Agentic AI for Ageing Healthcare Systems in Advanced Economies: A Structured Review of Evidence, Institutional Barriers, and a Sociotechnical Implementation Roadmap

1. Introduction

1.1. The Structural Demographic Challenge

By 2050, one in three citizens in Japan will be over 65, with Italy and South Korea projected to approach 30–40% over the same period.[1] Sustained sub-replacement fertility, including South Korea’s 0.75 total fertility rate (2025 estimate; 0.72 in 2023)[4], combined with rising longevity exceeding 84 years in Japan and Singapore[5] are driving this transition. Its speed is as consequential as its scale, with demographic ageing that took over a century in France occurred within decades in Japan and South Korea.[6,7]

Three structural pressures converge. First, workforce contraction: Japan faces an estimated 250,000 nursing shortfall by 2026[8] and Italy's long-term care workforce density remains well below the OECD average.[2,9] Second, fiscal strain: per-capita healthcare spending for older adults is 3 – 5 times that of working-age cohorts,[2] and ageing alone is projected to add 1.8 – 2.7 percentage points of GDP to public health expenditure by 2060.[9] Third, care complexity: multimorbidity and dependency increase coordination demands across already fragmented systems. Ageing also carries human costs obscured by aggregate statistics, including social isolation, solitary deaths, and intensified caregiving burdens that further constrain labour supply. [10]

Across the seven most demographically stressed advanced economies, healthcare spending ranges from 6.1% to 16.9% of GDP, and total ageing-related expenditure absorbs between 10% and 28% of GDP in the most affected economies.[2,9] The IMF's 2024 Fiscal Monitor projects public health spending in advanced economies will rise by 2.0–3.5 percentage points of GDP by 2050 under baseline scenarios, and by 3.5–6.0 points under adverse trajectories.[46] Country-level projections are stark: Japan's healthcare expenditure is forecast to rise 55% above its 2023 baseline by 2050; South Korea's is projected to double; Singapore's to nearly triple from S$22 billion (2018) to S$59 billion by 2030.[49] Demographic ageing is projected to reduce GDP growth by 0.3–1.2 percentage points per year by 2035, with Germany and Italy potentially recording growth below 1%,[9,46] while the worker-to-retiree ratio falls from approximately four contributors per retiree today to fewer than two by 2030–2040.[46,48] These projections motivate this review's central question: whether Agentic AI can contribute meaningfully to moderating expenditure trajectories that no conventional policy instrument has yet resolved.

The projected range of 0.3–1.2 percentage points annual GDP reduction by 2035 reflects country-specific estimates and should not be interpreted as a single aggregate trajectory. Germany and Italy sit at the higher end wherein the United States at the lower.[9,46]

1.2. The Insufficiency of Conventional Responses

Governments have responded across four domains, including pension and labour-market reform, healthcare restructuring, workforce expansion, and digital health integration yet each facing binding constraints. Pension reform provides limited fiscal relief and often provokes political resistance, as recent protests in France illustrate.[11] Preventive and community-based healthcare restructuring depends on the very workforce already under strain. Japan's long-term care insurance system and Singapore's Healthier SG programme both require sustained labour supply to scale.[8,16] Workforce expansion through training and recruitment faces long lead times and retention challenges.[2] Digital health strategies, including Germany's DiGA programme, remain unevenly adopted and lack validated system-level impact.[16] Consequently, even high-performing nations face pressures that existing policy instruments cannot fully resolve.[12]

1.3. Agentic AI Through a Sociotechnical Lens

This paper distinguishes three levels of healthcare AI. Narrow AI refers to single-task prediction or classification tools (e.g., diagnostic imaging models). Generative AI refers to large language models capable of natural language generation and reasoning across domains. Agentic AI (AAI) refers to systems that execute sustained, multi-step, goal-directed workflows across institutional boundaries, combining planning, memory, environmental interaction, and feedback under structured human oversight. The boundary between generative and agentic AI is functional rather than architecturally fixed and becomes increasingly fluid as persistent context and tool-use frameworks mature.

AAI offers capabilities well aligned with the pressures facing ageing health systems. Unlike narrow clinical AI limited to prediction tasks, AAI integrates planning, memory, environmental interaction, and feedback to execute sustained workflows across institutional boundaries.[13] For example, an AAI system could detect early clinical deterioration from home-monitoring data, initiate a telehealth consultation, and coordinate follow-up care under defined human oversight.

This review adopts sociotechnical systems theory, including the NASSS framework[35], as its analytical anchor. Technological success depends not only on technical performance but also on alignment with organisational structures, governance arrangements, financing models, and human roles. Applied to eldercare, this perspective generates a testable proposition that deployment outcomes are determined primarily by institutional conditions rather than algorithmic capability.

The analysis addresses three research questions:

−

RQ1 (Capability): What is the current and near-term evidential basis for AAI across remote monitoring, assistive robotics, diagnostic support, medication adherence, and workflow automation, and how does evidence vary with intervention complexity?

−

RQ2 (Implementation): Which regulatory, financial, organisational, and infrastructural factors determine whether AI capability translates into population-level benefit, and how do these vary across national contexts?

−

RQ3 (Equity and Sustainability): Under what conditions can AI be adopted to avoid widening inequities and contribute to fiscal moderation, and which governance and evidence gaps prevent validation at scale?

The sociotechnical lens positions institutional alignment rather than technical optimism as the organising principle for the roadmap that follows.

This review contributes to the emerging literature on AI and ageing societies by integrating demographic economics, implementation science, and agentic AI architectures into a unified sociotechnical framework for health system transformation.

2. Methodology

This study employed a structured narrative review (SNR) to synthesise heterogeneous evidence across demography, health economics, artificial intelligence, and policy. Meta-analysis was not feasible due to disciplinary diversity, outcome heterogeneity, and the inclusion of grey literature. Search and screening procedures followed PRISMA 2020 guidance. A structured review protocol was developed prior to data extraction but was not formally registered with a prospective registry; the review should therefore be interpreted as PRISMA-informed rather than fully PRISMA-compliant.

