Executive synthesis

Artificial intelligence is arriving in healthcare through thousands of apparently sensible decisions: a transcription tool here, an imaging algorithm there, a risk score, a patient chatbot, an automated prior-authorisation process. Each may work. Collectively, they may still fail to transform care.

The recurring error is to confuse technical capability with institutional capacity. A model can generate a useful answer in seconds; a health system must decide whether the answer is trustworthy, who may act on it, who remains liable, which old activity stops, how the saving is measured and who receives the benefit. Those are questions of organisation and power, not computation.

Ari Ercole's Artificial Intelligence, Natural Resistance supplies the sharpest recent formulation: resistance is often rational because the rules of healthcare make adoption personally costly, professionally risky or economically unrewarding.[1] Robert Wachter supplies the clinical history of unintended consequences; Clayton Christensen and colleagues explain why incumbent business models absorb disruptive innovations; Eric Topol and Tom Lawry show the human and managerial promise; Acemoglu and Johnson remind us that productivity is distributed by institutions; and McKinsey, BCG, Bain, Deloitte, Accenture and KPMG translate much of this into operating-model language.[3–14]

The combined conclusion is more demanding than "bring clinicians with you". AI adoption succeeds when an institution makes five explicit commitments:

1. Solve a defined patient or workforce problem, rather than purchase a fashionable capability.

2. Redesign an end-to-end pathway, including the work that will stop, not merely the task that will become faster.

3. Allocate decision rights, liability and escalation before deployment.

4. Share the gains credibly with patients and the workforce that creates them.

5. Measure realised outcomes over time, including hidden work, demand effects, equity and safety.

The machine may work. The system still may not.

Healthcare has a long history of mistaking installation for transformation. Wachter's The Digital Doctor documents how electronic health records improved legibility and access while also converting doctors into data-entry workers, enabling billing priorities to shape the clinical interface and creating new forms of error.[3] The lesson is not that digitisation failed. It is that a technology inherits the purposes and incentives of the institution into which it is placed.

Ercole extends that lesson to AI.[1] A technically capable model enters a setting organised around professional jurisdiction, departmental budgets, procurement rules, regulatory duties, liability, information asymmetries and historical mistrust. It is therefore never "just" an algorithm. It changes who knows, who decides, who checks and who can be blamed.

The seven systems surrounding every model

Implementation science makes this visible. Greenhalgh and colleagues' NASSS framework asks implementers to examine the condition, the technology, the value proposition, adopters, the organisation, the wider institutional context and their mutual adaptation over time.[15] A weakness in any one domain can prevent scale. This is a stronger diagnostic than blaming culture because it forces "culture" to be decomposed into observable constraints.

This reveals why the easiest technologies to adopt are not necessarily the most valuable. An ambient scribe may help one clinician without requiring colleagues to behave differently. A system that redesigns an entire diagnostic pathway requires many actors to adopt consistently before the benefit appears. Ercole describes this as the "value-adoption paradox": the more systemic the value, the greater the coordination burden.[2]

The partial-adoption equilibrium

The usual story treats partial adoption as a temporary stage. Ercole's "partial-adoption trap" is the more unsettling claim that it can be a stable destination.[2] Some people use the new system, others retain the old one, and the organisation preserves both to manage risk. Benefits never cross the threshold required to justify removing legacy work. The technology survives, but transformation does not.

1. A pilot demonstrates that the model can perform a task.

2. The organisation adds it without withdrawing the old process.

3. Clinicians create verification work to protect themselves.

4. Uneven adoption prevents pathway-level benefits.

5. The business case weakens; scepticism rises; the old equilibrium hardens.

BCG's warning that AI will not fix a health system without redesign, McKinsey's criticism of fragmented point solutions and Bain's finding that successful scalers redesign processes all converge here.[9–11] Consulting language calls this an operating-model problem. Ercole's language is sharper: operating models embody interests, and incumbents may benefit from preserving them.[1]

Productivity is a bargain, not a number

AI business cases commonly convert minutes saved into monetary value. That arithmetic is easy and often fictitious. Time is not cash unless the organisation can aggregate it, redesign capacity and change a budget or outcome. Ten minutes saved across many clinicians may be valuable as reduced fatigue and better conversations while producing no cashable saving. That is still real value, but it is a different claim.

The deeper problem is distribution. Acemoglu and Johnson's Power and Progress argues that technological gains are not automatically shared; institutions and bargaining power determine who benefits.[6] In healthcare, that question is unusually complex because patients, professionals, providers, commissioners, taxpayers and vendors may each create or capture different parts of the value.

