, and human-in-the-loop validation refines predictions, producing measurable before-and-after improvements in sepsis detection and time-to-antibiotics across 97,559 patient stays26. Both exemplars are exceptions rather than the rule. Most deployed AI has neither loop, with no systematic capture of decisions and outcomes, no workflow adaptation, and no model retraining, such that predictive accuracy and operational utility often erodes after launch27,28. “AI-enabled LHS” is becoming a badge to claim, not a mechanism to build, as institutions often adopt the name while skipping the infrastructure that gives it meaning. Scaling the concept requires closing the same infrastructure gaps that stalled LHS for two decades and now shape AI deployment. Otherwise, AI-enabled LHS will inherit various failure modes of both predecessors, and potentially add new ones in a rapidly disruptive environment. The field needs what neither data nor AI alone can deliver: the learning utility. Extending Haendel et al.‘s data-as-utility framework, the learning utility is the full institutional stack that converts data and AI into continuous, bidirectional, measurable learning at the speed of care. The learning utility treats learning not as an innovation project but as permanent civic infrastructure, with the same institutional seriousness as electricity or clean water. AI provides the computational engine; the learning utility provides everything else including the data infrastructure that feeds it, the federated architecture that lets it learn across institutional boundaries, the research-operational integration that connects insights to action, and the governance layer that ensures it runs safely. With the utility in place, cycle-closure becomes the default, not the exception. Without it, every AI deployment is another isolated project and every LHS initiative another grant-bounded pilot. With it, every prediction captures its outcome, every outcome updates evidence, and every update propagates across the ecosystem. Using the grid analog where power distributes continuously, reliably, and fairly, rather than each consumer running their own generator. Without this infrastructure, AI’s failure modes compound at scale, including automation bias, substitution of clinician judgment, and oversight strain as generative systems produce content existing governance was never designed to evaluate. The path to a learning utility begins in academic medical centers (AMCs), which represent the natural proving ground, where research, clinical operations, and informatics coexist and where federal LHS training programs (AHRQ/PCORI K12, LHS E-StaR Centers) are anchoring LHS workforce development. But the principles and mechanisms must extend further including to community health systems, safety-net hospitals, and rural and low-resource settings, so that the same infrastructure gaps do not compound existing disparities29. If a learning utility works only at AMCs, it widens the access gap rather than closing it. “The learning utility” defines the learning utility concept; “Infrastructure: the learning utility stack” describes the three-layer infrastructure stack it requires; and “Toward a learning compact” proposes a learning compact and research agenda for realizing it. The learning utility The concept: learning-as-utility The power grid analogy maps onto the components of an AI-enabled learning system. Clinical data are the fuel; traditional statistical methods and now increasingly AI algorithms are the generation technology; learning is the electricity, and the output is what powers better healthcare. Generation technology has evolved dramatically, from coal-fired boilers to wind turbines to nuclear reactors, each advance making electricity production more efficient, cleaner, and more scalable. AI is the analogous advancement in knowledge generation: it can convert clinical data into learning at a speed and breadth that manual analysis, structured surveys, or traditional statistical methods cannot match. Data infrastructure and compute form Layer 1 (Foundation), the physical grid of wires, substations, and storage that move power from generators to consumers. Feedback loops, monitoring systems, and workforce are Layer 2 (Machinery), the operating layer, the day-to-day work that keeps power flowing reliably. Governance, trust, and alignment mechanisms are Layer 3 (Enabling Environment), the regulatory layer, the public rules that make the utility safe, fair, and accountable. Learning, in this mapping, is the utility service, delivered continuously, reliably, and fairly. The three pillars of a public utility, as Haendel et al. define them, extend to learning utility, with adjustments. Public good: learning is non-rivalrous, one health system’s learning benefits patients across other systems. Distributed stewardship: no single institution can close the full learning loop alone, the utility is inherently multi-level, operating both within institutions (across departments and units) and across them (via shared infrastructure, standards, and networks; see The ecosystem and the translator problem). And distributed stewardship of learning requires sharing not just data but also models, outcomes, and inferences, a broader coordination problem than data alone. Public oversight: because learning is a dynamic output rather than a static resource, public oversight requires that it be measured, which motivates the “GDP (Gross Domestic Product) of learning” (Why the utility must be measurable: the “GDP of learning”) and the learning compact (A learning compact). Together these make the system accountable for whether learning actually happens. The LHS describes what should happen. The learning utility describes what you must build to make it happen. One is a vision. The other is an infrastructure commitment. The LHS cycle, from data to knowledge to practice to outcomes, and back to data, is widely accepted. What the field has not built is the permanent infrastructure that makes the cycle actually run. Building that infrastructure means funding what health systems typically do not: not just the visible layer of data systems, compute, and AI tools, but also the feedback loops that capture outcomes, the monitoring systems that detect drift, the governance bodies that decide what is learned and shared, the workforce that translates between research and operations, and the trust infrastructure that makes all of it usable. “Infrastructure: the learning utility stack” details these layers. The claim here is simpler: the utility requires all of them together, not any one alone. With feedback loops, monitoring, and governance built in rather than funded per-project, learning becomes fast, granular, and continuous rather than slow, aggregate, and discontinuous. What the utility changes The LHS cycle is bidirectional, two halves of the same loop. Direction 1 (evidence → practice) applies accumulated knowledge to individual care through guidelines, risk scores, AI-based decision support, and alerts; the FDA AI-Enabled Medical Device List now counts more than 1,000 cleared devices supporting this flow30. Direction 2 (data → evidence) turns individual outcomes, decisions, and clinical experience into new knowledge through trials, model retraining, and quality improvement cycles. Here “data” is broadly to include both information already captured in the EHR and information made available through systematic capture in clinical practice, such as patient-reported outcomes, social determinants of health, decision rationale, longitudinal follow-up. That capture step is a design choice, not a given, and is itself part of what the learning utility must build (Layer 2: machinery, what makes it learn). The bidirectional framing itself is not new; it is the LHS cycle as the field has always understood it. What is new is recognizing that neither direction, as currently delivered, operates at utility speed. Both directions today run at project speed, not utility speed. Direction 1 is delivered project-by-project, each CDS tool, each risk-score rollout, each alert integration funded and managed as its own initiative. Direction 2 is delivered study-by-study, each learning cycle scoped to a grant or trial, starting from scratch and ending when funding does. The utility model is different: continuous, always-on, shared. Every clinical event contributes to an ever-growing evidence pool; every evidence update reaches the point of care without a grant, a trial, or a publication lag. Learning becomes faster because no one starts from scratch: the pool persists, parallel accumulation runs across institutions, and