Abstract and hypothesis

A provocative leadership claim, tested against the implementation evidence

Healthcare organizations are investing rapidly in artificial intelligence while many of the processes selected for automation remain poorly mapped, weakly owned, and burdened by accumulated workarounds.

Four research questions, four executive answers

What the evidence base looks like

What this dashboard adds to the report

Every number below is computed live in your browser from the report’s own published figures and formulas. Nothing is sent anywhere.

Where the derived findings sit

All three are author-derived from the report’s published values and are labeled as such wherever they appear. They are arithmetic consequences of the report’s own numbers, not new empirical claims.

Central thesis

The report this dashboard is built from

The report runs to nine figures, twenty-nine tables, three original frameworks, a twelve-study evidence matrix, a twelve-metric dictionary, and a twenty-item pre-automation checklist. Every one of those elements appears in this dashboard, either as the published figure, as an interactive rebuild, or as a working tool.

Where a figure prints its own values clearly, it is shown as published rather than rebuilt. Where an interaction, a derived series, or legibility justifies a rebuild, the rebuild sits alongside a toggle that reveals the published original, so the two can always be compared.

The three original frameworks are labeled throughout as proposals requiring local testing. The report is explicit that they are not validated instruments, and this dashboard repeats that caveat at the point of use rather than burying it in a note.

Keywords

artificial intelligencehealthcare leadershipworkflow redesignsociotechnical systemshuman factorsimplementation scienceclinical governancepatient safetyworkforce well-beinghealth equity

The industrialization thesis

Automation changes the production characteristics of a workflow before it changes its purpose

Industrialization is used here in a precise operating sense. AI can make a process faster, more repeatable, cheaper per transaction, more scalable, and less dependent on individual effort. Those properties are valuable only when the process produces the right result. If the work is unnecessary, the decision target is wrong, exceptions are unsafe, or responsibility is diffuse, industrialization widens the defect’s reach.

The unit of analysis, therefore, cannot be the model alone. Sittig and Singh’s sociotechnical model describes health information technology as interacting with hardware and software, clinical content, people, workflow and communication, organizational policy, external rules, and measurement. The SEIPS tradition similarly links work-system design to process, patient, professional, and organizational outcomes. NASSS adds the condition, technology, value proposition, adopters, organization, wider context, and adaptation over time.

Across these frameworks, technology is one component of a complex adaptive system, not an independent treatment.

The same six capabilities, two different outcomes

Each AI capability listed below is neutral. What determines its effect is the condition of the workflow it enters. Use the switch to see the report’s paired reading of the same capability.

The two implementation sequences

The report’s first figure sets up the whole argument: the same organization, the same technology, two different orders of operations. The interactive rebuild lets you step through either track and see what each step assumes.

Product-first and workflow-first sequences, stepped

Select a track, then step through it. The dashed return line on the product-first track is the report’s point: the result feeds back into the tool decision rather than into the process.

Product-first trackWorkflow-first track

Figure rebuild. Interactive version of the report’s Figure 1. Source: Original author framework. The sequence is normative and has not been validated as a comparative instrument.

View the published Figure 1

How the evidence was weighted

The report used a structured critical synthesis rather than a de novo systematic review. Priority was given to peer-reviewed systematic and scoping reviews, real-world implementation studies, randomized and quasi-experimental evaluations, human factors experiments, consensus guidelines, and foundational sociotechnical frameworks. The 2026 American Hospital Association workforce scan was treated as industry context, not equivalent to peer-reviewed evidence. The 2025 JAMA Summit report was treated as a multidisciplinary consensus synthesis.

Limitations of the method include the absence of protocol registration, dual independent screening, formal risk-of-bias scoring for every included source, and quantitative meta-analysis. The report is suitable as an executive research synthesis and as a journal-ready foundation, but a target journal may require conversion into a formal systematic, scoping, or realist review.

Healthcare already has workflow debt

AI does not enter a clean production system

It enters layered policy, technology, and workarounds.

Workflow debt is the accumulated burden that arises when new requirements, fields, approvals, alerts, handoffs, and workarounds are added without removing obsolete work or redesigning the entire pathway. Like technical debt, it may be locally rational and globally damaging. Each workaround keeps care moving today while making tomorrow’s process harder to understand, measure, and govern.

The electronic health record makes the debt visible

Six debt patterns, and why each resists direct automation

Score your own debt exposure

Mark the patterns you can observe in the workflow you are considering for automation. This is a structured reading of the table above, not a validated instrument.