Four forms of value

A credible business case therefore needs an explicit conversion mechanism. If documentation time falls, will rotas change, overtime fall, visits lengthen, waiting lists shorten or retention improve? Each mechanism requires different management action and produces value for different people.

The trust bargain

Resistance becomes rational when the workforce expects efficiency to be punished. If every saved minute becomes another unit of demand, staff learn that innovation intensifies work. If an algorithm carries institutional authority but the clinician carries personal liability, "human oversight" becomes unpaid risk transfer. If a vendor captures recurring revenue while the provider supplies data, workflow expertise and verification, partnership becomes extraction.

Trust is not a communications campaign. It is a forecast people make from institutional behaviour.

It grows when commitments are specific, enforceable and honoured: protected learning time; no automatic headcount reduction during evaluation; transparent error reporting; shared governance; and a declared allocation of gains.

Demand can consume the gain

BCG highlights a second-order effect that many business cases omit: easier access can generate additional demand.[10] AI-enabled navigation, asynchronous consultation or continuous monitoring may uncover unmet need. That can be clinically desirable while increasing total workload and expenditure. A service can become more productive per encounter yet more expensive overall.

The correct question is therefore not simply "Does this save time?" but "What happens to the system when this becomes easier, cheaper or more available?" Capacity, demand and escalation must be modelled together.

The new clinical division of labour

Topol's Deep Medicine imagines AI returning medicine to its human core.[4] Lawry and the consulting firms translate this into task redesign: machines handle pattern recognition, drafting, retrieval and routine coordination; people contribute judgement, empathy, contextual understanding and accountability.[7,9–14] The aspiration is sound. The boundary is not self-executing.

A task does not arrive neatly labelled "automatable". It contains sensing, interpretation, decision, action, explanation and responsibility. AI may perform some components exceptionally and fail unpredictably at others. The practical unit of redesign is therefore the decision pathway, not the job title and not the model.

The practical unit of redesign is the decision pathway — not the job title, and not the model.

Beyond individual human oversight

"Human in the loop" is often used as an ethical safeguard, but it can disguise an accountability gap. A busy clinician asked to approve hundreds of AI outputs may possess formal authority without the time, information or realistic ability to challenge them. Oversight then becomes ceremonial.

Meaningful oversight requires that the human can understand the relevant basis for the output, refuse it without penalty, obtain additional evidence, escalate uncertainty and learn from aggregate performance. The organisation must in turn monitor drift, complaints, disparities, workarounds and near misses. WHO's principles — autonomy, wellbeing and safety, transparency, accountability, inclusion and sustainability — have to be made operational at these points.[16]

The hazards of a poor division of labour

• Automation bias — staff defer to an apparently authoritative recommendation.

• Deskilling — infrequently exercised capabilities deteriorate, especially in trainees.

• Verification burden — checking machine output takes longer than doing the task well.

• Responsibility without control — clinicians remain liable for systems they did not choose or configure.

• Exception blindness — the workflow performs well for typical cases and poorly for those who do not fit.

• Data feedback loops — decisions influenced by the model become future training data, concealing error.

• "Moral crumple zones" — in Madeleine Clare Elish's formulation, the nearest human absorbs blame for a failure produced by a distributed technological and organisational system.[19]

Professional identity is an implementation variable

Healthcare professions are organised partly around expertise. AI that makes specialist knowledge widely available can improve access while threatening established jurisdictions. Christensen's disruptive-innovation analysis helps explain why incumbents tend to incorporate new technology into high-cost models rather than allow it to create simpler ones. The response should not be to dismiss professional concerns: jurisdiction often carries genuine safety knowledge. But neither should professional status determine whether patients receive a more accessible service.

The practical answer is co-design with explicit decision rights. Clinicians should help determine which tasks transfer, what constitutes an exception, what evidence users see and how competence is maintained. Patients should determine what must remain relational and where automated access is preferable. Leaders must decide, visibly, who owns residual risk.

A practical compact for the healthcare AI ecosystem

No participant can solve institutional resistance alone. But every participant can stop reinforcing it. The following playbooks translate the synthesis into individual choices.