cycles do not pause between grants. But continuous, always-on, shared learning requires infrastructure, data capture, governance, funding, and workforce that does not expire at the end of a project. When either direction is deliberately built toward utility-grade infrastructure, progress becomes visible, though even the best current examples are closer to durable projects than to general-purpose infrastructure. For Direction 1 (evidence → practice), Kaiser Permanente’s Advance Alert Monitor (AAM) was refined through deployment: scores initially displayed every 6 hours on EHR dashboards were moved to hourly remote monitoring by a virtual nursing team that cascades validated alerts to rapid-response teams, institutionalizing 24/7 prediction-to-rescue across 21 hospitals that now handle roughly 16,000 alerts per year24,25,31. But AAM’s infrastructure is bespoke: the virtual nursing team, rescue workflows, and governance were purpose-built for this tool. For Direction 2, REMAP-CAP uses an adaptive platform design: randomization weights update monthly via a Bayesian inference model on accumulating outcomes, shifting allocation toward interventions that appear most favorable by stratum, 16,118 patients, 27,344 randomizations, and 240 active sites across 250+ institutions worldwide32. Adaptive platform trials are a growing trend (e.g., RECOVERY33, PRINCIPLE Trial Collaborative Group34, I-SPY235), reflecting an implicit recognition that evidence-generation benefits from persistent learning. These are the clearest current signals that each direction can move toward infrastructure. But utility-grade infrastructure, where deployment machinery serves new AI tools and adaptive platforms spawn the next trial at low marginal cost, remains unbuilt. Most Direction 1 (evidence → practice) deployments (the 1,000 + FDA-cleared AI tools) remain project-by-project; most Direction 2 (data → evidence) cycles remain study-by-study. When both directions are utility-grade, the mechanism yields benefits. The evidence pool accumulates continuously rather than being reset at the start of each grant cycle, so researchers generate evidence faster. Health systems draw on that shared evidence pool rather than rebuilding every tool from scratch, so operations run more efficiently. New evidence reaches the point of care without years of translation lag, so clinicians make better-informed decisions. Adjustments happen as evidence arrives rather than waiting for the next guideline, so patients receive more responsive care. Personalization emerges too, but as a property of the system, not a feature engineered into a model. Imagine a clinician asking, “What worked for patients like mine?” This is the Green Button concept, introduced more than a decade ago36,37. Without a learning utility, the pool of similar patients stays thin, decisions and outcomes go uncaptured, and the answer falls back on population averages. With a learning utility, every outcome from every similar patient feeds the pool, continuously. Over time, “patients like mine” becomes a larger, richer group, and the answer becomes more specific. The transition from concept to operating utility is now visible at Stanford University37, starting to scale the Green Button pathway into a continuously growing body of precision real-world evidence, surfacing it inside the EHR and ambient documentation workflows at the point of care38,39. becomes an addition to the evidence pool. The instance is one example, but the architectural lesson is general, when the pool, the workflow, and the quality assessment are all built into shared infrastructure, “patients like mine” stops being a research query and becomes a clinical service. Three design choices distinguish a learning utility from a clever query interface: every answer carries an evidence-quality assessment, so clinicians know when the pool is thick enough to act on; every clinician query, along with the decision made and the patient’s subsequent outcome, is captured and linked back so the loop closes; and every query triggers rapid-cycle, rigorously designed evidence-generation (e.g., causal inference and target trial emulation), turning accessible data into evidence rather than answers alone. This is not algorithmic magic; it is evidence density, accumulated, rigorous bidirectional learning across many similar cases. With learning as a utility, personalization is no longer a concierge feature available to a few; it becomes a civic property of the system. (Box 1) illustrates concretely with a GLP-1 receptor agonist example.) A primary care clinician asks: “For patients like mine (T2D, obesity, stage 3 CKD), should I start a GLP-1 receptor agonist, and if so, which one?“ Both directions of the LHS cycle today run at project speed, not utility speed. Direction 1 (evidence → practice) is more mature, but each deployment is funded and managed as its own project, not a shared infrastructure. Direction 2 (data → evidence) is where the loop most visibly fails to close. The left half of Direction 2 panel shows today’s study-by-study state. Green Button is a real step forward: it makes observational real-world data accessible at the point of care. But accessible data is not evidence. A Green Button-style query yields answers, but the clinician’s actual choice (which GLP-1, dose, or alternative) and the patient’s 6- and 12-month outcomes (HbA1c, weight, eGFR, MACE) are not explicitly captured or linked back to the query. Evidence requires linking the decision to the query and running rapid-cycle designed analyses with methodological safeguards (e.g., target trial emulation protocols with pre-specified outcomes and causal inference methods), so the answer improves the evidence pool with each cycle. Without that, the data grows but the learning doesn’t. The right half shows the utility-grade version: the prescribing decision and downstream outcomes (along with other patient context) are tagged and linked to the queries, rapid-cycle analyses run with rigorous study designs and methods built in (e.g., “mini-trials” via target trial emulation comparing GLP-1 options for this comorbidity profile), and the evidence pool grows with each cycle (847 → 848). The inset chart illustrates this dynamic: as evidence accumulates, “patients like mine” narrows (2000 → 47) while matching dimensions grow (5 → 200; e.g., kidney function, prior weight loss, GI tolerance, frailty, etc., beyond simple comorbidity diagnoses); thus, when both directions become utility-grade, the next clinician asking the same question for a similar patient receives real-world evidence grounded in 47 closely matched real-world chains rather than population averages over 2000 heterogeneous users. Personalization emerges as a civic property of the system, not an AI feature. Why the utility must be measurable: the “GDP of learning” Existing LHS frameworks, Friedman’s conceptual model of the LHS40; Foley & Vale’s five bodies of work for understanding, designing, and evaluating LHS2; Welch et al.‘s framework of LHS benefits3; and the Quintuple Aim, care experience, population health, cost, clinician wellbeing, and health equity41, collectively describe what a learning system should do, but share a gap: none provides a quantitative measure of whether a system is actually learning. Welch et al. (2024) explicitly call for future work to “develop a tool to evaluate an LHS’ impact.” The 17-year evidence-to-practice lag, the condition LHS was conceived to close, is cited everywhere as a symptom, but whether any LHS initiative has actually narrowed this gap remains unmeasured. What the field lacks is what economics had by 1937: a single, imperfect-but-useful number that tells you whether the system is working. The Haendel. et al. Science paper also recognizes this gap, “metrics that prioritize outcomes and real value, not just process adherence, institutional efficiency, or checkpoints, show better transformational performance“10. We call this missing measure the “GDP of learning.” Like GDP, it will not capture everything; it will miss dimensions of quality, equity, and sustainability, but it aligns diverse actors around a shared measure of progress. The distinction matters. Existing maturity models, for LHS (e.g., Lannon et al.