AI often enters too late in improvement

The product decision should follow process design, not substitute for it.

Many organizations begin with a capability inventory: summarization, prediction, computer vision, coding, scheduling, or agentic execution. They then seek use cases that justify a product already selected. This sequence imposes a design constraint on the current workflow. The organization asks where the technology fits rather than whether the work should exist, who should do it, and what result the patient needs.

Implementation reviews repeatedly identify workflow fit, stakeholder engagement, local adaptation, infrastructure, reflection, and evaluation as determinants of uptake. Preti and colleagues reviewed 34 empirical implementations and found recurring importance of design, relative advantage, complexity, access to knowledge, infrastructure, stakeholder engagement, and evaluation. Nair and colleagues combined 38 empirical cases with 69 stakeholder interviews, organizing barriers and strategies across leadership, change management, engagement, workflow, resources, legal issues, training, data, monitoring, maintenance, and ethics.

Reversing the buying sequence

Six questions that decide whether you are running product-first or workflow-first

The buying sequence is often reversed: a product is selected, use cases are solicited, and the current workflow becomes an immutable constraint. This converts transformation into tool insertion.

The report’s diagnostic is unusually practical. You do not need a maturity survey to find out which sequence your organization is running. You only need to listen to the question being asked in the room. Each product-first question below has a workflow-first replacement that answers something the first question cannot.

The full replacement table

What the replacement costs you

The E-R-A-A workflow-first AI model

A four-stage sequence for deciding whether, where, and how to automate

E-R-A-A is a sequencing discipline. It does not assume that every process requires a long redesign or that AI must be delayed until perfection.

It requires leaders to remove work that should not be done, stabilize the necessary process, make responsibilities explicit, and test the actual process before scaling automation. The navigator below runs that sequence as a gate: it will not report a later stage as reachable while an earlier gate is open, and it names the specific item that is holding the sequence.

The gate definitions the navigator enforces

Is this a good first workflow to attempt?

Sequencing tells you what order to work in. It does not tell you which workflow to start with. The report’s portfolio guidance names six characteristics of an early candidate. The screen below reports how many the selected workflow satisfies, and is deliberately separate from the gate logic above.

Scoring note: the six characteristics are scored here as present or absent for each named workflow, using the report’s own descriptions of those workflows. The report does not publish per-workflow scores, so these are editorial readings offered as a starting point for local discussion, not findings.

The efficiency evidence is more fragile than headlines suggest

Savings measured within a single task may disappear across the entire work system

Efficiency is one of the most common executive justifications for healthcare AI, but its measurement is frequently narrow. A model may shorten interpretation, drafting, or review while creating new work in verification, exception handling, patient communication, downstream testing, appeals, training, maintenance, or incident response. If the saved minutes cannot be translated into released capacity, lower cost, better access, or improved care, the result is local acceleration rather than system value.

Wenderott and colleagues screened 13,756 records and included 48 original studies of AI in routine medical imaging. Of 33 studies measuring task time, 67% reported reductions. Yet three meta-analyses using 12 studies found no statistically significant effect, and only three studies closely examined how time saved translated into workload. The review identified multiple workflow patterns, underscoring that the effect of AI depends on where and how it changes work.

What the funnel arithmetic shows

Five measurement levels, and what each one can miss

The report’s measurement ladder is the reason a task-time claim cannot stand alone. Select a level to see what it captures and what it cannot see.

Total work, not task time

The Net Work Ledger

The report sets a minimum standard for any efficiency claim. This ledger is that standard, made arithmetic.

An efficiency claim is permitted only when net total work per completed episode falls, and only when the released capacity is shown to have gone somewhere. The ledger below takes the task-level saving an organization is being sold, adds the work the report says accompanies it, and reports both numbers side by side so the difference is visible rather than assumed.

Planning without learning produces brittle implementation

Organizations are better at approving AI than at adapting it

A systematic review of 25 AI implementation frameworks mapped their content to the Plan-Do-Study-Act framework. Planning appeared in 84% of frameworks, study in 60%, doing in 52%, and acting in only 24%.

The pattern matters. Healthcare AI changes as data, clinical practice, interfaces, staffing, thresholds, and user behavior change. A framework that ends at go-live treats deployment as a finish line when it is actually the beginning of evidence generation.

Geography of the guidance itself

The same review found geographic inequity in authorship: only one of 172 authors was from a low-income country, and two were from lower-middle-income countries. Frameworks built primarily in well-resourced settings may understate constraints related to infrastructure, staffing, data completeness, maintenance, and access.

