From Pilot to Production: How AI Is Rewiring the Operating Discipline of Health Systems

Twenty-two per cent of health organisations now use some form of domain-specific artificial intelligence, a sevenfold increase over 2024, sixty-six per cent of United States physicians actively use AI tools in practice, and approximately one and a half billion dollars has been deployed at twice the rate of the broader economy. Yet eighty-three per cent of those efforts remain pilots and fewer than ten per cent of organisations have invested in genuine end-to-end production infrastructure. Dr. Arturo LoAIza-Bonilla, Systemwide Chief of Hematology and Oncology at St. Luke's University Health Network, Co-Founder and Chief Medical Officer at Massive Bio, and Associate Professor at the Lewis Katz School of Medicine at Temple University, presents a comprehensive analysis of the pilot-to-production gap in a 1ArtificialIntelligence session hosted by Glenn Tyranski, Partner at 1BusinessWorld and Executive in Residence at King's College. The conversation is structured around three operational layers, the economic flywheel of demonstrated return on investment, the clinical revolution from precision oncology to multi-agentic patient matching, and the human bottleneck of reskilling, governance, and frontline engagement that ultimately determines whether health systems make AI part of their operating discipline or leave it stranded in experimentation.

The pilot-to-production threshold

Healthcare AI has moved past the question of whether the tools work. Twenty-two per cent of health organisations now use domain-specific AI, a sevenfold rise over the prior year. Approximately one and a half billion dollars has flowed into deployment, at roughly twice the rate observed in the broader economy. Two-thirds of physicians in the United States are now active users of AI tools in their daily practice. The question is whether organisations are prepared to deploy at enterprise scale.

The same data tells the harder half of the story. Eighty-three per cent of those efforts remain pilots, and fewer than ten per cent of organisations have invested in the production-grade infrastructure required to scale. LoAIza-Bonilla organises the analysis around three layers of operational impact: the economic flywheel created by demonstrated return on investment, the clinical revolution underway in precision oncology and multi-agentic patient matching, and the human bottleneck of reskilling, governance, and frontline engagement that ultimately determines whether AI becomes part of operational discipline or remains a curiosity.

The economic flywheel and the undeniable return

Every dollar invested in healthcare AI now produces approximately three dollars and twenty cents of tangible return over a fourteen-month maturation period, with eighty-five per cent of healthcare leaders reporting active revenue increases directly tied to AI integrations. Each successful deployment creates the data, governance scaffolding, and organisational confidence required to deploy the next, which is why LoAIza-Bonilla treats the return as a flywheel rather than a one-time benefit case.

Legacy friction and AI-enabled velocity

The operational changes show up across three dimensions. In revenue cycle management, sixty per cent of denials historically went un-appealed because the manual cost of pursuing them outweighed the recovery, and ninety-three per cent of physicians have at some point been on the phone arguing prior authorisation cases. Cohere Health, working with Highmark, has demonstrated ninety per cent auto-approval rates with integrated AI tooling and an eighty-five per cent reduction in personnel involved in the prior-authorisation process. In supply chain, machine learning is replacing reactive responses to clinical shortages with predictive alerts that flag stockouts of medications, equipment, and spare parts before operations are affected. In documentation and coding, optical character recognition and natural language processing are now being layered with generative AI, deployed natively inside electronic medical records so that lab data, clinical notes, genomics, monitoring data, and medical images can be combined into foundation models that support decision-making rather than simply recording activity.

Operational friction in healthcare has been treated as inevitable for decades and is now negotiable. The same automation logic that reshaped revenue cycle a decade ago is now extending into clinical workflow, and LoAIza-Bonilla argues this is the year that organisations either close the production-readiness gap or fall structurally behind.

The electronic medical record enters the AI era

The largest electronic medical record vendors have moved from observers to operators in the AI economy. Epic, with approximately fifty-four per cent share of all hospital beds in the United States, now has around two hundred AI features in development, including ambient documentation in partnership with Abridge and Microsoft Dragon Copilot, the patient-facing assistant Emmy, the revenue cycle assistant Penny, and the clinical decision support assistant Art. Beneath these consumer-facing assistants sits Cosmos, an analytical environment built on approximately three hundred million records, and Comet, a foundation model pre-trained on roughly sixteen billion clinical events. Oracle, having acquired Cerner, is taking a parallel path with a cloud-native, voice-first agentic next-generation EHR that targets a thirty per cent reduction in clinical documentation time and deploys clinical agents across multiple specialties including oncology. The Food and Drug Administration has now cleared more than one thousand four hundred AI and machine learning devices, with projected growth of approximately three hundred and thirty per cent over the next five years.