2.1. Search Strategy and Source Selection

Searches were conducted in PubMed, Scopus, Web of Science, and IEEE Xplore, supplemented by arXiv, medRxiv, and grey literature from the OECD, WHO, United Nations, European Commission, and national health ministries. Publications from January 2020 to December 2025 were prioritised, and earlier landmark studies were included where sources identified as foundational are marked with a [Foundational] notation. Search terms covered four domains, including population ageing, healthcare system pressure, artificial intelligence and agentic systems, and geographic scope. Representative search terms are provided in Appendix A.

2.2. Screening and Selection Process

The search yielded 1,812 records, of which 440 duplicates were removed. Two reviewers independently screened 1,372 titles and abstracts, achieving strong inter-rater agreement (Cohen's κ = 0.81). Of the 355 full-text articles assessed, 56 met the inclusion criteria. An additional 25 sources identified through citation tracking resulted in a final sample of 81 included sources. The PRISMA flow diagram is presented in Figure 1.

Inclusion required relevance to population ageing or artificial intelligence in healthcare within OECD contexts, together with empirical evidence or methodologically transparent institutional data. Opinion pieces, non-healthcare AI studies, and commercial market research were excluded. Disagreements between reviewers were resolved through consensus.

2.3. Quality Appraisal, Data Extraction and Synthesis

Empirical studies were appraised using an adapted CASP framework assessing methodological rigour, relevance, outcome validity, follow-up adequacy, and reporting transparency. Systematic reviews were evaluated using AMSTAR-2 criteria, and institutional grey literature was assessed for data transparency and methodological clarity. Narrative synthesis followed a framework synthesis approach anchored to the NASSS framework,[35] with thematic mapping of evidence against the six NASSS domains (condition, technology, value proposition, adopter system, organisation, and wider system). This approach was selected over thematic or realist synthesis because the NASSS framework provides an established theoretical architecture directly applicable to complex health technology adoption — the central analytical task of this review.

Evidence strength was assigned at the domain level using pre-specified operational criteria applied independently by both reviewers with consensus resolution. The full classification framework is presented in Appendix Table B2.

3. Evidence Synthesis: Structural Pressures, AI Applications, and Institutional Constraint

3.1. Structural Pressures on Ageing Health Systems

Population ageing generates healthcare strain through three compounding mechanisms, including rising age-related expenditure, workforce contraction, and increasing care complexity. Per-capita costs for those aged 65 and above are several times higher than for working-age populations and rise sharply beyond age 80. OECD and European Commission projections indicate that ageing alone will add several percentage points of GDP to public health expenditure across member states.[9,15] Growth is concentrated among those aged 75 and above, the highest utilisation subgroup, and old-age dependency ratios in the most affected nations are projected to approach one older adult per working-age individual later this century.[1] Multimorbidity further compounds strain because most adults over 65 live with two or more chronic conditions requiring cross-specialty coordination.[10] Table 1 summarises these pressures. Note: France appears throughout this review as a comparative reference economy and in the cross-country data presented in Appendix E. It is not included as a primary case in the Section 3 synthesis tables, as its demographic and fiscal profile places it in an intermediate category relative to the seven economies where the evidence base is most developed.

Two cross-cutting patterns shape the remainder of the analysis. Evidence strength is inversely related to intervention complexity. The multi-domain applications most needed in ageing health systems therefore have the weakest evidence base, reflecting underinvestment in complex interventions. Implementation of context also matters more than technical performance. Factors such as workflow integration, clinician training, institutional readiness, and digital literacy influence outcomes more than algorithm accuracy.[25,35] Table 2 illustrates this inverse relationship across AI domains. Detailed cross-country demographic and fiscal projections for major ageing economies are provided in Appendix E (Tables E1–E2) to reduce table density in the main manuscript.

Three patterns from the cross-country demographic and fiscal projections in Appendix E (Tables E1–E2) are particularly important. South Korea and Singapore face the most severe projected dependency ratios by 2050, at 79.0 and 85.0 respectively, but their fiscal preparedness differs sharply. Singapore’s CPF compulsory savings model keeps ageing-related spending near 10 percent of GDP, compared with about 16 percent in South Korea. Italy and France already spend about 28–30 percent of GDP on ageing-related programmes, leaving limited fiscal space for new technology investments. The United States faces lower structural pressure because of a relatively younger population and immigration-supported workforce, despite having the highest per-capita health spending. These differences shape AI deployment sequencing. Countries with stronger fiscal capacity and high digital readiness, including Singapore, Germany, and South Korea, are better positioned for early Phase 2 pilots, while countries with higher fiscal risk, including Italy and France, may require fiscal consolidation before large-scale deployment.

Appendix E presents detailed demographic and fiscal projections across ageing economies. Table 3 summarises the relationship between GDP growth and healthcare expenditure pressures, and highlights where AI-enabled efficiency gains may help moderate fiscal strain.

3.2. AI Applications in Ageing Health Systems

Across five principal domains, a consistent pattern emerges. AI tools are most effective when used to support structured care pathways rather than replace human judgement.[42,43] Evidence for foundational applications such as remote monitoring and socially assistive robotics is largely based on trials conducted before 2018. Later studies, including systematic reviews published between 2020 and 2024, broadly replicate the directional findings of earlier work, though they do not extend follow-up duration or population scale sufficiently to resolve questions about long-term cost-effectiveness.[19,20,21,22,42,43] This pattern suggests continued underinvestment in complex intervention research and highlights the research gaps discussed in Section 4.6. Key methodological limitations and qualifications associated with widely cited studies are summarised in Table 4.