The challenge is not only to develop more guidance but to test whether governance is usable in the settings where AI will operate.

The lifecycle the frameworks stop short of

Where implementation frameworks weaken, the report substitutes a lifecycle with an evidence question and a required control at every stage. Select a stage.

Alerts at scale: sepsis as a sociotechnical case

The same clinical intent can produce very different operating consequences

Sepsis prediction illustrates why algorithm evaluation and implementation evaluation cannot be separated. In an external validation of a widely implemented sepsis model, Wong and colleagues analyzed 38,455 hospitalizations. Hospitalization-level AUROC was 0.63. At a threshold of six or higher, the model alerted on 18% of hospitalizations, failed to identify 67% of sepsis cases, and had a positive predictive value of 12%. The burden of alert evaluation was therefore part of the clinical effect.

A later before-and-after quasi-experimental study of COMPOSER involved 6,217 adult patients with sepsis across two emergency departments. Deployment was associated with a 1.9 percentage-point absolute reduction in in-hospital mortality, a 5.0 percentage-point absolute increase in bundle compliance, and lower organ dysfunction. The surrounding implementation included a nurse-facing alert, standardized response options, nurse-to-physician communication, stakeholder engagement, and continuous monitoring of data and performance. Only 5.9% of alerts were dismissed during the intervention period. Benefits were not uniform between the two hospitals, reinforcing the role of context.

What the integrated implementation added

The two implied baselines are arithmetic consequences of reporting the same effect in both absolute and relative terms. They reconcile the report’s text with its Figure 5 and are labeled author-derived. They are not values the source studies publish, and they are sensitive to the rounding of the relative figures.

The intervention is not the model

Generative AI: relief, expansion, and cognitive debt

The fastest-growing use cases show why output volume is not equivalent to value

Generative AI can relieve a meaningful burden, particularly when it converts an encounter into a draft note or provides a starting point for repetitive communication. It can also increase the quantity of text that humans must verify. The result may be lower creation effort but higher reading, correction, and information-retrieval burden.

This is a form of cognitive debt: future clinicians inherit more fluent material whose accuracy, relevance, and provenance must still be judged.

The studies behind the bars

Four questions to answer before scaling either use case

Human accountability cannot be automated

Human-in-the-loop is a location in a diagram, not a reliable control

Healthcare AI governance often relies on a reassuring phrase: a clinician remains in the loop. That phrase does not specify whether the clinician has time, expertise, independent information, authority, interface support, or a realistic opportunity to disagree. Oversight imposed as a final click can become ceremonial verification, especially when automation volume is high and errors are uncommon enough to weaken vigilance.

Gaube and colleagues studied 138 radiologists and 127 internal or emergency medicine physicians who reviewed chest radiographs with accurate or inaccurate advice, labeled as coming from either AI or a radiologist. Diagnostic performance was shaped by advice accuracy rather than by the purported source of the advice. Among internal or emergency medicine physicians, 41.73% followed all incorrect advice they received; 27.54% of radiologists did the same.

Eight dimensions that make oversight real

The report replaces the phrase with a specification. Each dimension below is a question that has an answer or does not; there is no partial credit in the source table.

The six-condition oversight test

The report states that meaningful human oversight requires independence, information, competence, time, authority, and feedback, and that if one is absent the organization should treat the control as weak and add system-level safeguards. That is a rule with no partial credit, so this tool does not award any.

AI can industrialize inequity

Bias is often a workflow and objective problem before it is a model problem

AI learns from recorded care, but recorded care is not the same as need. It reflects who gained access, which services were paid for, what was documented, which populations were studied, and how organizations defined success. A model can therefore be accurate for a target that reproduces unequal systems.

Obermeyer and colleagues examined a widely used population health algorithm that used cost as a proxy for health need. Because less money was spent on Black patients with the same level of illness, the algorithm assigned them a lower risk. Correcting the objective would have increased the share of Black patients receiving additional help from 17.7% to 46.5%. The defect was not simply statistical. It was a mis-specified operational objective embedded in a resource-allocation workflow.

Six sources of inequity, each with a workflow manifestation

The report’s structure is deliberate. Every row names where the problem becomes visible in the work, not only where it originates in the data.

Objective validity check

The report’s central equity argument is about the target variable, so this check asks about the target rather than about model performance. It returns a reading, not a score.

Proposed AI Workflow Readiness Index

A 100-point internal governance screen with noncompensable safety gates

The AWRI is an original, provisional instrument for comparing candidate workflows and documenting governance readiness.