AI inside the medical record is no longer a feature delivered by external vendors. It is becoming the connective tissue of the record itself, and health systems that treat AI capability as separable from EHR strategy will find themselves negotiating that separation against vendors who have already integrated.

Ambient AI, the first mass adoption win

Ambient AI is the clearest example of healthcare's first mass adoption success. Kaiser Permanente, in partnership with Abridge, has deployed ambient documentation across forty hospitals and six hundred medical offices, saving approximately fifteen thousand seven hundred and ninety-one clinician hours of typing in the first year alone. Across the broader market, ambient scribe revenue has reached approximately six hundred million dollars at two and a half times year-over-year growth, with Abridge, Microsoft Nuance, and Dragon Copilot among the leading platforms. Reductions in clinician burnout are substantive, with seventy-four per cent lower odds of burnout reported among ambient AI users and seventy per cent improvement in workflow performance reported across organisations deploying AI in process redesign.

What ambient AI gives back is harder to measure than what it saves. Ambient AI restores the clinician's eye contact with the patient, removing the keyboard from the consultation and allowing the physician to listen rather than transcribe. LoAIza-Bonilla treats this as the most important operational impact AI will have, the restoration of humanity to the clinical encounter that the previous generation of electronic medical records inadvertently displaced.

The clinical revolution and seeing the unseen

Generative models alone are not enough for clinical work. In radiology, convolutional neural networks operating on actual data sets rather than generated images are now outperforming human readers in tasks such as pancreatic cancer detection, where AI achieves approximately thirty-eight per cent fewer false positives and identifies subtle early signs that would otherwise be difficult to determine. In radiation oncology, the replanning workflow that historically required several days has been compressed to minutes, with AI handling the patient-anatomy adjustments and breathing-cycle adaptations that allow radiation beams to remain tightly focused on changing targets. In progressive cancer research, image analysis is replacing manual caliper measurements of tumour nodules with AI-assisted criteria for response assessment.

The clinical-trials gap is the most consequential opportunity. An analysis of approximately twelve million patients in the National Cancer Institute database found that fewer than one tenth of one per cent ever enrol in a clinical trial. Even at mission-driven research centres, enrolment rates run between five and fifteen per cent of patients despite philanthropic and federal investment, because eighty-five per cent of cancer patients are treated in community settings rather than at academic enclaves and because clinicians face data drift, alert fatigue, and time-bound priorities that obscure the connection between a patient and a trial they would qualify for. LoAIza-Bonilla and his co-founders at Massive Bio built Synergy AI as a multi-agentic architecture combining large language model extraction, an oncology knowledge graph, and real-time recommendation. Validation work presented at the American Society of Clinical Oncology meeting matched three thousand eight hundred and four patients in approximately one hour, work that would conventionally have taken approximately fifty thousand human-hours to complete. A subsequent collaboration with the American Cancer Society now offers any cancer patient in the United States access to the program at no cost.

From generation to orchestration

The next frontier is what LoAIza-Bonilla describes as the move from generation to orchestration. Single-prompt large language models give way to multi-agentic systems in which agents collect database information, communicate with autonomous goal-driven scheduling, billing, and care-coordination systems, and execute on the full operational workflow rather than producing isolated answers. Eighty per cent of healthcare executives expect value from agentic AI, but only three per cent of agents are currently deployed in live workflows. Closing that gap is what separates the systems that will lead from those that will follow.

Health systems setting the pace

The leaders are already visible. Mayo Clinic has invested approximately one billion dollars in foundation models and federated learning architecture and is currently regarded by external rankings as the leading institution in the world. Cleveland Clinic has anchored its strategy in EMR-integrated documentation and clinician workflow. HCA Healthcare has documented an eighty-three per cent reduction in charge entry time and a ninety per cent nurse approval rate for its AI deployments. Nebraska Medicine has reduced length of stay by five per cent for complex cases. Providence has added approximately six thousand additional surgical cases through operating-room optimisation. CommonSpirit has achieved a fivefold increase in care gap closures through AI cancer screening. Tampa General has reported approximately one hundred million dollars in operational savings from sepsis detection. UC San Diego has reduced sepsis mortality by approximately seventy per cent through early-warning models. Where health systems treat AI as core operating capability rather than a tactical experiment, the outcomes compound across cost, quality, and clinician experience.