A 12-question pre-deployment challenge

1. What patient, workforce or system problem are we solving, and what is the baseline?

2. Why does AI improve this problem rather than merely digitise it?

3. What complete pathway surrounds the model output?

4. Which existing task or control will stop, and when?

5. Who changes behaviour before value appears?

6. Who receives the benefit and who carries the transition cost?

7. What new demand will easier access or lower friction create?

8. Who holds the decision right, who is liable and who can halt use?

9. Can the designated human actually exercise meaningful oversight?

10. How will performance, drift, equity, workload and workarounds be monitored?

11. What happens if the vendor changes, fails or becomes unaffordable?

12. What evidence after 6 and 12 months would justify stopping, adapting or scaling?

    
    

Six propositions worth debating

The safest AI is not necessarily the most explainable model; it may be the institution most able to detect and correct its mistakes.

• A pilot that preserves every legacy control is designed to prove technical feasibility, not economic value.

• Clinician resistance is sometimes conservatism, sometimes safety intelligence and sometimes rational bargaining. Leadership must distinguish among them.

• Time saved for recovery and humane care is productive value, even when it cannot be removed from a budget.

• The more a technology promises system transformation, the less credible a procurement-led implementation becomes.

• If benefits are private but errors and transition costs are socialised, mistrust is not a barrier to adoption; it is an accurate diagnosis.

  
  

Build the bargain before the model

Healthcare AI is often discussed as a contest between technological optimism and professional resistance. The literature suggests a more useful framing. Technology creates a new set of possible arrangements; institutions determine which arrangement becomes real.

The optimistic case remains strong. AI can reduce clerical burden, make expertise more available, improve coordination, discover patterns beyond unaided cognition and support patients between encounters. But none of those benefits arrives merely because a model performs well. Each depends on changes to workflow, authority, responsibility, finance and behaviour.

The practical task is to construct an institutional bargain in which adoption is safe and rational: patients receive demonstrable benefit and recourse; clinicians gain workable tools, voice and protection; organisations obtain measurable value; vendors carry appropriate responsibility; and society retains legitimate control over data and standards.

That bargain is harder to build than a pilot. It is also where the real innovation lies.

Sources

This is a critical synthesis, not a chapter-by-chapter précis. Numbered references identify the principal sources for distinctive concepts and attributed claims. Consulting reports are useful for operating-model patterns and survey signals but also promote advisory services; their quantitative claims should be read with attention to sample, geography and definitions. Ercole's institutional argument is analytically powerful, but it should not be used to minimise genuine limitations in validation, generalisability, regulation, data quality or model safety.

1. Ari Ercole. Artificial Intelligence, Natural Resistance: How the Institution of Medicine Resists the Technologies That Could Save Us (Latent Press, 2026). ↗

2. Ari Ercole. The partial adoption trap: coordination failure, trust, and cultural lock-in in health AI adoption (2026). ↗

3. Robert M. Wachter. The Digital Doctor (2015); and A Giant Leap (2026). ↗

4. Eric Topol. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again (2019). ↗

5. Clayton M. Christensen, Jerome H. Grossman and Jason Hwang. The Innovator's Prescription (2009). ↗

6. Daron Acemoglu and Simon Johnson. Power and Progress (2023). ↗

7. Tom Lawry. AI in Health: A Leader's Guide (2020). ↗

8. Peter Lee, Carey Goldberg and Isaac Kohane. The AI Revolution in Medicine (2023). ↗

9. McKinsey & Company. The coming evolution of healthcare AI toward a modular architecture (2025); Generative AI in healthcare (2026); Rewired, 2nd ed. (2026). ↗

10. Boston Consulting Group. AI Won't Fix Your Health System. Redesigning It Will (2026); EPRs as enablers, not endpoints (2026). ↗

11. Bain & Company. The Human Imperative: Scaling AI Across Life Sciences (2026); Want More Out of Your AI Investments? Think People First (2026). ↗

12. Accenture. Scaling Gen AI for Healthcare Productivity (2025). ↗

13. Deloitte. From pilots to practice: Scaling AI use in federal health (2025); Health care's quest for an enterprisewide AI strategy (2022). ↗

14. KPMG. From Symptoms to Solutions: Treating AI Implementation Challenges in Healthcare (2025); Intelligent Healthcare. ↗

15. Trisha Greenhalgh et al. Beyond Adoption: the NASSS framework (Journal of Medical Internet Research, 2017). ↗

16. World Health Organization. Ethics and governance of artificial intelligence for health (2021). ↗

17. NHS England. Planning and implementing real-world AI evaluations: lessons from the AI in Health and Care Award (2024). ↗

18. Nuffield Trust. How are GPs using AI? Insights from the front line (2025). ↗

19. Madeleine Clare Elish. Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction, Engaging Science, Technology, and Society 5 (2019), 40–60. ↗