Remote monitoring systems are associated with reduced hospital admissions and emergency utilisation among chronic disease populations.[19,20] However, results depend on context and evidence for mortality reduction or delayed institutionalisation remains limited. Socially assistive robots improve agitation, mood, and engagement in dementia care.[21,22,40] Their use also raises governance questions about augmentation versus substitution of human care. Diagnostic AI achieves specialist-level accuracy in imaging tasks.[17,18,44] However, it does not generalise well to complex geriatric clinical decision-making. Medication adherence tools show modest and mixed improvements.[14] Ambient documentation systems show early evidence of reduced clinician administrative burden, although studies remain short term and observational.[41]

3.3. Economic Evidence: Microeconomic Gains and Macroeconomic Uncertainty

Economic evidence is limited by the absence of validated macroeconomic evaluation. The central question is whether AI can meaningfully moderate demographic expenditure growth at population scale. This question remains unanswered in the peer-reviewed literature. The IMF Fiscal Monitor (2024) projects that public health spending will rise by 2.0–3.5 percentage points of GDP by 2050 under baseline scenarios and by 3.5–6.0 points under adverse trajectories.[46] Claims that AI could moderate these trends should therefore be treated as hypotheses until longitudinal evidence emerges.

At the microeconomic level, studies report incremental savings from reduced admissions through remote monitoring and care coordination.[19,20] Some studies report long-term care gains from delayed institutionalisation.[31] Emerging evidence also suggests productivity gains from documentation automation.[41] Cost-effectiveness findings differ across financing models. In Japan’s publicly funded system, savings from avoided institutionalisation align incentives. In Germany’s Bismarckian model, reimbursement reform across statutory funds is required. In Singapore’s Medisave-based system, central governance may accelerate implementation, but equity risks remain. The economic case for AI in eldercare is therefore system specific and cannot be derived from aggregate evidence without considering financing models, incentives, and population baseline.

Table 5. Economic Evidence on AI in Ageing Health Systems.

Economic Layer	Mechanism	Evidence Strength	Key Limitation	Implication
Healthcare spending	Reduced admissions through remote monitoring and coordination	Moderate evidence from pilot and controlled studies [19,20]	Effects vary by context and are not validated nationally	Promising but requires system-level evaluation
Long-term care costs	Delayed institutionalisation through ageing-in-place technologies	Moderate evidence with sustained deployment [31]	Demographic growth may offset efficiency gains	Scale required for fiscal impact
Workforce productivity	Automation of documentation and coordination tasks [41]	Emerging observational evidence	Implementation costs and workflow risks unclear	Near-term productivity gains possible
Macro-fiscal impact	AI-augmented care workforce	No validated population-level evidence	Current estimates rely on extrapolation	IMF projects health spending increase of 2.0–6.0% GDP by 2050 [46]

Table 5. Economic Evidence on AI in Ageing Health Systems.

Economic Layer	Mechanism	Evidence Strength	Key Limitation	Implication
Healthcare spending	Reduced admissions through remote monitoring and coordination	Moderate evidence from pilot and controlled studies [19,20]	Effects vary by context and are not validated nationally	Promising but requires system-level evaluation
Long-term care costs	Delayed institutionalisation through ageing-in-place technologies	Moderate evidence with sustained deployment [31]	Demographic growth may offset efficiency gains	Scale required for fiscal impact
Workforce productivity	Automation of documentation and coordination tasks [41]	Emerging observational evidence	Implementation costs and workflow risks unclear	Near-term productivity gains possible
Macro-fiscal impact	AI-augmented care workforce	No validated population-level evidence	Current estimates rely on extrapolation	IMF projects health spending increase of 2.0–6.0% GDP by 2050 [46]

3.4. Governance, Regulation, and the Algorithmic Bias Problem

Under the EU AI Act (2024), AI systems that qualify as medical devices, including diagnostic, clinical decision support, or patient management tools, are classified as high risk. This classification requires conformity assessment, human oversight, transparency, data governance, and post market monitoring.[34] Agentic eldercare systems that execute multi step clinical workflows fall within this category, so compliance timelines must be integrated into implementation planning from the beginning. Japan, Singapore, and South Korea face similar governance challenges but lack the EU’s coordinated regulatory structure. The OECD has also identified responsible AI deployment in healthcare as a policy priority.[33] Large language models introduce additional risks because they can generate clinically plausible but incorrect outputs. This reinforces the need for human oversight at critical clinical decision points.[23]

Algorithmic bias represents another institutional constraint. Obermeyer et al. demonstrated systematic racial bias in a US risk scoring algorithm.[24] Subsequent studies in Nordic and UK systems identify similar risks linked to underrepresentation of rural, minority, and oldest old populations.[30] Bias patterns are context specific. East Asian datasets may appear ethnically homogeneous but can still underrepresent socioeconomically marginalised elderly groups. Auditing must therefore rely on locally representative data. In geriatric care this issue is particularly important because underrepresented populations often have the highest care needs and the least alternative access to services.

3.5. Equity as a Structural Barrier, Not an Afterthought

Equity implications extend beyond algorithmic bias to structural determinants shaping whether deployment narrows or widens disparities. Three inequality axes are central. Urban rural divides show rural older adults face weaker digital infrastructure and higher illiteracy, with remote monitoring uptake consistently lower in rural settings[14,32] meaning those most isolated from conventional care are least likely to benefit from AI alternatives. Digital literacy gradients show telehealth utilisation among adults aged 75 and above remains substantially lower than among younger cohorts across OECD nations[14,30] risking exclusion of the oldest old with the highest care needs. Socioeconomic stratification shows out of pocket costs for assistive technologies concentrate quality gains among already advantaged populations[24,30]. Equity must shape deployment sequencing prioritising underserved groups in pilot design rather than treating them as post deployment audit categories.

3.6. Comparative Institutional Readiness

Institutional readiness patterns reinforce the sociotechnical thesis that adoption trajectories are shaped as strongly by governance design infrastructure maturity and public trust as by technical capability. Japan combines acute demographic urgency with sustained robotics investment but faces workforce shortages and fragmented EHR systems[8]. South Korea benefits from advanced digital infrastructure yet confronts the fastest demographic transition globally which compresses the implementation window[4,39]. Singapore's centralised governance enables rapid digital health coordination though its small population limits evidence generalisation[16,38]. Germany and Nordic states lead EU digital health adoption, and the EU AI Act uniform high-risk classification extends regulatory timelines across member states[34].