‘s Network Maturity Grid42) and separately for AI readiness (e.g., HAIRA43), assess whether a health system can learn: inputs like data quality, technical readiness, organizational capacity. The GDP of learning measures whether it is learning: outputs like cycle speed, loop closure, evidence flow. Both input and output measures are needed; what is missing is the output metric. The methodological building blocks for an output measure already exist; what is missing is their assembly and validation. For example, net benefit and decision-curve analysis evaluate models in terms of the actions taken on them rather than discrimination alone, anchoring a “useful” axis to harms and benefits rather than to AUROC44,45. The FURM (Fair, Useful, Reliable Model) framework for AI operationalized this at health-system scale, treating impact as the joint product of model output, decision protocol, and the capacity of the workforce to act on the output46. A GDP of learning is, in this sense, a principled natural aggregation of these per-tool measures across the deployed portfolio, weighted by the populations actually reached and the frequency of the decisions involved: what fraction of our deployed AI tools deliver positive net benefit, in production, for the populations we actually serve; how is that fraction changing over time; and how much population-weighted benefit do they generate in aggregate? Without that aggregate, “AI-enabled LHS” remains a label; with it, the field acquires the output measure it has called for over two decades. Learning is also the common denominator across otherwise divergent stakeholders. Health-system operations, clinical teams, research, patients, and communities each value different things, efficiency and ROI, patient outcomes and reduced burden, evidence generation and scientific impact, trust and equitable access. All benefits when learning accelerates. External stakeholders, payers, regulators, vendors, professional societies, shape the environment the utility operates in, and the learning compact (A learning compact) proposes aligning their incentives. Without a measurable learning metric, each stakeholder optimizes for its own metric in isolation; with one, the GDP provides the common denominator. If the learning rate is zero, all stakeholders lose, regardless of what they individually value. Making learning measurable exposes the accountability gap that has stalled the LHS for 20 years. Making learning measurable, however, depends on infrastructure that most health systems have yet to develop. Without routine capture of decisions, actions, and outcomes, feedback loops remain incomplete. Without governance, learning efforts remain fragmented and misaligned. “Infrastructure: the learning utility stack” examines these infrastructure gaps directly. Infrastructure: the learning utility stack Drawing on frameworks related to LHS and AI (e.g., Friedman, Welch, CFIR 2.0, HAIRA, and others reviewed earlier), we identified more than a dozen infrastructure dimensions required for an AI-enabled learning utility. We grouped these dimensions by function into a three-layer stack (Fig. 1): Layer 1 (Foundation), what we build (data, compute, AI tools); Layer 2 (Machinery), what makes it learn (feedback loops, monitoring, maintenance); Layer 3 (Enabling Environment), what makes it sustainable (governance, trust, alignment). Health systems have invested heavily in Layer 1. The learning utility requires two additional layers they have mostly ignored.Fig. 1The learning utility stack.Three infrastructure layers: Foundation (Layer 1: data, compute, AI tools), Machinery (Layer 2: feedback loops, monitoring, maintenance), and Enabling Environment (Layer 3: governance, trust, alignment), each spanning four cross-cutting dimensions (Data, Technology, Workforce, Governance). Cell shading and status badges mark investment intensity: Layer 1 is built but project-by-project; Layer 2 is a documented gap; Layer 3 is largely absent. The translator band on the right highlights the cross-layer workforce the field has not yet professionalized. The matrix operationalizes the wrong-investment thesis: the field built Layer 1 and must now build Layers 2–3 and the translators who connect them. The learning utility stack. Three infrastructure layers: Foundation (Layer 1: data, compute, AI tools), Machinery (Layer 2: feedback loops, monitoring, maintenance), and Enabling Environment (Layer 3: governance, trust, alignment), each spanning four cross-cutting dimensions (Data, Technology, Workforce, Governance). Cell shading and status badges mark investment intensity: Layer 1 is built but project-by-project; Layer 2 is a documented gap; Layer 3 is largely absent. The translator band on the right highlights the cross-layer workforce the field has not yet professionalized. The matrix operationalizes the wrong-investment thesis: the field built Layer 1 and must now build Layers 2–3 and the translators who connect them. The wrong-investment thesis Critics and health-system leaders often talk past each other about “infrastructure.” Health systems have invested substantially in Layer 1, the physical grid: EHRs, data warehouses, compute, and analytics at considerable scale. They have barely begun building Layer 2, the operations that make the grid generate and deliver power. Layer 3, the governance that makes the grid trustworthy and sustainable, remains almost entirely absent. We cannot have a utility with only the physical plant; we need all three layers. Health systems are purchasing AI tools, investing in Layer 1, and calling it “learning.” But technology adoption is not learning. Without the feedback loops to capture outcomes (Layer 2) and the governance to align incentives (Layer 3), Layer 1 spending produces more AI deployed but no more learning achieved. Digital maturity scores have been shown to offer false comfort about AI readiness47; the same risk applies to the “LHS” label, and to maturity models that primarily assess Layer 1 inputs while Layers 2–3 go unmeasured. The pattern has a structural explanation: Layers 2 and 3 are infrastructure whose returns accrue diffusely and over time, making them harder to measure, attribute, and fund than visible Layer 1 investments, especially when health systems are under short-term financial pressure, even though, once built, Layers 2–3 would compound those same operational gains. Another pattern is that health systems invest in AI projects individually, with each project developing its own data pipeline and governance structure. A collection of projects is not a utility, just as a collection of generators is not a power grid. Hospital command centers illustrate the gap between successful AI projects and learning utility (Box 2). The point is not that Layer 1 infrastructure is unimportant, but that it is insufficient. Once a health system has basic data capacity, the limiting factor often shifts to whether it can govern, implement, monitor, and adapt AI-enabled learning. Therefore, the marginal return on additional Layer 1 investment may be lower than the return on initial investments in Layer 2 or Layer 3 infrastructure. The recommendation is to invest differently, not necessarily more. Panel 1 shows Michigan Medicine’s M2C274 command center as a successful AI project: Layer 1 done well within the project’s own scope, with strong data infrastructure (Epic integration, in-house AI bed-placement tool), real ROI ($19.5 M net over two years; −33% adult bed wait time; −37% ED wait for beds; −8% length of stay; 63 virtual beds), and local operational learning (Isolated Success). Panel 2 reframes M2C2 as one of many projects all served by a shared three-layer learning utility: Layer 1 (shared data, compute, AI tooling, common standards), Layer 2 (feedback loops, monitoring, shared data and model lifecycle), and Layer 3 (governance, trust, alignment, learning compact) together form the utility backbone; M2C2 sits alongside peer projects (sepsis early warning, imaging triage AI, discharge planning) all connected upward into the stack; translators bridge research and operations; and an inter-institutional compact connects the institution to the broader field (Connected Learning). The infrastructure, all three layers together, serving many projects, is the utility, not any one project. The bottleneck is the system, not the algorithm. Health systems invested in data and AI (Layer 1). The learning utility requires two additional layers they have mostly ignored: feedback loops and monitoring (Layer 2), and governance, trust, and alignment (Layer 3).The learning utility is civic infrastructure, not another framework. The full institutional stack that converts data and AI into continuous, bidirectional, measurable learning at the speed of care, treated with the same seriousness as electricity or clean water.Both directions of learning exist, but neither runs at utility speed. Direction 1 (system-to-individual) is widely deployed ( > 1000 FDA-cleared AI tools) but rolled out project-by-project. Direction 2 (individual-to-system) is slower still, cycles gated by grants and trials, and includes capturing decisions, outcomes, and clinical experience as data, not just analyzing already-stored data. The learning utility turns both into continuous, shared infrastructure.The GDP of learning: a measurable alignment mechanism. Without a shared metric, each stakeholder optimizes in isolation. Learning is the common denominator that aligns research, operations, clinicians, patients, and payers.A collection of AI projects is not a utility. The utility requires an ecosystem, differentiated institutional roles plus shared, connected infrastructure across all three layers (held by a neutral convener where one exists, federated where none does), operating at both intra- and inter-institutional levels.Translators have the highest multiplier effect. Cross-layer bridge roles connect research to operations to governance. One translator creates value across the full stack; one more AI tool without translation creates another silo.The speed of learning is the speed of trust. Barriers labeled as regulatory, legal, or organizational are frequently trust problems in disguise. Trust infrastructure is the rate-limiting factor for everything else.Generalizability must be designed in, not assumed. Learning itself does not make knowledge transferable. The utility must capture not just “what worked” but “for whom, in what context, under what conditions.”The learning compact organizes obligations and supports the ecosystem. It requires learning-loop closure (not just data access) and distributes obligations across five ecosystem roles (generators, transmission, distribution, regulators, prosumers). Enforcement is non-legal, making the cost of not learning visible through published learning metrics.Learning rigor must be built into continuous learning. Faster learning without rigorous methods, confounding control, outcome validation, bias detection, is faster noise. The learning utility must ensure speed does not sacrifice quality. The bottleneck is the system, not the algorithm. Health systems invested in data and AI (Layer 1). The learning utility requires two additional layers they have mostly ignored: feedback loops and monitoring (Layer 2), and governance, trust, and alignment (Layer 3). The bottleneck is the system, not the algorithm. Health systems invested in data and AI (Layer 1). The learning utility requires two additional layers they have mostly ignored: feedback loops and monitoring (Layer 2), and governance, trust, and alignment (Layer 3). The learning utility is civic infrastructure, not another framework. The full institutional stack that converts data and AI into continuous, bidirectional, measurable learning at the speed of care, treated with the same seriousness as electricity or clean water. The learning utility is civic infrastructure, not another framework. The full institutional stack that converts data and AI into continuous, bidirectional, measurable learning at the speed of care, treated with the same seriousness as electricity or clean water. Both directions of learning exist, but neither runs at utility speed. Direction 1 (system-to-individual) is widely deployed ( > 1000 FDA-cleared AI tools) but rolled out project-by-project. Direction 2 (individual-to-system) is slower still, cycles gated by grants and trials, and includes capturing decisions, outcomes, and clinical experience as data, not just analyzing already-stored data. The learning utility turns both into continuous, shared infrastructure. Both directions of learning exist, but neither runs at utility speed. Direction 1 (system-to-individual) is widely deployed ( > 1000 FDA-cleared AI tools) but rolled out project-by-project. Direction 2 (individual-to-system) is slower still, cycles gated by grants and trials, and includes capturing decisions, outcomes, and clinical experience as data, not just analyzing already-stored data. The learning utility turns both into continuous, shared infrastructure. The GDP of learning: a measurable alignment mechanism. Without a shared metric, each stakeholder optimizes in isolation. Learning is the common denominator that aligns research, operations, clinicians, patients, and payers. The GDP of learning: a measurable alignment mechanism. Without a shared metric, each stakeholder optimizes in isolation. Learning is the common denominator that aligns research, operations, clinicians, patients, and payers. A collection of AI projects is not a utility. The utility requires an ecosystem, differentiated institutional roles plus shared, connected infrastructure across all three layers (held by a neutral convener where one exists, federated where none does), operating at both intra- and inter-institutional levels. A collection of AI projects is not a utility. The utility requires an ecosystem, differentiated institutional roles plus shared, connected infrastructure across all three layers (held by a neutral convener where one exists, federated where none does), operating at both intra- and inter-institutional levels. Translators have the highest multiplier effect. Cross-layer bridge roles connect research to operations to governance. One translator creates value across the full stack; one more AI tool without translation creates another silo. Translators have the highest multiplier effect. Cross-layer bridge roles connect research to operations to governance. One translator creates value across the full stack; one more AI tool without translation creates another silo. The speed of learning is the speed of trust. Barriers labeled as regulatory, legal, or organizational are frequently trust problems in disguise. Trust infrastructure is the rate-limiting factor for everything else. The speed of learning is the speed of trust. Barriers labeled as regulatory, legal, or organizational are frequently trust problems in disguise. Trust infrastructure is the rate-limiting factor for everything else. Generalizability must be designed in, not assumed. Learning itself does not make knowledge transferable. The utility must capture not just “what worked” but “for whom, in what context, under what conditions.” Generalizability must be designed in, not assumed. Learning itself does not make knowledge transferable. The utility must capture not just “what worked” but “for whom, in what context, under what conditions.” The learning compact organizes obligations and supports the ecosystem. It requires learning-loop closure (not just data access) and distributes obligations across five ecosystem roles (generators, transmission, distribution, regulators, prosumers). Enforcement is non-legal, making the cost of not learning visible through published learning metrics. The learning compact organizes obligations and supports the ecosystem. It requires learning-loop closure (not just data access) and distributes obligations across five ecosystem roles (generators, transmission, distribution, regulators, prosumers). Enforcement is non-legal, making the cost of not learning visible through published learning metrics. Learning rigor must be built into continuous learning. Faster learning without rigorous methods, confounding control, outcome validation, bias detection, is faster noise. The learning utility must ensure speed does not sacrifice quality. Learning rigor must be built into continuous learning. Faster learning without rigorous methods, confounding control, outcome validation, bias detection, is faster noise. The learning utility must ensure speed does not sacrifice quality. Layer 1: foundation, what you build Layer 1 is the data and technical foundation, built up at progressively larger scales: EHRs and clinical data warehouses within individual health systems (e.g., the Mayo Clinic Platform); common data models (OMOP) and federated research networks at national scale (PCORnet, NIH’s All of Us Research Program, the NIH Health Care Systems Research Collaboratory); and emerging international infrastructure such as the European Health Data Space. Federated learning has matured as a parallel technical advance, training models across institutions without centralizing patient data48,49. Interoperability standards such as FHIR, cloud compute, security, and the AI tools and CDS that sit on top of this foundation complete the layer50. This is where investment has concentrated, and it matters. While it is necessary, it is not sufficient. Layer 2: machinery, what makes it learn Layer 2 is the operational infrastructure that converts data and AI into actual learning, and it might be the largest gap in the field51–55. Feedback-loop infrastructure captures decisions and outcomes