The human bottleneck and the seventy per cent

The technical layer is the smallest part of the transformation. Approximately ten per cent of the work is the algorithms themselves. Approximately twenty per cent is the data and technology infrastructure that supports them. The remaining seventy per cent is processes, governance, and the human capacity to absorb new ways of working, with sixty per cent of organisations reporting reskilling as their top challenge. Security and clinical informatics teams operate under appropriate caution, but caution applied uniformly to every algorithm regardless of its risk profile creates bottlenecks that prevent legitimate deployments from progressing.

AI's greatest organisational impact will be in giving clinicians their humanity back. LoAIza-Bonilla anchors the case in Moravec's Paradox: capabilities such as language and arithmetic, which seem distinctively human, are evolutionarily recent and computationally tractable, while capabilities such as nuanced empathy, context-sensitive judgement, and complex physical coordination are evolutionarily old and computationally intractable. Algorithms should absorb the recent and tractable work of computation, freeing clinicians to do the older and harder work of empathy and presence. LoAIza-Bonilla calls the resulting model collaborative intelligence.

Ethics, governance, and the data provenance imperative

Three governance pillars must be satisfied for any deployment to earn trust. The first is explainability, including the data provenance and decision pathway behind any AI tool, with different levels of scrutiny calibrated to whether the application is a revenue-cycle tool or a regulated decision-support device. The second is data set integrity, since training data drawn from one population may not generalise to another. LoAIza-Bonilla cites the example of mammography AI trained on one demographic deployed on patients with materially different breast density, where the underlying performance assumptions break down. The third is the human in the loop, since no matter how strong the algorithm, the patient relationship is fundamentally a human one and decisions affecting care must remain anchored in human judgement.

At the executive and board level, the most effective adoption pathways combine champion clinicians who understand both the clinical and operational frames, structured forums in which leading institutions share their work, and clear demonstration that even the most respected institutions in the world are now investing fully. Where boards see Mayo Clinic and Cleveland Clinic committing at scale, the conversation shifts from whether to invest to how to invest responsibly.

The four trends defining 2026 and beyond

LoAIza-Bonilla closes his presentation with four trends that will define the year ahead. First, ambient AI becomes table stakes, present in essentially every health system and no longer a differentiator. Second, revenue cycle AI scales, with the highest measured return on investment of any current category, and organisations that have not deployed will face structural disadvantage. Third, agentic AI moves from pilot to production in the systems that take the transition seriously, with Mayo Clinic, Cleveland Clinic, and St. Luke's among those advancing buy-and-build strategies. Fourth, AI governance becomes critical at the C-suite and board level, replacing the previous era of software implementation with a faster, more rigorous deployment cadence and explicit accountability for explainability, data provenance, and the human in the loop.

Strategic takeaways for health system leaders

Five conclusions carry direct implications for health system strategy. First, the institutions that translate pilots into production capability this year will create durable structural advantages over those that do not, and the experimentation window will not stay open indefinitely. Second, the economic case is established at approximately three dollars and twenty cents of return per dollar invested over a fourteen-month maturation, and revenue cycle, prior authorisation, and ambient documentation are the categories where that return arrives fastest. Third, multi-agentic orchestration is the next operational frontier, and the gap between executive expectation and live deployment is the precise opportunity space for the systems that move first. Fourth, the human bottleneck is the binding constraint: ten per cent algorithms, twenty per cent data and technology, seventy per cent processes and people. Fifth, governance is now a leadership obligation rather than a delegated technical function, with explainability, data provenance, and the human in the loop as the operational pillars that translate AI deployment into trusted institutional practice.

Patient first is the principle that holds the architecture together. Technology has its highest value when it is invisible to the patient, when the clinician is freed to listen, and when the operational system around the encounter delivers the right care to the right person at the right time. AI is no longer futuristic. It is actionable now, and the institutions that recognise that will define the next decade of healthcare delivery for the communities they serve.