A consistent pattern emerges because demographic urgency often diverges from institutional readiness. Italy and Southern Europe face sharp ageing pressures alongside constrained fiscal and digital capacity. Japan technological leadership coexists with workforce and interoperability constraints. These governance mismatches rather than technical limitations define the central challenge for the roadmap that follows.

Case Study Singapore as a Governance First Benchmark for Ageing System Transformation Singapore provides one of the clearest examples of governance aligned preparation for population ageing. The country faces one of the steepest projected dependency ratios increases globally with the old age dependency ratio expected to rise from 24.2 in 2023 to approximately 85 by 2050. At the same time national healthcare expenditure is projected to reach about S$59 billion by 2030. In response the government has implemented a coordinated national strategy that integrates prevention financing workforce policy digital infrastructure and cross sector governance.[49,54] Preventive System Transformation – The Healthier SG The Healthier SG initiative launched in 2022 restructures primary care around longitudinal preventive management. All residents aged 40 and above are enrolled with a designated primary care physician and receive personalised preventive plans supported by digital monitoring tools including the Healthy 365 application. The government has committed about S$1 billion over five years to support this preventive shift which aims to reduce costly hospital utilisation before the demographic peak.[50] Financing Architecture – The ‘3M’ Framework Singapore financing structure combines three complementary mechanisms. Medisave provides mandatory individual health savings contributions of roughly 8 to 10.5 percent of income. MediShield Life provides universal catastrophic insurance coverage. MediFund operates as a public safety net for patients unable to meet residual costs. This layered design distributes responsibility across individuals’ insurance pools and government while containing fiscal risk and limiting moral hazard.[51] Workforce Longevity Policy - Progressive Retirement Reform To mitigate labour shortages associated with ageing the government has progressively extended workforce participation. The statutory retirement age has been raised to 63 with a planned increase to 65 by 2030 while the re-employment age will rise to 70. Labour force participation among adults aged 65 to 69 increased from 31.4 percent in 2013 to 43.3 percent in 2023 which demonstrates that policy incentives and regulatory reform can significantly extend productive ageing.[52] Digital Health Infrastructure Singapore has established national digital health infrastructure through the National Electronic Health Record system which integrates clinical data across the healthcare system. The Digital Health Blueprint supports nationwide interoperability and data governance. AI enabled diagnostic tools for conditions such as diabetic retinopathy and stroke detection are being deployed while robotic assistants have been piloted in community and long-term care environments. These systems collectively represent operational foundations for the Integrated Perception pillar of the AAF 3.0 framework.[53] Whole of Government Governance – APSA 2023 The Action Plan for Successful Ageing coordinates more than 200 initiatives across health housing, employment, transport and community participation. Active Ageing Centers operate nationwide to support community engagement and care coordination. The Seniors Go Digital initiative has trained more than 200000 older adults in digital literacy which directly addresses digital exclusion risks associated with technology-based healthcare delivery.[54] Implications for Comparative Policy Singapore institutional model is not directly transferable because its city state scale centralised governance and fiscal reserves create enabling conditions that larger federated systems do not possess. Nevertheless, the sequencing of reforms provides a clear policy lesson. Preventive care financing digital infrastructure and cross sector governance were established before demographic pressures reached their peak. In contrast many ageing economies attempt to deploy technological solutions after fiscal and workforce pressures have already intensified. From a NASSS perspective Singapore demonstrates favourable conditions across all six domains including technology readiness value proposition adopter acceptance organisational embedding and system level governance. For larger ageing economies including Japan South Korea Germany and Italy the principal barrier to AI enabled healthcare adoption therefore lies less in algorithmic capability and more in institutional alignment.

Table 6. Comparative Institutional Readiness for AI-Enabled Healthcare in Ageing Societies.

Country / Region	Digital Health Maturity	Governance Architecture	Strategic Advantage	Structural Constraint	Policy Implication
Japan	Very high	Moderately centralised national policy with local implementation	Deep investment in eldercare robotics and monitoring	Severe workforce shortages and fragmented EHR interoperability	Scale automation while prioritising national interoperability standards
South Korea	Very high	Highly centralised digital governance	Advanced ICT ecosystem and rapid national deployment capability	Fastest demographic ageing globally compresses policy response window	Accelerate AI assisted care models to offset labour shortages
Singapore	Very high	Highly centralised and coordinated governance	Unified national digital health strategy	Small population limits large scale clinical evidence generation	Position as a global regulatory and innovation testbed
European Union	High but uneven	Multi-level governance across EU and national authorities	Strong regulatory leadership through the EU AI Act	Fragmented interoperability and extended compliance timelines	Harmonise digital health infrastructure and accelerate cross border data sharing
Italy / Southern Europe	Moderate	Decentralised healthcare governance	Demographic urgency creating reform momentum	Fiscal constraints and limited long-term care workforce	Prioritise digital health infrastructure and workforce augmentation through AI

Table 6. Comparative Institutional Readiness for AI-Enabled Healthcare in Ageing Societies.