and links them back to models, and triggers consequent action: model retraining, workflow adjustment, or de-implementation when warranted. It is the infrastructure that makes the return path (What the utility changes) systematic rather than ad hoc. Most sites deploy AI without closing this loop. This is where the Green Button pathway36,37,56 may break: the decision and outcome are not captured. Feedback-loop infrastructure also requires its own workforce, data engineers to build the capture systems, monitoring specialists to maintain them, and governance leaders to define what must be tracked. Monitoring and maintenance infrastructure detects model drift, surveils performance, and raises alerts. Vanderbilt’s VAMOS is an institutional exemplar, purpose-built monitoring that tracks performance, process, outcomes, and fairness57. Davis’s 360-degree maintenance model, developed for deployed clinical prediction tools, describes four modes, preventive, preemptive, responsive, and reactive58. The learning utility must support all four, and apply them across the full stack, not only to models. If AI is infrastructure, monitoring is maintenance. Utilities have maintenance budgets, maintenance teams, and accountability structures; projects do not. Monitoring without intervention is observation, not learning. Detected drift, performance decline, or fairness shifts must trigger action: retrain, retire, recalibrate, escalate to human review, or adjust workflow. Without such infrastructure, deployed AI can degrade undetected, as the Epic sepsis case (Introduction: the promise, the progress, and the gap) illustrates. Research-operational integration is the connective tissue between evidence generation and clinical action. Research IT is optimized for batch processing, long-horizon studies, and IRB-governed data access; operational IT is optimized for real-time patient care, regulatory compliance, and clinical workflow integration. The LHS cycle requires both simultaneously, in what Coiera called the “last mile” problem: technically high-performing systems that fail at the point of integration into clinical workflows, not because algorithms are wrong but because the institutional infrastructure connecting algorithmic output to clinical action and workflow is inadequate59. Most health-system architectures have not assigned any organizational unit to this connective role60. It demands interdisciplinary translators who are fluent across clinical, technical, research, and operational domains, who match solutions to problems rather than speak only one discipline (The ecosystem and the translator problem). Pragmatic-trial infrastructure embedded in the EHR (e.g., running AI-tool deployments as continuous comparative-effectiveness experiments rather than one-time validations) is a Layer 2 capability the utility could support natively29. These capabilities are demonstrated with the University of Minnesota’s “RapidEval” pragmatic trial unit enabled through a range of IT capabilities including adaptive evaluation design. Similarly, The University of Wisconsin’s Pragmatic Trial Operations (PTOps) playbook for monitoring and evaluating ambient AI in clinical practice is an institutional exemplar of this capability61. Layer 3: enabling environment, what makes it sustainable Layer 3 comprises the governance, trust, alignment, and sustainability infrastructure that makes learning durable. Governance and accountability determine who decides and who is responsible. Yet governance priorities today are often misaligned with learning goals: only 17% of health systems “always measured” health equity as an AI evaluation metric, while 74% always measured basic uptake62. If AI-LHS is a utility, equity is an infrastructure-access problem, not just an algorithmic-bias problem. Trust infrastructure operates at multiple levels, clinician (whether to accept AI-assisted recommendations and delegate tasks), patient (whether to consent to AI-augmented care and data use), and inter-organizational (whether to share data, models, and outcomes across institutional boundaries), and across institutional functions such as legal, compliance, IT, and payer relations, each of which can accelerate or block learning on the basis of trust. Empirically it is central: in one institutional study, trust accounted for 56% of variance in clinician AI acceptance63; across 36 academic medical centers, lack of clinician trust was cited by 50% of institutional leaders as a top concern in generative AI implementation64. Health systems move at the speed of trust. Without trust, nothing gets adopted regardless of performance. Generative AI sharpens this challenge: its novel failure modes, hallucinations, non-determinism, emergent behavior, are often invoked, alongside anecdotes and a handful of high-profile failures, to argue that it is categorically different and therefore untrustworthy. The response cannot be indefinite caution; it must be more learning-utility infrastructure, continuous monitoring, bounded use cases, and feedback loops that surface failures early rather than waiting for headlines. This also requires new evaluation paradigms: predictive-performance metrics are insufficient for systems that synthesize, reason, and act, and tiered, risk-based frameworks for generative and agentic AI (that also bridge benchmark performance to deployed net benefit) remain an active research need. Early-stage infrastructure for this is emerging. MedHELM offers a clinician-validated taxonomy of 121 medical tasks across five categories, plus a benchmark suite for apples-to-apples comparison of frontier models on real-world clinical work rather than licensing-exam questions65. MedAgentBench extends this toward agentic AI, benchmarking multi-step clinical reasoning and action in a virtual EHR environment20. Benchmarks of this kind are necessary but not sufficient: they evaluate model capability in isolation, whereas the learning utility ultimately needs evaluation of model-plus-workflow-plus-action, i.e., model output, decision protocol, and workforce capacity to act, the same triad FURM established for predictive AI46. Crucially, many barriers to learning velocity are commonly labeled as regulatory, legal, organizational, or technical, but are trust problems in disguise. Legal, IT, operations, and research each invoke legitimate constraints, HIPAA, security, bandwidth, data quality, but the discretion in how those constraints are practiced is where trust operates. When trust is high, the same policy is interpreted permissively; when low, it becomes a wall. Trust infrastructure is therefore not optional; it is the rate-limiting factor for everything in Layers 1 and 2. Alignment and incentives establish shared metrics and lifecycle funding rather than one-time acquisition. The alignment gap is structural: research optimizes for publications, operations for efficiency, IT for system stability. When measured against learning output, the divergence resolves, which is why the GDP of learning (Why the utility must be measurable: the “GDP of learning”) is not merely a measurement aspiration but the mechanism that resolves the alignment failure. Sustainability also requires that acquisition and maintenance be funded; capital-equipment budgeting is incompatible with subscription-based AI. The ecosystem and the translator problem The learning utility is operated by an ecosystem of differentiated institutional roles (Fig. 2). In power-grid terms: generators (any clinical encounter, any patient experience, any operational event, learning generation is inherently distributed across all actors and settings); transmission (informatics institutes, health information exchanges (HIEs), research networks, model registries, professional societies, journals, and standards bodies, the channels that move data, models, and validated evidence within and between institutions); distribution (the settings and workflows where learning reaches end users, from clinics to home health and pharmacies, and the workflows that act on evidence at each site); regulators (the Food and Drug Administration [FDA], the Office of the National Coordinator for Health Information Technology [ONC], the Centers for Medicare & Medicaid Services [CMS], and other federal/state regulators and accreditation bodies); and prosumers (producer-consumers, patients, clinicians, caregivers, and community health workers who both contribute observations and receive learning). AMC’s outsized current role in the utility stems from transmission and capture capacity, not generation: learning happens wherever care happens, but converting observations into shared evidence