It is not a clinical score, a procurement ranking, or a validated prediction model. Organizations should test reliability, validity, gaming risk, and decision usefulness before using it for incentives or external comparison. The calculator below reproduces the published weights exactly and enforces the published stop gates as stops, not as deductions.

The published weights and bands

Derived finding

The band table and the stop gates are two different instruments

The report states both rules plainly, and they are not in conflict when read carefully: the band is an interpretation of readiness, and the gates are a permission to proceed. In practice, a governance committee reading a single number can easily treat the top band as approval. The enumeration below shows how far apart the two rules can sit.

Proposed Industrialized Dysfunction Risk model

A conceptual screening equation, not a validated prediction instrument

The model multiplies four domains: workflow waste, automation scale, propagation speed, and opacity. Its published decision rule is short and unambiguous: if two domains are high, redesign before scale. The report is explicit that the multiplicative form is illustrative and that this is not a calibrated risk score.

Derived finding

The multiplication and the rule rank differently

Scored 1 to 5, the four domains produce 625 profiles. If the product were used as a ranking, it would order those profiles differently from the published rule. The two disagree on 174 of the 625, which is 27.84%.

Governance for the human-plus-AI system

The board governs the clinical operating model, while management runs a continuous control loop

The 2025 JAMA Summit report emphasized holistic, continuous, multistakeholder lifecycle management and noted that widely adopted tools often have limited evaluation of health effects. The report also observed that effectiveness depends on the human-computer interface, user training, and setting, and that no overarching accountability structure spans the many stakeholders. FUTURE-AI similarly organizes trustworthy AI across fairness, universality, traceability, usability, robustness, and explainability, with practices spanning design through monitoring.

Six governance layers, and the question each one owns

Select a layer to see the accountability it holds and the questions it must be able to answer.

Stakeholder benefit, burden, and the protection each one requires

The 2026 American Hospital Association workforce scan states that AI works best when paired with redesigned processes. The report treats this as industry context rather than peer-reviewed evidence, and notes that it aligns with the peer-reviewed implementation literature.

The board dashboard

Measure value, work, and harm

Adoption is not the outcome. Activity counts can reward the industrialization of the wrong process.

Measures should be stratified by site, service, population, urgency, role, and time when sample size permits. A favorable average can conceal failure at a single hospital, in a burdensome specialty, or within an underserved population. Trend limits and decision thresholds should be established before leaders know whether results are favorable.

Ten domains a board should see

Vanity metric translator

The report pairs five commonly reported metrics with what should replace them. Select the metric your reporting pack currently leads with.

Metric dictionary

Proposed operational definitions for local governance and benefits realization. Filter by the direction each metric should move, or by owner.

Implementation roadmap

A staged operating model that begins with elimination and evidence, not enterprise licensing

The first portfolio should favor workflows with a clear patient outcome, high avoidable burden, measurable baseline, stable data, bounded action, strong ownership, and reversible deployment. Avoid beginning with high-autonomy, high-severity decisions, workflows with weak data provenance, or processes whose demand and exception patterns are not understood.

Discussion

The decisive question is not whether healthcare should use AI, but what operating system AI should amplify.

The central hypothesis is supported by a consistent pattern: healthcare AI value emerges from the interaction of model, workflow, people, resources, interface, accountability, and learning. This pattern appears in foundational systems theory, contemporary implementation reviews, human factors experiments, generative AI field studies, governance consensus, and contrasting real-world outcomes. The evidence does not justify rejecting AI. It justifies rejecting automation as a substitute for improvement.

The title’s force comes from a genuine asymmetry. Once automation is embedded, scale and sunk cost create pressure to preserve it. Users adapt, policies are written around it, vendors integrate it, and metrics begin to reward activity. A flawed manual workflow is visible in queues and complaints. A flawed automated workflow can appear efficient because the activity is rapid and its costs are dispersed across review, exceptions, downstream care, patient effort, and rare harms.

Conclusion

Do not automate accumulated dysfunction and call it transformation.

Artificial intelligence can improve healthcare. The evidence supports real gains in documentation burden, prediction-enabled response, and selected clinical processes. But those gains are conditional. AI does not decide whether a task is necessary, whether a proxy is morally and clinically valid, whether an exception has an owner, whether a reviewer has time, or whether released capacity returns to patients. Leaders must make those decisions.

The workflow-first sequence is straightforward: eliminate low-value work, redesign the complete pathway, assign clinical and operational accountability, and automate only the validated process. This sequence changes the leadership conversation from how quickly a tool can be deployed to whether the organization is prepared to amplify the process it has built.