Country / Region	Digital Health Maturity	Governance Architecture	Strategic Advantage	Structural Constraint	Policy Implication
Japan	Very high	Moderately centralised national policy with local implementation	Deep investment in eldercare robotics and monitoring	Severe workforce shortages and fragmented EHR interoperability	Scale automation while prioritising national interoperability standards
South Korea	Very high	Highly centralised digital governance	Advanced ICT ecosystem and rapid national deployment capability	Fastest demographic ageing globally compresses policy response window	Accelerate AI assisted care models to offset labour shortages
Singapore	Very high	Highly centralised and coordinated governance	Unified national digital health strategy	Small population limits large scale clinical evidence generation	Position as a global regulatory and innovation testbed
European Union	High but uneven	Multi-level governance across EU and national authorities	Strong regulatory leadership through the EU AI Act	Fragmented interoperability and extended compliance timelines	Harmonise digital health infrastructure and accelerate cross border data sharing
Italy / Southern Europe	Moderate	Decentralised healthcare governance	Demographic urgency creating reform momentum	Fiscal constraints and limited long-term care workforce	Prioritise digital health infrastructure and workforce augmentation through AI

4. Discussion and Policy Implications for AI Implementation in Ageing Health Systems

4.1. The Central Institutional Thesis

The synthesis across Sections 3.1 to 3.6 supports a consistent conclusion. Where AI underperforms relative to technical potential, four institutional factors limit impact more reliably than model performance: fragmented data architectures, misaligned incentives, regulatory uncertainty, and exclusionary deployment design. In ageing systems, institutional fragility amplifies technological limitations.[35,43].

The NASSS framework maps directly onto observed constraints including condition complexity linked to multimorbidity technology complexity from cross domain orchestration uncertain value propositions without macroeconomic validation adopter resistance interoperability limits and regulatory and financing structures that mediate translation from effectiveness to system utility[35]. A further tension concerns relational care quality. Evidence from socially assistive robotics shows that measured improvements in agitation or mood do not determine whether care is enhanced or diminished. Outcomes depend on governance decisions regarding augmentation versus substitution[21,22,40]. Technological capability is therefore necessary but insufficient. Population level benefits depend on institutional alignment and capability without governance produces pilots that cannot scale.

4.2. The Agentic AI Framework (AAF 3.0): Derivation and Differentiation

The AAF 3.0 is more precisely described as an agentic AI domain-mapping matrix, an applied implementation tool derived deductively from the structural pressures identified in Section 3.1 and positioned as an extension of the NASSS framework specifically for agentic AI architectures. The designation “3.0” reflects a conceptual progression in healthcare AI capability from narrow clinical AI (primarily diagnostic and single-task predictive systems) to generative AI (multi-domain language reasoning) and finally to agentic AI (sustained multi-step workflow orchestration with environmental feedback). This versioning is original to this paper and should be understood as a descriptive periodisation of evolving AI capability in healthcare contexts rather than a reference to a previously established framework lineage; no AAF 1.0 or 2.0 frameworks exist in prior literature.

AAF 3.0 differs from NASSS in scope and purpose. NASSS is a general-purpose implementation framework applicable to any complex health technology. AAF 3.0 uses the NASSS analytical structure as its foundation but organises it around the specific functional capabilities of agentic AI systems — each derived from a distinct structural pressure. Where NASSS asks 'which implementation factors matter?', AAF 3.0 asks 'which agentic capabilities address which system pressures, at what evidence stage, and under what governance conditions?'. This applied specificity is its contribution.

Each pillar corresponds to a system constraint: Integrated Perception addresses information fragmentation; Autonomous Orchestration augments coordination capacity under workforce contraction; Adaptive Goal-Seeking responds to multimorbidity beyond static single-disease protocols. Systems are characterised as agentic when they execute sustained multi-step, cross-system workflows toward defined care objectives under structured human oversight.

Table 7 indicates a consistent temporal gradient. Integrated Perception is most developed in bounded domains but constrained by interoperability gaps. Autonomous Orchestration shows the strongest near-term trajectory in single domain automation. Adaptive Goal Seeking remains the most experimentally distant. Near term implementation should prioritise Perception and bounded Orchestration while treating Adaptive Goal Seeking as a medium-term research investment.

4.3. Operational Bottlenecks and AI Entry Points in Ageing Health Systems.

Three high-impact, deployment-ready targets are identified. These form the operational context for the phased roadmap that follows in Section 4.4. (1) Documentation burden in geriatric and community care is the most defensible initial target, combining a well-documented inefficiency with commercially available solutions and relatively contained workflow redesign.[41] (2) Discharge coordination and readmission prevention for multimorbid older adults form the second priority.[19,20] (3) Early deterioration detection through remote monitoring represents a third entry point, targeting high-cost acute events. All three targets share a critical property: they permit structured workflow redesign before AI deployment, reducing reliance on technological substitution.

4.4. A Sociotechnical Phased Roadmap

The roadmap translates the institutional thesis into a governance sequenced implementation strategy across a ten-year horizon. Phase durations are indicative and draw on observed timelines for large scale digital health reform regulatory pathway development cycles for high-risk adaptive technologies and implementation science evidence on pilot to scale transitions. Higher readiness systems may progress at the lower end of these ranges while lower readiness contexts should anticipate longer timelines.

Figure 2. A Sociotechnical Roadmap for AI Implementation in Ageing Health Systems.

Figure 2. A Sociotechnical Roadmap for AI Implementation in Ageing Health Systems.

Phase 1 — Foundational (~0–2 Years): Governance and Infrastructure Readiness

Goal: Create an enabling institutional environment for the safe equitable introduction of AI enabled healthcare platforms.

Minimum conditions for Phase 1 completion include at least one interoperable regional data environment, a regulatory sandbox or controlled experimentation pathway, and validated workflow mapping for priority bottlenecks. These function as gate criteria for Phase 2 investment.

Key Actions:

−

Governance: Establish coordinated national AI in health governance linking demographic strategy with ethical and deployment oversight[16,54]

−

Data Infrastructure: Mandate HL7 FHIR adoption and secure federated architecture positioning legacy migration as capital investment rather than operating expense[14]

−

Regulation: Create regulatory sandboxes for supervised testing of adaptive AI systems and mandate post deployment performance monitoring[34]

−

Procurement: Develop AI-specific procurement frameworks requiring interoperability compliance and bias audits.

−

Public Engagement: Initiate early co-design with older adults’ caregivers and clinicians to build trust and inform workflow specifications[25]

Operational exit criteria for Phase 1 are detailed in Appendix D (Table D1).

