1 Overview
At most academic medical centers right now, AI is being deployed faster than the governance structures meant to oversee it. The 2025 CHIME/Censinet survey found that 84 percent of U.S. health systems have some form of AI steering committee — but only 10 percent maintain an automated inventory of which AI tools are actually running in clinical environments (College of Healthcare Information Management Executives and Censinet 2025). The committee exists. The operational visibility does not.
That gap is not a small administrative oversight. Ambient documentation tools are now used by more than 600,000 clinicians across U.S. health systems. Epic has embedded AI across the clinical workflow at thousands of hospitals without requiring most of those hospitals to make a deliberate procurement decision. Research teams are running LLM-assisted literature synthesis, grant applications, and protocol drafts whether or not anyone has worked through the research integrity implications. Administrative staff are using consumer AI for tasks that touch sensitive data, often without knowing what the privacy exposure actually is. The question is no longer whether AI has been deployed at your AMC. The question is whether anyone knows what has been deployed, and whether the governance structures are in place to ensure it is being used well.
This book exists because the answer to that second question, at most institutions, is not yet yes.
This book is fully compliant with the llms.txt standard. Every page is available in plain Markdown format for use with AI assistants and large language models. See the Quarto llms.txt documentation for details.
1.1 What changed between 2020 and 2024
Previous waves of AI in healthcare did not land like this. Expert systems in the 1980s and 1990s required custom development and remained isolated experiments. The early machine learning wave of the 2010s produced promising models in research settings — genomic risk prediction, sepsis alert systems, radiology AI — but deployment at scale was slow, expensive, and required deep integration with clinical informatics teams. The tooling was hard. The data pipelines were fragile. Most models never left the institution that built them.
What changed was the API. Starting in 2023, Microsoft Azure OpenAI, AWS Bedrock, and Google Vertex AI began offering foundation model access through enterprise APIs with signed Business Associate Agreements, U.S.-only data residency options, and zero-data-retention configurations for prompt content. The computational barrier to deploying a capable language model collapsed. An AMC with no machine learning infrastructure could connect a clinical workflow tool to a frontier model in weeks rather than years.
The second change was the nature of the tools. Previous clinical AI was episodic and discrete: this radiology image, this EHR record, this risk score calculated at a specific decision point. The new AI is ambient and continuous. An ambient documentation system is active in every patient encounter, listening to a conversation, and generating a clinical note that the physician then attests to as their professional documentation. A predictive readmission model runs on every patient in the hospital, updating continuously as new data arrives. A care gap identification algorithm touches every patient in the panel, every night. These tools do not generate outputs at discrete moments when a clinician is paying attention. They operate continuously, at scale, in the background.
That shift from episodic to ambient changes everything about governance. A governance model designed for discrete, intentional AI queries does not cover a system that is continuously analyzing every patient encounter without any individual triggering event.
1.2 The evidence: what is actually working
The ambient documentation case is the clearest current evidence of clinical AI value. Studies of early adopters at academic medical centers have found consistent reductions in documentation time — typically 10 to 15 minutes per patient encounter — alongside improvements in note quality and physician satisfaction (Tierney et al. 2024). The AMA’s 2023 survey found that most physicians who reported using AI tools had positive perceptions of their impact on efficiency, though they were more skeptical about the tools’ accuracy and their own ability to verify AI-generated content (American Medical Association 2023). The burnout implications are real: documentation burden is a primary driver of physician attrition, and attrition at AMCs is measured in millions of dollars per departing physician. Ambient documentation is not just a convenience tool. It is a workforce retention intervention.
Diagnostic AI in radiology and pathology has accumulated the strongest clinical evidence base outside of documentation. FDA-cleared AI tools for diabetic retinopathy screening, mammography triage, pulmonary nodule detection, and stroke identification have demonstrated performance at or near specialist-level accuracy in prospective validation studies. The evidence for diagnostic AI reducing time-to-diagnosis and, in some cases, improving outcomes in underserved populations where specialist access is limited is compelling enough that the liability question has begun to cut both ways: institutions may face exposure not only for harms caused by deploying AI but, increasingly, for failing to deploy AI tools that have become part of the standard of care for specific diagnostic tasks.
1.3 The evidence: where it has gone wrong
The counterpart to this evidence base is a set of high-profile failures that share a common diagnosis. IBM Watson for Oncology was deployed at major cancer centers with confident marketing claims about its ability to recommend treatment plans. Physicians at several institutions found its recommendations unsafe, based on synthetic training cases rather than real patient records, and in direct conflict with clinical judgment. The product was eventually discontinued. The failure was not primarily algorithmic. It was a governance failure: inadequate validation against the populations and workflows where the tool was actually deployed, and institutional decisions made on the basis of vendor claims rather than independent evidence.