Claims ledger

What the evidence supports, and what it does not

The report grades its own claims rather than presenting them as uniform findings. Two of the six are explicitly not established, including a weaker form of the report’s own title.

Evidence matrix

The twelve studies most directly informing the hypothesis, each with the limitation the report records against it. Filter by design.

Limitations and research agenda

The practical framework precedes direct comparative evidence and should be tested accordingly.

The report’s hypothesis is not supported by randomized comparisons of E-R-A-A against product-first implementation. Many cited studies are observational, pre-post, quality improvement, simulation, or review designs. Real-world AI interventions are bundles, making it difficult to isolate the model from training, interface, staffing, communication, and concurrent improvement. Publication bias may favor successful implementations. Vendor evolution also makes results time-dependent.

The original models are provisional. E-R-A-A has face validity from implementation and systems evidence but requires empirical evaluation. The Industrialized Dysfunction Risk model is conceptual, and its multiplicative form is illustrative. The AWRI weights and thresholds are proposed for internal testing and should not be treated as psychometrically established.

Verification ledger for this dashboard

Every figure this dashboard derives, and whether it reproduces the report’s published value.

References

Peer-reviewed and authoritative sources cited in the report. DOI and source links are live.

• Adams, R., et al. (2022). Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nature Medicine, 28, 1455-1460. doi.org/10.1038/s41591-022-01894-0
• Angus, D. C., Khera, R., Lieu, T., et al. (2025). AI, health, and health care today and tomorrow: The JAMA Summit report on artificial intelligence. JAMA, 334(18), 1650-1664. doi.org/10.1001/jama.2025.18490
• Arndt, B. G., Beasley, J. W., Watkinson, M. D., et al. (2017). Tethered to the EHR: Primary care physician workload assessment using EHR event log data and time-motion observations. Annals of Family Medicine, 15(5), 419-426. doi.org/10.1370/afm.2121
• Beede, E., Baylor, E., Hersch, F., et al. (2020). A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1-12. doi.org/10.1145/3313831.3376718
• Boussina, A., Shashikumar, S. P., Malhotra, A., et al. (2024). Impact of a deep learning sepsis prediction model on quality of care and survival. npj Digital Medicine, 7, 14. doi.org/10.1038/s41746-023-00986-6
• Cabitza, F., Rasoini, R., & Gensini, G. F. (2017). Unintended consequences of machine learning in medicine. JAMA, 318(6), 517-518. doi.org/10.1001/jama.2017.7797
• Carayon, P., Schoofs Hundt, A., Karsh, B. T., et al. (2006). Work system design for patient safety: The SEIPS model. Quality & Safety in Health Care, 15(Suppl 1), i50-i58. doi.org/10.1136/qshc.2005.015842
• Duggan, M. J., Gervase, J., Schoenbaum, A., et al. (2025). Clinician experiences with ambient scribe technology to assist with documentation burden and efficiency. JAMA Network Open, 8(2), e2460637. doi.org/10.1001/jamanetworkopen.2024.60637
• Gaube, S., Suresh, H., Raue, M., et al. (2021). Do as AI says: Susceptibility in deployment of clinical decision aids. npj Digital Medicine, 4, 31. doi.org/10.1038/s41746-021-00385-9
• Garcia, P., et al. (2024). Artificial intelligence-generated draft replies to patient inbox messages. JAMA Network Open, 7(3), e243201. doi.org/10.1001/jamanetworkopen.2024.3201
• Greenhalgh, T., Wherton, J., Papoutsi, C., et al. (2017). Beyond adoption: A new framework for theorizing and evaluating nonadoption, abandonment, and challenges to scale-up, spread, and sustainability of health and care technologies. Journal of Medical Internet Research, 19(11), e367. doi.org/10.2196/jmir.8775
• He, J., Baxter, S. L., Xu, J., Xu, J., Zhou, X., & Zhang, K. (2019). The practical implementation of artificial intelligence technologies in medicine. Nature Medicine, 25, 30-36. doi.org/10.1038/s41591-018-0307-0
• Holmgren, A. J., Hendrix, N., Maisel, N., Everson, J., & Bazemore, A. (2024). Electronic health record usability, satisfaction, and burnout for family physicians. JAMA Network Open, 7(8), e2426956. doi.org/10.1001/jamanetworkopen.2024.26956