Phase 2 — Pilot and Validation (~2–5 Years): Targeted Deployment and Evidence Generation

Goal: Generate rigorous, reproducible evidence on clinical effectiveness, workflow integration feasibility, and cost economics in high-need settings.

Key Actions:

−

Targeted Pilots: Prioritise deployment in structurally constrained settings including rural elderly isolation and long-term care workforce shortages while selecting sites with varying institutional readiness[8]

−

Evidence Standards: Employ rigorous study designs including randomised trials where feasible that follow CONSORT AI[36] and TRIPOD AI[37] guidance with prospective registration and adequate power for clinically meaningful outcomes

−

Workflow Redesign: Specify required workflow changes prior to deployment and recognise that neutral results without redesign indicate implementation misalignment rather than technological failure

−

User Adoption Monitoring: Treat adoption rates, workflow adherence, and clinician override frequency as primary implementation outcomes

−

Workforce Development: Integrate AI literacy and human AI collaboration training into professional education and budget training as a core pilot cost[25]

Structured validation metrics for Phase 2 are summarised in Appendix D (Table D2).

Phase 3 — Scaled Deployment (~5–10 Years): Integration and System Reform

Goal: Transition validated AI-enabled service platforms into routine, population-level healthcare delivery within reformed payment models and national data governance frameworks.

The transition from pilot to scale is historically the most failure prone stage of digital health deployment[35]. Scaled deployment should be conditional on Phase 2 evidence.

Key Actions:

−

National Rollout: Link procurement to demonstrate effectiveness and equity appraisal through health technology assessment bodies while sequencing expansion from sites with the highest digital readiness[14]

−

Legacy Integration: Require interoperability with dominant national EHR systems as a procurement condition and fund system integration as core health infrastructure

−

Adaptive Surveillance: Mandate continuous monitoring for safety bias and performance drift in alignment with FDA and EU AI Act standards[24,34]

−

Payment Reform: Permanently aligned reimbursement with validated AI enabled value-based care models recognising that sustained scale depends on financing reform[9]

4.5. Stakeholder Recommendations

AI implementation will fail if governance, provider institutions, and technology development advance independently. Recommendations below specify concrete mechanisms for each stakeholder group.

Table 8. Stakeholder Recommendations for AI Implementation in Ageing Health Systems.

Stakeholder	Priority Recommendations (Evidence-Derived)
Policymakers and Regulators	1. Replace fixed regulation with adaptive licensing based on a Demographic Urgency Index – A composite index combining dependency ratios, workforce decline, and fiscal capacity should guide regulatory timelines. Countries with higher scores could receive faster approval pathways for eldercare AI, while still requiring post-market monitoring within 12 months. The EU AI Act’s uniform classification system is the main target for reform. (§3.4, §3.6, §4.4) 2. Create sovereign AI-in-health data agreements with shared validation rights – Different privacy laws such as GDPR, PDPA, PIPA, APPI limit cross-border validation of AI systems. Governments should create federated agreements that allow model validation without transferring raw data, including shared audit rights and bias-reporting rules. OECD and WHO should develop a common template. Without such agreements, the evidence base will remain too weak for national-level investment. (§3.4, §4.6, §4.7.3) 3. Replace activity-based reimbursement with Outcome-Linked Technology Contracts – Payment for AI should depend on verified outcomes such as reduced readmissions, delayed institutional care, and clinician time saved. Multi-year contracts should measure these outcomes and be audited annually by an independent health technology body. This approach produces missing economic evidence and aligns commercial incentives with health system sustainability. (§3.3, §3.6, §4.4)
Healthcare Systems and Providers	1. Establish AI Clinical Governance Boards including elder patients and caregivers – Dedicated boards, separate from general informatics committees, should approve AI systems used for patients aged 65+. They should have authority to redesign workflows, suspend poorly performing tools, and request safety reviews. This addresses the governance gap where decisions about replacing or supporting human care are often made by procurement teams rather than clinicians. (§3.2, §4.1, §4.3) 2. Require 90-day shadow deployment before clinical use, with override data reported nationally – Before activation, AI outputs should be reviewed for 90 days but not used in clinical decisions. Override rates and disagreements should be submitted to a national registry. This creates real-world validation data, establishes baseline performance, and triggers review if post-activation results differ significantly. (§3.2, §3.4, §4.4) 3. Replace basic AI literacy training with human-AI teamwork programmes – Training should focus on practical skills: judging when to trust or override AI, maintaining human-led care roles, and monitoring system performance. These competencies should be included in mandatory professional development for geriatricians, nurses, and care coordinators, and required before unit-level deployment approval. (§3.4, §3.5, §4.4)
Technology Developers	1. Publish geriatric-specific model cards showing performance in older populations –Standard model cards hide weaknesses in key groups such as patients aged 80+, people with multiple conditions, and those with cognitive impairment. Developers should report to subgroup performance and training dataset demographics before market entry. This moves bias detection from post-deployment audits to pre-procurement validation. (§3.4, §3.5) 2. Design eldercare AI systems with human-resumable workflows – All multi-step AI processes should allow clinicians to pause, review, or complete tasks manually without losing data. Each handover should generate a plain-language clinical summary. This addresses governance risks in autonomous AI systems and meets EU AI Act human oversight requirements. (§3.4, §4.2, §4.4) 3. Replace proprietary deployment with shared validation partnerships – Real-world performance data should be jointly owned by developers and health systems, with anonymised results published within 24 months. In return, health systems provide controlled access to longitudinal data. This turns commercial deployment into a research partnership and produces the long-term evidence currently missing. (§3.3, §4.6, §4.7.2)

Table 8. Stakeholder Recommendations for AI Implementation in Ageing Health Systems.

