Health systems are entering a new phase of digital change. Artificial intelligence (AI) and generative AI (GenAI) no longer sit at the edge of care as “helpful tools”. They increasingly shape the clinical pathways, workflows and operating systems that determine how care gets delivered, measured and improved.
AI capability is moving fast: from earlier expert systems and machine learning to foundation models and, now, “agentic” systems that can plan tasks, use tools and act towards goals across multiple steps.
If demand keeps rising, but capacity and budgets stay tight, what changes the equation?
The answer is increasingly structural: AI is beginning to shape how patients enter the system, how clinicians decide, and how services coordinate across episodes of care. In healthcare, uptake has accelerated quickly: the American Medical Association reported that 66% of physicians used AI tools in 2024, up from 38% in 2023. By contrast, European Commission data suggest that in Italy AI adoption (including GenAI) is 18% versus a European average of 25%, and 46% of firms neither use nor plan to use AI, compared with 40% across other European countries.
Nonetheless, healthcare has already seen progress in discrete tasks such as image interpretation, appointment handling and documentation support. In the UK, ML lung-cancer detection tools such as Behold.ai can speed diagnosis. Administratively, Microsoft’s ambient listening tool, DAX Copilot, can transcribe clinician–patient conversations directly into the health record.
Yet adoption remains slow. Many organisations still treat AI as a collection of pilots rather than an industrial capability, deploying AI in narrow pockets: a model that flags clinical risk, a chatbot that answers simple questions, or automation in back-office processes.
These initiatives can create local value, but they rarely change the system’s binding constraints; newer GenAI and agentic approaches make broader redesign possible because they can interpret unstructured information, coordinate actions across systems and learn from feedback.
The next shift is structural. AI begins to shape how patients enter the system, how clinicians make decisions, and how services coordinate across episodes of care. Value comes less from individual tools and more from redesigning end-to-end pathways and the operating model that sustains them.
From pilots to pathways: three use cases that rewire care delivery
Early wins come from embedding AI (predictive, generative and agentic) in end-to-end use cases that span quality, experience and productivity. The article from BCG Managing Director and Partner Maria Lopez, “How AI Is Helping to Heal Patients and Hospitals” reports three “flagship” use cases of a European private hospital group with 200+ hospitals and care centres, which have improved care, expanded capacity and reduced costs by implementing AI and ML. The strategy centred on scaling access to unstructured data and translating it into standardised, clinically validated, AI-enabled workflows to support more personalised care.
These examples also show a wider lesson: strong results appear when organisations redesign workflows around AI rather than inserting AI into unchanged processes.
1) Redesigning clinical pathways
Care delivery often fragments into steps owned by separate teams and systems. AI enables pathway intelligence that follows the patient across the episode, highlights deviations early, and standardises decisions where evidence is strong, reducing unwarranted variation. In practice, this means shifting attention from “which use case looks exciting” to “which pathway should be redesigned first”.
In the European hospital group example, an AI support tool read and interpreted electronic health record information and proposed pathway assignments, helping address very low baseline pathway adoption. The programme focused on the end-to-end pathway: identifying patients, selecting a pathway, setting the starting point and adjusting the clinical plan as the patient progresses. By automating identification, recommendations and feedback loops, adoption increased, and capacity was unlocked while keeping clinicians in control.
A similar logic applies in imaging. Providers face rising demand and limited radiologist capacity, plus long waits and variability in reporting. An AI-enabled radiology workflow combines algorithms (for example triage and segmentation) with automated preliminary report generation and smart assignment of studies using rules and AI. Early results indicate a 20–40% reduction in reporting time and improved quality and consistency of diagnostics. Here, the “use case” is a redesigned pathway from prescription to reporting that reduces delay and variability.
Pathway redesign matters because it makes preventive action feasible at scale and builds in an explicit “human checkpoint”, so clinical teams can confirm, amend or reject suggestions. This creates a learning loop: pathway owners can track deviations, relate them to outcomes, and update protocols accordingly, supporting human–machine teaming and a clinical-grade validation approach.
2) Human–machine teaming and clinical-grade validation
Healthcare does not need “autonomous clinicians”. It needs teams that combine machine speed with clinical judgement. Human–machine teaming means designing workflows where AI drafts, flags or prioritises, and clinicians approve, correct and learn, with performance measured against outcomes and safety, not only task speed.
The pattern appears in everyday applications. Ambient documentation can draft notes and discharge summaries, but clinicians must confirm accuracy and relevance. Imaging tools can triage studies and generate preliminary report text, but radiologists validate findings. Patient communication agents can handle routine scheduling and FAQs, but staff oversee escalation for complex or sensitive needs.
“The average human-AI system performs better than a human alone”
In the European provider example, a virtual agent programme reduced average handling time by about 20% while maintaining service levels, illustrating how automation can complement, not replace, frontline teams. In practice, AI supports three functions: 1) sense: interpret signals (notes, imaging, vitals, laboratory results, call transcripts) and detect risk or deviation; 2) Decide: recommend actions aligned to evidence, pathways and local constraints; 3) Act: trigger tasks (schedule, route, draft, and escalate) while keeping humans in control of clinical decisions.
Human–machine teaming works best when AI handles pattern recognition, information retrieval and draft outputs, while clinicians retain accountability for decisions that affect diagnosis and treatment. Operationally, this design assumes supervision, escalation and handover, rather than unattended autonomy.