• Kamel Rahimi, A., Pienaar, O., Ghadimi, M., et al. (2024). Implementing AI in hospitals to achieve a learning health system: Systematic review of current enablers and barriers. Journal of Medical Internet Research, 26, e49655. doi.org/10.2196/49655
• Khan, A. A., et al. (2024). Implementation frameworks for artificial intelligence translation into health care practice: A systematic review. PLOS Digital Health, 3, e0000514. doi.org/10.1371/journal.pdig.0000514
• Kwong, J. C. C., Nickel, G. C., Wang, S. C. Y., & Kvedar, J. C. (2024). Integrating artificial intelligence into healthcare systems: More than just the algorithm. npj Digital Medicine, 7, 52. doi.org/10.1038/s41746-024-01066-z
• Lekadir, K., Frangi, A. F., Porras, A. R., et al. (2025). FUTURE-AI: International consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ, 388, e081554. doi.org/10.1136/bmj-2024-081554
• Liu, T. L., Hetherington, T. C., Dharod, A., et al. (2024). Does AI-powered clinical documentation enhance clinician efficiency? A longitudinal study. NEJM AI, 1(12), AIoa2400659. doi.org/10.1056/AIoa2400659
• Nair, M., Svedberg, P., Larsson, I., & Nygren, J. M. (2024). A comprehensive overview of barriers and strategies for AI implementation in healthcare: Mixed-methods design. PLOS ONE, 19(8), e0305949. doi.org/10.1371/journal.pone.0305949
• Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. doi.org/10.1126/science.aax2342
• Preti, L. M., Ardito, V., Compagni, A., Petracca, F., & Cappellaro, G. (2024). Implementation of machine learning applications in health care organizations: Systematic review of empirical studies. Journal of Medical Internet Research, 26, e55897. doi.org/10.2196/55897
• Sendak, M. P., Ratliff, W., Sarro, D., et al. (2020). Real-world integration of a sepsis deep learning technology into routine clinical care: Implementation study. JMIR Medical Informatics, 8(7), e15182. doi.org/10.2196/15182
• Sinsky, C., Colligan, L., Li, L., et al. (2016). Allocation of physician time in ambulatory practice: A time and motion study in 4 specialties. Annals of Internal Medicine, 165(11), 753-760. doi.org/10.7326/M16-0961
• Sittig, D. F., & Singh, H. (2010). A new sociotechnical model for studying health information technology in complex adaptive healthcare systems. Quality & Safety in Health Care, 19(Suppl 3), i68-i74. doi.org/10.1136/qshc.2010.042085
• Sriharan, A., Sekercioglu, N., Mitchell, C., et al. (2024). Leadership for AI transformation in health care organizations: Scoping review. Journal of Medical Internet Research, 26, e54556. doi.org/10.2196/54556
• Tai-Seale, M., Baxter, S. L., Vaida, F., et al. (2024). AI-generated draft replies integrated into health records and physicians’ electronic communication. JAMA Network Open, 7(4), e246565. doi.org/10.1001/jamanetworkopen.2024.6565
• van de Sande, D., Chung, E. F. F., Oosterhoff, J., et al. (2024). To warrant clinical adoption, AI models require a multi-faceted implementation evaluation. npj Digital Medicine, 7, 51. doi.org/10.1038/s41746-024-01064-1
• Vasey, B., Nagendran, M., Campbell, B., et al. (2022). Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nature Medicine, 28, 924-933. doi.org/10.1038/s41591-022-01772-9
• Wang, A., Freeman, S., & Magrabi, F. (2026). Governance for safe and responsible AI in healthcare organizations: A scoping review of frameworks. npj Digital Medicine, 9. doi.org/10.1038/s41746-026-02679-2
• Wenderott, K., Krups, J., Weigl, M., & Wooldridge, A. (2024). Effects of artificial intelligence implementation on efficiency in medical imaging: A systematic literature review and meta-analysis. npj Digital Medicine, 7, 265. doi.org/10.1038/s41746-024-01248-9
• Wenderott, K., Krups, J., Weigl, M., & Wooldridge, A. (2026). Navigating the complexity of artificial intelligence implementation in healthcare: Ten recommendations grounded in implementation science. BMJ Open Quality, 15, e003639. doi.org/10.1136/bmjoq-2025-003639
• Wong, A., Otles, E., Donnelly, J. P., et al. (2021). External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Internal Medicine, 181(8), 1065-1070. doi.org/10.1001/jamainternmed.2021.2626
• American Hospital Association. (2026). 2026 Health Care Workforce Scan. aha.org/aha-workforce-scan

Showing all 33 references.