Stakeholder	Priority Recommendations (Evidence-Derived)
Policymakers and Regulators	1. Replace fixed regulation with adaptive licensing based on a Demographic Urgency Index – A composite index combining dependency ratios, workforce decline, and fiscal capacity should guide regulatory timelines. Countries with higher scores could receive faster approval pathways for eldercare AI, while still requiring post-market monitoring within 12 months. The EU AI Act’s uniform classification system is the main target for reform. (§3.4, §3.6, §4.4) 2. Create sovereign AI-in-health data agreements with shared validation rights – Different privacy laws such as GDPR, PDPA, PIPA, APPI limit cross-border validation of AI systems. Governments should create federated agreements that allow model validation without transferring raw data, including shared audit rights and bias-reporting rules. OECD and WHO should develop a common template. Without such agreements, the evidence base will remain too weak for national-level investment. (§3.4, §4.6, §4.7.3) 3. Replace activity-based reimbursement with Outcome-Linked Technology Contracts – Payment for AI should depend on verified outcomes such as reduced readmissions, delayed institutional care, and clinician time saved. Multi-year contracts should measure these outcomes and be audited annually by an independent health technology body. This approach produces missing economic evidence and aligns commercial incentives with health system sustainability. (§3.3, §3.6, §4.4)
Healthcare Systems and Providers	1. Establish AI Clinical Governance Boards including elder patients and caregivers – Dedicated boards, separate from general informatics committees, should approve AI systems used for patients aged 65+. They should have authority to redesign workflows, suspend poorly performing tools, and request safety reviews. This addresses the governance gap where decisions about replacing or supporting human care are often made by procurement teams rather than clinicians. (§3.2, §4.1, §4.3) 2. Require 90-day shadow deployment before clinical use, with override data reported nationally – Before activation, AI outputs should be reviewed for 90 days but not used in clinical decisions. Override rates and disagreements should be submitted to a national registry. This creates real-world validation data, establishes baseline performance, and triggers review if post-activation results differ significantly. (§3.2, §3.4, §4.4) 3. Replace basic AI literacy training with human-AI teamwork programmes – Training should focus on practical skills: judging when to trust or override AI, maintaining human-led care roles, and monitoring system performance. These competencies should be included in mandatory professional development for geriatricians, nurses, and care coordinators, and required before unit-level deployment approval. (§3.4, §3.5, §4.4)
Technology Developers	1. Publish geriatric-specific model cards showing performance in older populations –Standard model cards hide weaknesses in key groups such as patients aged 80+, people with multiple conditions, and those with cognitive impairment. Developers should report to subgroup performance and training dataset demographics before market entry. This moves bias detection from post-deployment audits to pre-procurement validation. (§3.4, §3.5) 2. Design eldercare AI systems with human-resumable workflows – All multi-step AI processes should allow clinicians to pause, review, or complete tasks manually without losing data. Each handover should generate a plain-language clinical summary. This addresses governance risks in autonomous AI systems and meets EU AI Act human oversight requirements. (§3.4, §4.2, §4.4) 3. Replace proprietary deployment with shared validation partnerships – Real-world performance data should be jointly owned by developers and health systems, with anonymised results published within 24 months. In return, health systems provide controlled access to longitudinal data. This turns commercial deployment into a research partnership and produces the long-term evidence currently missing. (§3.3, §4.6, §4.7.2)

4.6. Future Research Agenda

Sections 3.2 to 3.5 reveal three key research gaps. If these gaps remain, evidence will not grow fast enough to support ageing health systems. This agenda proposes three research tracks.

Track 1: Long-term Clinical and Economic Evaluation

The largest gap is the lack of long-term evidence. Multiyear studies are needed to measure effects on healthy life expectancy, mortality, quality of life and long-term healthcare costs in older populations. Economic impact is hard to measure for three reasons. Healthcare spending changes due to ageing wages and policy at the same time. Pilot projects often occur in high-capacity systems and may not reflect national conditions. Efficiency gains may also be offset by growing demand from ageing populations. Future studies should include economists from the start use difference in differences methods that compare regions adopting AI at different times[9,32] link health and social data and compare results across countries using OECD data standards[2,33]. Until stronger evidence exists claims that AI will reduce healthcare spending should be treated cautiously.

Track 2: Institutional Implementation Research

AI implementation varies across countries due to governance financing models and digital infrastructure. However, research comparing these factors is limited. Studies should compare centralised and decentralised systems using common frameworks such as NASSS[35] and CFIR[45]. Research across early adopter countries including Japan South Korea and Singapore can provide lessons for later adopting countries such as Italy and France[32,33]. Another key gap is how reimbursement and procurement systems affect adoption. Payment models often determine whether proven AI tools become funded healthcare services[9,16].

Track 3: Human–AI Care Integration

Efficiency gains may hide declines in human care quality. This risk is institutional rather than purely technical. Future studies should measure outcomes such as patient dignity, social connection human contact time and continuity of care in all Phase 2 pilots[22,40]. Research should also examine how AI tools such as ambient documentation affect clinician attention reasoning and job satisfaction over time[41,42]. Caregiver impact is another important gap. Studies should assess whether ageing in place technologies reduce or shift informal caregiving burden within families[10,31]. Together these three research tracks form the evidence needed to turn the proposed roadmap into a tested policy framework.

4.7. Limitations

4.7.1. Review Methodology Limitations

This study uses a structured narrative review to integrate evidence from multiple disciplines including health policy economics artificial intelligence and gerontology. Full PRISMA compliance was not achievable given the heterogeneous grey literature included. The review is better described as PRISMA-informed. A structured review protocol was developed prior to data extraction but was not formally registered; readers should interpret findings accordingly and future replication studies should register prospectively.

The evidence base is concentrated in high resource digitally mature health systems which limits generalisability to lower capacity contexts. Most primary studies remain short term pilot deployments which restrict inference about sustained population scale cost effectiveness. Consequently, the macroeconomic moderation hypothesis proposed in this paper remains unvalidated on the national scale.

Another limitation concerns the generalisability of AI systems. Many models are trained on specific clinical datasets and may not perform consistently across different populations health systems or care environments. Performance may vary due to differences in data quality, clinical practice patterns, demographic structure and health system organisation. This heterogeneity limits transferability of pilot results in broader healthcare contexts.

