9 min readInterim Leadership · Advisory & Consulting27 August 2026

Healthcare Leadership in AI Transformation: Why Leadership Becomes the Limiting Factor

AI has arrived in healthcare. The decisive bottleneck now lies in leadership, process design and governance. How leadership teams turn AI into measurable process and outcome improvement while actively preventing skill erosion.

AI technology has arrived in healthcare. The decisive bottleneck now lies in leadership, process design and governance. This Impulse shows how leadership teams turn AI into measurable process and outcome improvement while actively preventing skill erosion.

Executive Summary

Thesis	Key statement
Technology and organisation	The technology question is largely settled. The organisational question determines value creation. 50% of surveyed US healthcare organisations use generative AI in production, while 81% of physicians use AI professionally.
Leadership gaps	Only 33% of frontline employees consider leadership communication about AI clear. 36% feel adequately qualified. 66% receive little or no direction on how to use time saved.
Process redesign	The value step-change comes from redesigning whole value streams. 42% of organisations already lead AI as collective process change.
Skill erosion	A multicentre observational study points to a potential deskilling risk. Following introduction of an AI detection system, adenoma detection in procedures without AI fell from 28.4% to 22.4%.
Management and product	Change management becomes a product and management requirement. The Digital Omnibus changes timelines and clarifies obligations, while responsibility for safe use and training remains.

1. Starting Point: Technology Has Arrived, Organisation Has Not

The adoption debate in healthcare is largely over. McKinsey's survey of 150 US healthcare organisations reports that 50% use generative AI in production. More than 80% have rolled out initial use cases to end users. Eighty-two per cent expect a positive return on investment and 45% can already quantify it.[1]

The user perspective tells a similar story. The AMA survey of around 1,700 physicians reports professional AI use of 81%. Seventy-six per cent see benefits for care quality and 70% expect relief in workload-driving activities.[2]

Use alone does not create value. BCG's fourth survey, covering 11,749 respondents in 14 markets, shows that only 42% of organisations treat AI as collective change across end-to-end processes. Most still use distributed individual tools without shared process logic. This is where leadership determines whether AI delivers outcomes or disappointment.[3]

2. The Real Bottleneck Is Leadership, Not Technology

The BCG data make the practical gap visible: clarity, direction and accountability. Models, computing capacity and licences do not replace this leadership task.[3]

Leadership gap	Finding	Organisational consequence
Strategic clarity	28% see alignment between their stated AI strategy and everyday practice.	Parallel initiatives, tool sprawl and no prioritisation.
Communication	33% of frontline employees consider leadership communication about AI clear.	Rumours, uncertainty and defensive use.
Qualification	72% experience changed skill requirements and 36% feel adequately qualified.	The capability gap grows faster than the training offer.
Time dividend	66% receive little or no direction on how to use time saved.	Efficiency gains dissipate and benefits remain invisible.
Governance	50% report missing rules for human-AI collaboration.	Responsibility, liability and control remain unclear.

The central insight is clear: employees with a clear strategy and limited access to tools achieve better outcomes than employees with broad tool access but no direction. The next licence package does not replace the decision about which process should work differently in the future.[3]

Deloitte's survey of 100 US technology leaders from health systems and payers also describes a widening gap. Sixty-one per cent are building agentic AI or have secured budgets. Eighty-five per cent plan to raise investment over the next two to three years. Among early adopters, 59% expect cost effects above 20%; among waiting organisations, only 13% do.[5]

3. What This Means for Workflows: Three Maturity Levels

The most effective leadership step is an honest assessment of maturity. The following framework serves as a diagnostic grid.

Maturity level	Characteristics	Typical metric	Core leadership task
1. Tool level	AI tools are provided. People use them individually. Processes remain unchanged.	Usage rate, licence utilisation	Permit, safeguard and observe.
2. Workflow redesign	Individual process chains are redesigned. Steps are removed and roles shift.	Cycle time, error rate, staffing per case	Redefine processes and formally retire old steps.
3. Operating-model redesign	Service delivery, role model and steering are reconsidered. Agentic systems take over process chains.	Cost per episode of care, capacity, outcome quality	Shape target operating model, decision rights, governance and people strategy.

The move from level 1 to level 2 is most often skipped. It requires active deprecation management. During validation, running an AI-supported and a manual route in parallel is appropriate. It creates additional load when no date, criterion and accountable role exist for retiring the old process.

The time dividend also needs leadership. AI releases time in many small units. Without a clear decision, existing work absorbs it and its impact remains unmeasurable. Leadership teams should decide upfront whether the time goes into additional volume, quality work, training or relief, and make that choice measurable.[3]

4. The Underestimated Risk: Skill Erosion

The key question is not only whether AI improves performance. It is also how regular AI use affects the ability to carry out critical tasks independently. A multicentre observational study in four Polish centres examined 19 experienced endoscopists, each with more than 2,000 previous colonoscopies. Across 1,443 non-AI colonoscopies, the adenoma detection rate fell from 28.4% to 22.4% after routine introduction of an AI detection system. The absolute difference was minus 6.0 percentage points.[4]

The study is observational, limited to four centres in one country, and does not establish causality. It still indicates a relevant direction: regular AI use may affect professionals' ability to perform patient-relevant tasks without support. Eighty-eight per cent of physicians surveyed by the AMA report concerns about skill loss. Seventy per cent are concerned about training gaps among medical students and residents.[2]

Skill preservation therefore becomes an actively managed metric. Organisations should identify critical manual capabilities. AI-free training and assessment intervals, calibration of human judgement, and measures against automation bias belong in implementation. These safeguards ensure that efficiency gains contribute sustainably to care quality.