Therefore, the aim is not to replace clinicians. The aim is to industrialise human–machine teaming and clinical-grade validation, where AI handles repetitive interpretation and coordination, and clinicians focus on judgement, empathy and complex trade-offs.
3) Industrialisation and enterprise orchestration
The hardest part of AI is not the model; it is scaling reliably across sites, specialties and technologies. Industrialisation means building an enterprise capability that turns pilots into products, and products into an operating model change.
In the European hospital group example, the programme moved from proof of concept to a minimum viable product (MVP) and scale-up. Across three flagship initiatives - clinical pathway support, imaging workflow optimisation and a patient service virtual agent - the portfolio targeted improved care, expanded capacity and reduced costs.
Orchestration also avoids “AI sprawl”. Domain-based roadmaps, shared data foundations and a governance process for model approval allow reuse across functions, while keeping clinical ownership clear. A BCG framework often summarised as “10/20/70” reflects this imbalance, suggesting that value depends far more on people, process and change than on the technology layer alone.
What AI changes in healthcare delivery: a blueprint for executives and clinical leaders
Technology alone delivers limited value. Organisations that treat AI as an IT side project often scale pilots but fail to change outcomes. It is the redesign of care that substantially shifts the equation.
From reactive to preventive/proactive
Generative AI can support a shift from reactive episodes to preventive care by detecting risk earlier, standardising follow-up and reducing avoidable delays. Predictive algorithms already flag deterioration and prioritise high-risk patients; generative AI adds a conversational layer that can summarise complex records, draft care plans and translate clinical intent into coordinated actions across teams.
From episodic to continuous
Digital-first pathways supported by AI can maintain continuity between visits, using structured follow-ups, symptom tracking, and targeted escalation rules. Remote monitoring, virtual wards and command functions can extend clinical oversight beyond hospital walls and detect deterioration earlier.
From variable to standardised
AI-supported standardisation can reduce unwarranted variation in documentation, triage, and pathway adherence. That reliability matters in quality improvement, safety, and financial sustainability.
From capacity-constrained to capacity-optimised
Operational AI can improve theatre scheduling, bed management, discharge planning, and staffing deployment, releasing time for direct care.
From manual to automated
AI can reduce time spent on documentation, scheduling, reporting and other tasks that consume staff time without improving outcomes.
Five operating-model shifts matter most:
1) Prioritise transformations, not dozens of pilots: Leaders should focus on a small set of end-to-end opportunities across a few functions, rather than dispersing efforts across 100+ use cases
2) Build a data and AI platform with guardrails: A GenAI layer needs orchestration, model gateways and safety controls, underpinned by data governance and secure integration
3) Establish an AI delivery office: Central teams can set targets, track value, coordinate releases, and manage risk, while functions own workflow change
4) Invest in skills and new roles: Programmes need clinicians who can define pathways and validate outputs, product owners who translate needs, and teams trained in AI literacy and supervision
5) Embed change management: Adoption requires redesigning roles, updating standard operating procedures and supporting staff through learning-by-doing rather than one-off training.
What can go wrong and how to prevent it
AI in healthcare raises the bar on safety, privacy and accountability. Regulation and ethics add further constraints, requiring clear accountability, privacy controls and documented validation.
A responsible AI (RAI) programme should define risk taxonomy and appetite, assign roles and committee structures, set “no-fly zones” for unacceptable uses, and monitor outputs with key risk indicators.
Leaders can move faster without compromising safety by following four steps:
1) Define outcomes and value pools: Tie AI to capacity, access, quality and staff experience
2) Select a small portfolio of pathways: Imaging and access use cases that connect end-to-end and standardise toolkits and playbooks to support smooth integration
3) Build foundations early: Platform, data governance, safety controls and a delivery office
4) Scale with evidence: By creating “deployment networks”. Run pilots with clinical-grade validation, standardise what works, then expand while monitoring performance.
AI and generative AI are becoming the backbone of future care models
So, what changes the equation? The opportunity extends beyond isolated tools: it lies in redesigned clinical pathways, validated decision support and enterprise orchestration that scales safely. Health systems that industrialise AI with clinical-grade validation and responsible governance can unlock capacity, improve sustainability and deliver patient-centred and personalised performance at scale.
The most advanced vision is a learning care system where data from routine care improves pathways continuously. In this model, AI supports decision-making, operations, and patient engagement while governance ensures safety, equity, and accountability.
References
[1] “AI@Providers: 3 BCG flagship AI use cases at European private hospital group” – BCG, (October 2025).
[2] “GenAI revolution and agentic AI: enterprise foundations and responsible AI.” – BCG, (2024-2025).
[3] Impact Advisors (2025). The Generative AI Revolution: 9 Trends Reshaping Healthcare in 2025 (investment trends and outlook).
[4] Scottsdale Institute / JAMIA (Fall 2024 survey; published 2025). Adoption of artificial intelligence in healthcare: survey of health systems (ambient documentation and imaging deployment).
[5] “Health system digital front end: Reimagining early patient pathways.” - BCG, February 2025.
[6] “When humans and AI work best together — and when each is better alone” - MIT Management Sloan School, February 2025.
[7] Laying the Tech Foundation for GenAI Success – BCG, December 2024
[8] Unlocking AI’s potential in the NHS: Moving from pilots to scale – HSJ for healthcare leaders, February 2025
[9] AI e leadership femminile in sanità opportunità, sfide e crescita – BCG, July 2025
[10] How AI Is Helping to Heal Patients—and Hospitals – BCG, January 2026