4.7.2. Implementation and Infrastructure Risks

Large scale AI deployment requires substantial computational and energy infrastructure often overlooked in health policy discussions. Capital costs for servers’ storage network upgrades and cooling infrastructure must therefore be included in health system investment planning. Jurisdictions such as the European Union and Singapore also operate under net zero climate commitments which may constrain expansion of energy intensive digital infrastructure without renewable energy sourcing.

Health systems may also face vendor locking risks when procuring AI platforms from a small number of dominant providers. Over time, this concentration may reduce strategic autonomy over clinical data infrastructure workflow design and procurement leverage particularly under demographic pressure to scale solutions rapidly[55]. The OECD has noted that digital public infrastructure requires governance mechanisms that prevent monopolistic concentration in procurement markets[55]. Mitigation strategies include mandatory open API interoperability requirements performance linked contracts with independent audit rights and phased pilot structures that preserve competitive tendering.

4.7.3. Global Data Governance Constraints

Cross-border AI evaluation is constrained by fragmented international privacy and data governance frameworks. Regulations such as the European Union’s General Data Protection Regulation (GDPR) and the emerging European Health Data Space differ significantly from Singapore’s Personal Data Protection Act (PDPA), Japan’s Act on the Protection of Personal Information, and South Korea’s Personal Information Protection Act.[33,56,57] These differences complicate cross-border model validation, federated learning, and multi-country evaluation designs identified as priorities in Section 4.6.

Until internationally coordinated frameworks for responsible cross border health data exchange emerge potentially under WHO or OECD leadership the global evidence for AI enabled healthcare will remain fragmented across jurisdictions. Taking together these limitations indicate that the roadmap proposed in this study should be interpreted as a governance-oriented implementation hypothesis that requires empirical validation through the research agenda outlined in Section 4.6.

4.8. Policy Impact and Potential Policy Outcomes

Section 4.5 outlines the institutional responsibilities of policymakers, healthcare systems, and technology developers. This section considers the system-level outcomes that could plausibly follow from implementation of the proposed roadmap. While validated macroeconomic proof is not yet available, the evidence reviewed provides a reasonable basis for estimating potential policy impact.

Healthcare System Outcomes

−

Avoidable admissions: Remote monitoring shows 20–30% reductions in emergency utilisation in structured pilots[19,20]. Evidence moderate at pilot level and plausible at population level.

−

Clinician administrative burden: Ambient documentation reduces documentation time by at least 15 percent and increases patient facing time[41]. Evidence emerging and mainly observational.

−

Early deterioration detection: Validated AI algorithms enable earlier escalation of care which may reduce severity and cost of acute episodes[19,20]. Evidence moderates structured settings and emerging in community eldercare.

−

Care coordination: The Autonomous Orchestration pillar of AAF 3.0 could reduce fragmented specialist management in interoperable systems. Evidence emerging though the mechanism is theoretically grounded within Phase 3 horizons[35,42].

Economic Outcomes

−

Healthcare expenditure moderation: If AI interventions reduced projected expenditure growth by 5 to 10 percent relative to IMF projections[46] advanced economies could avoid about 0.1 to 0.35 percentage points of GDP annually by mid-century. Scenario is based on plausible extrapolation from moderate micro level evidence.

−

Workforce productivity: AI supported documentation and coordination could increase effective capacity of care workers by roughly 10 to 20 percent under optimistic but plausible conditions which could partially offset Japan projected workforce shortfall[8,41].

−

Long-term care costs: Delaying institutionalisation by six to twelve months through ageing in place technologies could generate substantial individual long term care savings[31]. Evidence is moderate at the individual level but dependent on population scale adoption.

Workforce and Social Outcomes

−

Workforce shortage mitigation: AI workflow tools may extend staff productivity and reduce burnout related attrition in geriatric care settings[28,41]. Evidence plausible with partial empirical support.

−

Ageing in place: Integrated remote monitoring and community care can extend independent living, reduce institutional demand and improve wellbeing[31,32]. Evidence is moderate but dependent on system design.

−

Caregiver burden and rural access: AI enabled telehealth can expand geographic access to care and may reduce caregiving pressure on family networks. Evidence emerges and requires further evaluation.

These outcomes support investment in the phased roadmap even under conservative assumptions. Healthcare and social benefits are supported by evidence from controlled settings while economic impacts require the longitudinal research programme described in Section 4.6 to move from plausible scenarios to validated projections.

5. Conclusion

This review examined whether Agentic AI could address the structural pressures of demographic ageing in advanced economies. AI demonstrates measurable capability in bounded clinical and administrative domains. However, a consistent pattern across the reviewed literature is that system-level impact is constrained more reliably by institutional than by technical factors, and macroeconomic fiscal moderation remains empirically unvalidated.

AI is neither a substitute for demographic policy nor a self-executing technological solution. Its contribution depends on alignment between governance reimbursement interoperability and workforce structures so that augmentation technologies can be integrated without degrading relational care or widening inequities. Under demographic compression, governance capacity becomes as important as algorithmic capability.

The AAF 3.0 domain-mapping matrix distinguishes deployable capabilities from near term and speculative horizons and links these capabilities to structural pressures within ageing health systems. The phased roadmap positions AI adoption within governance sequencing and recognises institutional reform as a prerequisite for sustainable scale. Singapore integrated five domain strategy provides a benchmark for governance first sequencing as discussed in Sections 4.5.1 and 4.5.2 and illustrated in the policy impact scenarios in Section 4.8.

The research agenda in Section 4.6 outlines the longitudinal comparative and relational evidence needed to transform AI deployment from a plausible hypothesis into a validated policy instrument. Structural constraints including energy infrastructure requirements, vendor dependency risks and fragmented international data governance must also be addressed during implementation planning. The future of ageing societies will depend less on what AI systems can do than on whether institutions are prepared to deploy them responsibly equitably and sustainably.