5. Change Management in Practice: Four Levers

Lever	Concrete action	Metric	Typical mistake
Participation and decision rights	Formally involve clinical staff in selection, piloting and acceptance. Define veto rights.	Share of use cases with documented professional involvement	Treating AI selection solely as an IT or procurement decision.
Deprecation management	Set a retirement date, retirement criterion and accountable role for the old process for every use case.	Number of retired parallel processes per quarter	Indefinite parallel operation without accountability.
Tiered qualification	Build AI baseline competence for all, application competence for users and accountability competence for approvers.	Coverage by level, evidence of completion	One-off mandatory training without role relevance.
Governance and liability	Define a responsibility matrix for human-AI decisions, escalation pathways and deviation documentation.	Share of AI-supported decisions with documented human approval	Unclear liability when a system recommendation is wrong.

The first lever deserves particular attention. Eighty-five per cent of surveyed physicians want to be consulted about decisions on AI use. The AMA states a core condition: AI should support, not replace, physician work.[2] Deloitte's survey also suggests a favourable window: 38% of leaders report less resistance to change and 35% report stronger leadership support.[5]

The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It changes timelines and selected implementation obligations of the AI Act. For organisations using high-risk AI in clinical settings, qualification remains central to safe use and defensibility in the event of liability claims.[6]

6. What This Means for Product Development

For manufacturers of AI-enabled software, particularly Software as a Medical Device, three requirement levels emerge.

Level	Previous practice	Expectation from 2026 onward
Use safety	Usability engineering under IEC 62366-1 focuses on use errors.	Automation bias and over-reliance become distinct use-related risks in risk management. Explainability and uncertainty indication become design requirements.
Market surveillance	PMS and PMCF focus mainly on model performance and incidents.	Real-world use patterns, acceptance rates of system recommendations and indications of user skill drift are systematically monitored.
Implementability	The product ends at the technical interface.	Reference workflow, role model, training curriculum and retirement plan for old processes become part of the deliverable.

Implementability is commercially decisive. When organisations do not lead AI as process change, the success of a technically compelling product depends on the customer's ability to adopt it. Manufacturers who provide implementation methodology, a measurement framework and qualification material connect software to a traceable outcome improvement.[3] [5]

The revised timelines create planning room. They do not replace early preparation. For stand-alone high-risk systems under Annex III, 2 December 2027 applies. For AI as a safety component in regulated products under Annex I, including applications in the MDR and IVDR context, 2 August 2028 applies. Transparency duties under Article 50 of the AI Act apply from 2 August 2026. For systems already placed on the market before that date, the Digital Omnibus provides a transition period until 2 December 2026.[6]

7. A 90-Day Agenda for Leadership Teams

Priority	Task	Concrete result after 90 days
1	Assess maturity	All ongoing AI use cases are assigned to one of the three maturity levels.
2	Define the time dividend	Use of saved time is decided and communicated for the three largest use cases.
3	Create a retirement plan	Every parallel operation has a date, criterion and accountable role.
4	Map capability risks	Critical manual capabilities and AI-free training and assessment intervals are defined.
5	Approve a responsibility matrix	Approval, documentation and liability are defined for each class of AI-supported decision.
6	Establish qualification	Baseline, application and accountability competence are established with evidence of completion.
7	Change the metrics	Outcome measures such as cycle time, error rate, capacity and outcome quality complement usage rates.

Conclusion

AI transformation in healthcare is a leadership task. The tools are available and users are already adopting them. Sustainable value emerges when leadership teams provide direction, redesign workflows, retire old processes consistently and actively preserve capabilities.

Strategic clarity creates more impact than a broad tool offering. Process redesign achieves more than individual productivity. Managed skill preservation ensures that efficiency gains strengthen care quality. Manufacturers gain further advantage when they provide the organisational change alongside the product and make implementability a true product attribute.

Sources: [1] McKinsey & Company: Generative AI in healthcare: Current trends and future outlook, 16 April 2026. mckinsey.com · [2] American Medical Association: 2026 Physician Survey on Augmented Intelligence, March 2026. ama-assn.org · [3] Boston Consulting Group: AI at Work: Why Strategy Matters More Than Tools, 2 June 2026. bcg.com · [4] Budzyń K. et al.: Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy. The Lancet Gastroenterology & Hepatology 2025; 10(10): 896 to 903. DOI 10.1016/S2468-1253(25)00133-5 · [5] Deloitte: Many health care leaders are leaning into agentic AI as adoption hurdles ease, 11 February 2026. deloitte.com · [6] Regulation (EU) 2026/1744, Digital Omnibus on AI, EUR-Lex.

Dr. Patrik Scholler

Consultant for Digital Health, Life Sciences and Managed Delivery

More about me