requires infrastructure and workforce that AMCs have built over time.Fig. 2The learning ecosystem.Five differentiated roles in a power-grid analogy: Generators produce raw learning material (clinical encounters, patient experiences, operational events, outcomes) across all care settings; Transmission moves data, models, and validated evidence (HIEs, research networks, registries, societies, standards bodies); Distribution flows validated evidence into action (guidelines, decision support, EHR alerts, care pathways); Regulators set constraints (FDA, ONC, CMS, public-health and accreditation bodies); Prosumers both contribute and receive (patients, clinicians, caregivers, community health workers). The shared backbone is shared infrastructure spanning all three stack layers (Layer 1 standards/registries/APIs; Layer 2 monitoring/feedback/outcome capture; Layer 3 governance/trust/compact); a multi-party connecting body (e.g., regional HIE, coordinating center of a research network, independent research institution) can hold it where one exists, otherwise it must be built federated, multi-party, governed jointly. Four gold dashed arcs represent translator bridges (clinical/technical, research/operations, governance/technical, care/patient). The green return arrow makes bidirectional learning visible on an ecosystem scale. The bottom inset shows the same framework at intra- and inter-institutional scales. The learning ecosystem. Five differentiated roles in a power-grid analogy: Generators produce raw learning material (clinical encounters, patient experiences, operational events, outcomes) across all care settings; Transmission moves data, models, and validated evidence (HIEs, research networks, registries, societies, standards bodies); Distribution flows validated evidence into action (guidelines, decision support, EHR alerts, care pathways); Regulators set constraints (FDA, ONC, CMS, public-health and accreditation bodies); Prosumers both contribute and receive (patients, clinicians, caregivers, community health workers). The shared backbone is shared infrastructure spanning all three stack layers (Layer 1 standards/registries/APIs; Layer 2 monitoring/feedback/outcome capture; Layer 3 governance/trust/compact); a multi-party connecting body (e.g., regional HIE, coordinating center of a research network, independent research institution) can hold it where one exists, otherwise it must be built federated, multi-party, governed jointly. Four gold dashed arcs represent translator bridges (clinical/technical, research/operations, governance/technical, care/patient). The green return arrow makes bidirectional learning visible on an ecosystem scale. The bottom inset shows the same framework at intra- and inter-institutional scales. Shared, connected infrastructure is the underappreciated structural requirement. Each entity invests in its own infrastructure tuned to its needs, but also must invest in the connective layer that lets the ecosystem learn — and the connective layer spans all three layers of the stack:Layer 1 connective: common data standards, shared registries, interoperable APIs, model registries.Layer 2 connective: shared monitoring infrastructure, joint feedback loops, common evaluation pipelines, cross-institution outcome capture.Layer 3 connective: shared governance frameworks, aligned incentives, common trust agreements, the learning compact (A learning compact). Layer 1 connective: common data standards, shared registries, interoperable APIs, model registries. Layer 1 connective: common data standards, shared registries, interoperable APIs, model registries. Layer 2 connective: shared monitoring infrastructure, joint feedback loops, common evaluation pipelines, cross-institution outcome capture. Layer 2 connective: shared monitoring infrastructure, joint feedback loops, common evaluation pipelines, cross-institution outcome capture. Layer 3 connective: shared governance frameworks, aligned incentives, common trust agreements, the learning compact (A learning compact). Layer 3 connective: shared governance frameworks, aligned incentives, common trust agreements, the learning compact (A learning compact). In some settings, a multi-party connecting body (e.g., regional HIEs, coordinating centers of research networks, independent research institutions connected across parties) can hold the connective layer on behalf of the ecosystem. Such a body does not need to be incentive-free (e.g., network coordinating centers publish, HIEs collect fees, and research institutes seek grants). What makes it work is multi-party governance that creates incentives aligning each stakeholder’s benefit with the ecosystem mission. Where no such body exists, the connective layer must still be built: federated, multi-party, governed jointly. The utility is nested: the same roles and learning can apply within institutions Every unit, clinical, research, and operational, is a generator. Research rigor, methods, inference, bias control, must be embedded in clinical settings; the long protocols and grant cycles of traditional research cannot keep up with the speed of care. Internal informatics, research IT, and data warehouses serve as transmission; service lines and point-of-care workflows are the distribution layer; AI governance committees, institutional review boards (IRBs), and quality/safety bodies act as regulators. This scaling is enforced by leadership accountability within institutions and by professional standards, funding conditions, and published metrics across them. Translators are the missing workforce Every layer has its own expertise, data engineers (Layer 1), monitoring specialists (Layer 2), governance experts (Layer 3). The layers do not connect because no one’s job is to be the connection. The learning utility requires translator roles that bridge across layers and domains: clinical–technical translators (AI capabilities ↔ clinical workflow), research-operations translators (evidence generation ↔ practice change), governance–technical translators (policy requirements ↔ system design), and care–patient translators (care delivery ↔ patient activation and lived experience). Power grids do not run on enthusiasm and training modules. They have linemen, dispatchers, engineers, and inspectors, permanent, professionalized roles. The learning utility similarly requires not just experts within each layer, but translators who connect them. Currently, nobody’s job title is “the person who makes Layer 1, Layer 2, and Layer 3 talk to each other.” It should be. Clinical informatics board certification, which now requires Accreditation Council for Graduate Medical Education (ACGME)-accredited fellowship training, is one step; Mahajan & Bates66 argue for a dedicated physician role in clinical AI. Nevertheless, physicians with informatics training alone might not be enough. Effective translators are multilingual, fluent in research methodology, clinical workflow, AI/data, and institutional operations, able to move between physician, researcher, informatician, and administrator without losing meaning. Clinical informatics as a field sits at this intersection, and it does not depend on degree (MD, PhD, or MD/PhD); what matters is demonstrated cross-domain competency, not the title. For that matter, the translator may not even need to be human. LLM-based systems can already absorb routine translation tasks: drafting clinical documentation67, translating clinical content for patient consumption68, and turning real-world evidence into point-of-care answers56. What they cannot replace is the institutional work of translation, governance accountability, organizational diplomacy, contextual judgment about what to build and what to retire, and the trust relationships that enable cross-layer coordination. Translators are LLM-augmented operators of the learning utility, not a workforce that LLMs displace. In practice this may take a team with complementary backgrounds, some rooted in clinical practice, others in research methods, others in operations, but all multilingual, each fluent enough in the others’ languages to translate across them, supported by governance, legal, and IT expertise. The broader translator workforce remains undefined and unfunded. Direct evidence for the translator workforce is fragmentary precisely because the role is informal and untracked. The best available signal comes from Li et al.