The Epic Sepsis Model story is more instructive because it involves a tool that was widely deployed across real clinical environments and subjected to rigorous external validation. When Wong and colleagues at the University of Michigan validated the model against their own patient population, they found that its area under the curve — 87 percent in the vendor’s reported validation — dropped to 63 percent in their environment (Wong et al. 2021). More significantly, when they analyzed what the model was actually predicting, they found it was largely capturing patients who were already suspected of having sepsis and who had already had diagnostic cultures ordered. As a predictive tool that could trigger earlier intervention, it was performing close to chance. The model was doing something, just not what the deployment decision assumed it was doing.
This pattern — a tool that performs well on a vendor-provided validation set and underperforms in the real clinical environment — is not an Epic-specific failure. It is a predictable consequence of deploying models without independent local validation. The validation set reflects the population and workflow context where the model was developed. Your population and workflow are different. Sometimes the difference is small. Sometimes it is the difference between an 87 percent AUC and a 63 percent AUC.
The algorithmic bias literature documents a third failure mode that is less about performance in aggregate and more about who the performance failures fall on. Obermeyer and colleagues’ demonstration that a commercial risk stratification algorithm systematically underestimated the health needs of Black patients — because it used healthcare cost as a proxy for health need, encoding unequal access into the model’s outputs — remains the clearest published example of how a technically functional AI tool can produce inequitable outcomes (Obermeyer et al. 2019). The algorithm was working as designed. The design encoded an injustice.
1.4 Where governance stands right now
The governance response to this evidence — both the successes and the failures — has been substantial in scope and uneven in implementation. A handful of academic medical centers have built genuinely operational AI governance programs. Duke Health published a framework for Algorithm-Based Clinical Decision Support oversight that treats deployed algorithms as clinical assets with full lifecycle management requirements: a clinical owner, a technical owner, a silent evaluation phase before any tool influences clinical decisions, and a registry that maintains visibility into every algorithm in the environment (Bedoya et al. 2022). UCSF developed a Trustworthy AI playbook grounded in six operating principles — Fair, Robust, Transparent, Responsible, Privacy, and Safe — with mandatory checkpoints at data validation, pilot deployment, and enterprise scale. Vanderbilt built a REDCap-based intake process that applies structured triage to every AI tool proposal before it consumes governance committee bandwidth.
These are working models. They are not yet the norm. The same CHIME/Censinet survey that found 84 percent of health systems with AI governance committees found that only 59 percent have a formal intake process for evaluating new AI tools, and only 10 percent have automated inventory of what is actually deployed (College of Healthcare Information Management Executives and Censinet 2025). The governance aspiration is widespread. The operational machinery is not.
At the same time, the regulatory environment has moved from guidance to enforcement. The ONC Health Data, Technology, and Interoperability rule took effect in 2025, requiring EHR vendors to surface 31 structured source attributes — training data provenance, demographic performance breakdowns, known limitations — for every certified AI-enabled clinical decision support tool in the workflow. The HHS Section 1557 nondiscrimination rule now holds covered entities liable for deploying patient care decision-support tools that produce discriminatory outcomes. Colorado’s AI Act requires annual impact assessments for high-risk AI. These rules are described in detail in Chapter 10. For present purposes, the point is that the option of deploying AI and revisiting governance later is closing.
1.5 What this book is and isn’t
This is a working framework, not a finished playbook. It was developed for a specific context — an academic medical center trying to organize the deployment of AI tools across four semi-independent organizational domains (clinical care, research, education, and business operations) while maintaining coherent governance — and that context shapes every recommendation in it.
The framework is organized around the recognition that an AMC is not a single AI deployment environment. It is four. Clinical AI governance is shaped by patient safety obligations, FDA regulation, and EHR integration realities that have nothing to do with research AI governance. Research AI governance is shaped by IRB requirements, data sharing agreements, and publication integrity standards that are irrelevant to the education domain. Each domain has its own risk profile, its own leadership structure, its own budget authority, and its own pace of adoption. A governance program designed for clinical AI and applied wholesale to educational uses will miss things that matter. The converse is equally true.
Across all four domains, the same five operational questions recur: who can access what data under what conditions, how is the technical infrastructure governed and secured, what are the ethical and legal obligations, how does the workforce develop the competency to use AI responsibly, and who manages the AI program across its full lifecycle from intake to decommission. These five questions are the workstreams that cross-cut the domain structure, and they are the organizational scaffolding of this book.
The chapters that follow are written to be useful independently as well as together. A CMIO who needs to stand up a clinical AI governance program can read the clinical chapter, the infrastructure chapter, and the project management chapter without reading everything in between. A CHRO who needs to build a workforce AI literacy program can read the workforce chapter without needing the data governance chapter. The cross-references are there for context, not for prerequisite reading.
I wrote the first version of this framework in 2023, when the dominant institutional question was still “should we let people use ChatGPT?” That question has been overtaken by events. The question now is how to govern AI programs that are already real, already large, and already affecting patients, researchers, students, and staff in ways that most institutions do not yet have full visibility into. This version of the book tries to be useful to that question. Whether it succeeds is something you will be better positioned to judge than I am.