‘s meta-analysis of 105 AI deployment studies: clinician-led deployments were far more likely to succeed (OR 7.79, p = 0.039), ~70% of successful deployments came from mixed clinician-technologist teams, and technologist-only teams fared worst (60% vs. 94% impact, p < 0.001); the authors frame this in Team Science terms and describe study teams as “translating AI into reality”69. Senior authorship is only a proxy for leadership; the data cannot distinguish whether clinician leaders personally bridged domains or convened teams that did, but either reading points to cross-domain bridging as the predictor of success. Beyond deployment mechanics, translators do work no single-layer expert can: brokering trust across clinicians, patients, regulators, and administrators, and aligning incentives that otherwise diverge between research (rigor, novelty) and operations (throughput, cost). Absent a formally defined and funded workforce, this bridging happens by luck, concentrated wherever a gifted individual or an unusually mixed team exists. Two faces of the same question: who gets to participate in the learning utility, and who maintains it. Toward a learning compact Redirect investment The first recommendation is to spend differently, not necessarily more. All layers need investment, including Layer 1 (Foundation), data infrastructure, compute, and AI tools. But if budgets are constrained, Layer 1 already dominates health-system IT budgets; the marginal return on the next Layer 1 dollar is diminishing while Layer 2 (Machinery: feedback-loop infrastructure for decision capture and outcome tracking, monitoring, and maintenance) and Layer 3 (Enabling Environment: governance, trust-building, alignment, and lifecycle funding) remain largely unbuilt. Redirecting, or at minimum rebalancing, dollars toward these layers is where additional investment will do the most work. Every AI deployment should budget for the full stack, not just the model. If you cannot fund the feedback loop, the monitoring, and the governance, you cannot fund the deployment. Use the GDP of learning (Why the utility must be measurable: the “GDP of learning”) as the measure of whether you are spending right. Among all investments, the translator workforce, cross-layer bridges introduced in “The ecosystem and the translator problem”, has the highest multiplier effect. A single translator strengthens all three layers simultaneously: a research-operations translator improves feedback loops, builds inter-organizational trust, and makes data infrastructure more fit-for-purpose. Without translators, layers remain siloed: data infrastructure exists but does not connect to clinical workflows; governance exists but does not reflect technical realities; monitoring exists, but no one acts on its findings. One translator role creates value across the full stack; one more AI tool without translation creates another silo. If a health system could do only one thing, it should be this. A learning compact Haendel et al. proposed a “regulatory compact” for data: pair obligations (data sharing, interoperability, nondiscrimination) with supports (infrastructure investment, economic sustainability, regulatory modernization beyond HIPAA), grounded in distributed stewardship and public oversight and governed through federated, standards-based, community-driven boards10. We extend this in two ways. First, the learning compact mandates more than data access: it requires learning-loop closure, capture of decisions and outcomes, systematic feedback to models, and measurable improvement over time. Second, the compact distributes obligations and supports the five ecosystem roles from “The ecosystem and the translator problem” (generators, transmission, distribution, regulators, and prosumers) rather than treating data holders as a single class. Table 1 details how this plays out at the stakeholder level, the values each actor brings, how learning serves their interests, what they contribute, and what they receive in return.Table 1Stakeholder alignment and the learning compactStakeholderWhat they valueHow learning serves themContributesReceivesINTERNALHealth system operations (incl. IT)Efficiency, ROI, stability, regulatory complianceFaster identification of what works; evidence-based resource allocationClose learning loops; fund Layers 2–3Shared infrastructure; liability clarityClinicalBetter patient outcomes, reduced burden, autonomyBetter decision support; less alert fatigue; evidence at the point of careCapture decisions and outcomes at point of careEvidence-based tools tuned to their populationResearchPublications, grants, scientific impactFaster evidence from real-world data; Direction 2 (evidence to practice) infrastructureGenerate and validate evidence; train translatorsData access; implementation partnershipsPatients & communitiesBetter outcomes, trust, transparency, equitable accessCare informed by latest evidence; personalization as emergent propertyPatient-reported outcomes (PROs); co-design participationTransparency; right to benefit from learningEXTERNALPayersCost-effectiveness; quality metricsEvidence for what worksPayment incentives tied to learning metricsBetter cost-effectiveness data; reduced wasteRegulators & policySafety, compliance, accountabilityBetter safety dataMandate monitoring; create safe harborsStandardized reporting; reduced burdenVendors & industryMarket share; adoptionPlatform stickinessEnable interoperability; not captureMarket access; integration standardsProfessional societiesStandards, qualityField-wide improvementAccreditation; quality registriesInfluence over utility governanceEight stakeholders, four internal (operating the learning utility: health system operations, clinical, research, patients and communities) and four external (shaping its environment: payers, regulators, vendors, professional societies), with what each values, how continuous learning serves those values, what the compact obliges them to contribute, and what they receive in return. Learning is the common denominator across otherwise divergent stakeholder interests; the compact organizes their obligations and benefits into a reciprocal deal. Table 1Stakeholder alignment and the learning compactStakeholderWhat they valueHow learning serves themContributesReceivesINTERNALHealth system operations (incl. IT)Efficiency, ROI, stability, regulatory complianceFaster identification of what works; evidence-based resource allocationClose learning loops; fund Layers 2–3Shared infrastructure; liability clarityClinicalBetter patient outcomes, reduced burden, autonomyBetter decision support; less alert fatigue; evidence at the point of careCapture decisions and outcomes at point of careEvidence-based tools tuned to their populationResearchPublications, grants, scientific impactFaster evidence from real-world data; Direction 2 (evidence to practice) infrastructureGenerate and validate evidence; train translatorsData access; implementation partnershipsPatients & communitiesBetter outcomes, trust, transparency, equitable accessCare informed by latest evidence; personalization as emergent propertyPatient-reported outcomes (PROs); co-design participationTransparency; right to benefit from learningEXTERNALPayersCost-effectiveness; quality metricsEvidence for what worksPayment incentives tied to learning metricsBetter cost-effectiveness data; reduced wasteRegulators & policySafety, compliance, accountabilityBetter safety dataMandate monitoring; create safe harborsStandardized reporting; reduced burdenVendors & industryMarket share; adoptionPlatform stickinessEnable interoperability; not captureMarket access; integration standardsProfessional societiesStandards, qualityField-wide improvementAccreditation; quality registriesInfluence over utility governanceEight stakeholders, four internal (operating the learning utility: health system operations, clinical, research, patients and communities) and four external (shaping its environment: payers, regulators, vendors, professional societies), with what each values, how continuous learning serves those values, what the compact obliges them to contribute, and what they receive in return. Learning is the common denominator across otherwise divergent stakeholder interests; the compact organizes their obligations and benefits into a reciprocal deal. Stakeholder alignment and the learning compact Eight stakeholders, four internal (operating the learning utility: health system operations, clinical, research, patients and communities) and four external (shaping its environment: payers, regulators, vendors, professional societies), with what each values, how continuous learning serves those values, what the compact obliges them to contribute, and what they receive in return. Learning is the common denominator across otherwise divergent stakeholder interests; the compact organizes their obligations and benefits into a reciprocal deal. Enforcement is not by mandate but by making the cost of not learning visible. The GDP of learning (Why the utility must be measurable: the “GDP of learning”) is the enforcement mechanism: publishing learning metrics exposes non-learners. Non-legal levers include payer incentives (CMS quality metrics), accreditation (Joint Commission), professional standards, and reputational pressure from published metrics. Professional societies are already moving from guidelines to operational infrastructure, radiology provides the clearest operational template: multi-society statements now cover the full AI lifecycle from development to monitoring70, and the ACR launched the ARCH-AI program in June 202471, a precursor to formal practice accreditation anticipated in spring 2027, with attestation requirements across governance, model selection, acceptance testing, and monitoring, the last requiring participation in the Assess-AI registry for benchmarking local performance71. Analogous operational infrastructure remains to be built in other specialties. The learning compact is, crucially, not legislation but an analytical framework that does two things: it aligns existing and proposed incentive mechanisms into a reciprocal deal that makes collective action rational, and it extends the set of obligations from data access to learning-loop closure. This is what utility compacts did for electricity in the 1930s, and what pieces of a data compact are beginning to do now: ONC information-blocking rules, CMS Promoting Interoperability, HIPAA Right of Access, and TEFCA, though these remain fragmented, incomplete, and oriented toward data access rather than learning. Research agenda Three priorities follow from the framework. Priority 1: Validate a composite learning metric, the “GDP of learning.” Five candidate dimensions offer a starting point, not an exhaustive list. Learning velocity measures how fast evidence reaches practice; its limiting factor is the speed of trust. Learning completeness measures what fraction of learning cycles actually close (AI recommendation → decision → outcome → fed back), the signature metric distinguishing a learning utility from a decision-support system. Learning rigor measures the quality of the evidence learned, guarding against “fast but wrong” learning. Learning equity asks whether the utility generates evidence for all populations, not just about them, extending the Quintuple Aim’s health-equity pillar41 from care to learning. Learning sustainability asks whether learning persists after grants end, extending LHS concerns about project-dependent initiatives (pilotitis, Introduction: the promise, the progress, and the gap) to utility-scale persistence. Velocity, completeness, and rigor lack direct analogs in existing frameworks. Additional candidates include bidirectionality (whether both Direction 1 evidence-to-practice and Direction 2 data-to-evidence flows are running, see What the utility changes), adaptiveness (the ability to respond to new conditions such as pandemics), and generalizability (whether learning transfers beyond the local context). Existing maturity models across LHS (e.g., Lannon9), AI (e.g., HAIRA43), and real-world data domains serve a different and complementary role, assessing inputs (can you learn?) rather than outputs (are you learning?); synthesizing them into a learning-utility maturity model is a parallel research need. Operational candidates for several of these dimensions already exist (Why the utility must be measurable: the “GDP of learning”, e.g., net benefit44,45, and FURM-style assessment46). The engineering work is to assemble and validate them into a portfolio-level metric across institutions and federated networks: tractable, but it has not yet been done at utility scale. The core call of this paper is to build the GDP of learning: the output measure the field has advocated around for two decades but never had. Without it, accountability for learning remains aspiration; with it, learning becomes something the field can track, compare, and improve. Priority 2: Design and test closed-loop feedback infrastructure The concrete engineering challenge: how do you link an AI recommendation → the clinician’s decision → the patient’s outcome → back into the evidence pool, at scale, within existing EHR workflows? Early efforts, including the Green Button pathway (Layer 2: machinery, what makes it learn), attempted pieces of this. Isolated successes like Kaiser AAM, REMAP-CAP, and ImproveCareNow72 demonstrate that portions of the cycle can be built as infrastructure; none has been extended to a general, utility-grade feedback system across institutions. EHR-embedded pragmatic trials of AI-software interventions are one concrete form this infrastructure can take, where the loop and the experiment are the same system; the University of Wisconsin’s PTOps playbook for ambient AI is an early operational instance61. The utility should reduce the marginal cost of running such trials to where pragmatic AI-tool RCTs become routine evidence-generation, not exceptional events. Closing the loop is not enough on its own: generalizability should be a design feature of the learning utility, not an assumed by-product. Learning itself does not make knowledge generalizable. A learning utility must capture not just “what worked” but “for whom, in what context, under what conditions.” Without that design choice, even excellent local learning stays local. Priority 3: Professionalize the translator workforce The translator workforce is the single highest-leverage workforce investment in the learning utility (Sections 3e, 4a), yet remains undefined and underfunded. Questions: What organizational structures enable cross-layer bridges, joint appointments, embedded roles, or standalone translator units? What training pathways (beyond existing clinical informatics certification) are needed for research-operations and governance–technical translators? What career trajectories, promotion ladders, salary bands, tenure-equivalent recognition, would make translator roles institutionally durable? What is the return on investment of translator roles compared to additional Layer 1 spending? A parallel research direction is whether LLM-based agents can offload routine translation tasks (e.g., evidence summarization, cross-domain communication, documentation), freeing human translators for the institutional functions, governance, contextualization, and accountability for cross-layer trade-offs, that remain irreducibly human. Without answers, translator impact remains dependent on heroic individuals rather than becoming a reproducible design principle. Conclusion Two decades of investment in algorithms, then data, then tools. The bottleneck persists because it was never any of these. It is the infrastructure that connects them: the feedback loops, the governance, the trust, the workforce, and the alignment mechanisms that convert data and AI into continuous learning. The learning utility is a commitment to permanence, a decision that learning from clinical data is as core to a health system’s mission as delivering care (Box 3). The window is now to engage the broader healthcare and research community, especially the informatics community, which is primed to advance this agenda73. Data infrastructure has matured, AI capabilities have accelerated (Introduction: the promise, the progress, and the gap), yet the infrastructure gap is widening, not closing. The technical knowledge to build a learning utility exists; project plans can be assembled. What is missing is the alignment, a shared output measure (the GDP of learning) and a shared commitment to deploy it. If the field does not build Layers 2–3 (Machinery and Enabling Environment, the feedback loops, monitoring, governance, and trust) now, AI deployments will calcify into decision-support tools without learning capability, locking in a generation of infrastructure that tracks efficiency but cannot learn. The longer we wait, the harder it becomes to retrofit feedback loops, governance, and trust in systems designed without them. The specific call is simple: start measuring learning. Adopt or develop a GDP-like metric for health systems. If we cannot answer “how fast is our system learning?” and “are our learning loops closing?”, we do not have a learning utility; we have a collection of AI tools. The IOM’s vision of a system that learns at the speed it delivers care remains the right aspiration. But the speed of learning is the speed of trust, and trust is built by the hard institutional work this editorial has described, not by better algorithms.