Open-source AI offers transparency, customization and privacy advantages while proprietary systems dominate on reliability and vendor support; policymakers should adopt a hybrid, tiered governance and certification regime to capture complementary benefits without entrenching inequities or market concentration.
Framework for Government Policy on Agentic and Generative AI in Healthcare: Governance, Regulation, and Risk Management of Open-Source and Proprietary Models
Open-source and proprietary healthcare AI each bring complementary strengths and distinct risks, and a hybrid, tiered governance framework that combines transparency, validation, and vendor accountability is recommended to maximize benefits while containing harms.
This paper provides a comprehensive review and strategic framework to navigate this complex ecosystem of open-source and proprietary models for healthcare. We analyze the technical capabilities, implementation challenges, and governance requirements of both AI paradigms through a systematic and organnized survey of current literature and emerging trends. Our findings indicate that while open-source models offer superior transparency, customization, and data privacy—increasingly rivaling proprietary performance in diagnostics—proprietary systems maintain advantages in reliability, support, and integration. However, AGI also introduces complex risks ranging from algorithmic bias (if uncontrolled) to regulatory fragmentation (lack of regulation). Evidence shows concerning patterns in automated decision appeals and significant financial barriers to implementation that could limit accessibility. To address these challenges, we propose a tiered risk-management and governance framework that synthesizes the strengths of both open and closed-source approaches. Our recommendations include the adoption of international certification protocols aligned with global explainability standards, federated learning architectures to ensure privacy while enabling collaboration, and adaptive policymaking to balance innovation with patient safety. This integrated approach aims to maximize the benefits of both open-source and proprietary AI while focusing on remodification of unique risks posed by agentic systems.
Summary
Citation: Joshi, S. (2026). Framework for Government Policy on Agentic and Generative AI in Healthcare: Governance, Regulation, and Risk Management of Open-Source and Proprietary Models. International Journal of Innovative Research in Computer Science and Technology, 14(1), 94–115. DOI: https://doi.org/10.55524/ijircst.2026.14.1.12
Main Finding
The paper argues that a hybrid, risk‑tiered governance and deployment framework—combining the transparency and adaptability of open‑source models with the reliability, validation, and support of proprietary systems—is the most practical public‑policy approach for safe, effective uptake of agentic and generative AI (AGI) in healthcare. It recommends international certification, federated learning for privacy‑preserving collaboration, and adaptive regulation to balance innovation, equity, and patient safety.
Key Points
- Technical convergence: Recent open‑source models now closely match proprietary models on many diagnostic tasks (reported AUC ranges: open‑source 0.92–0.95 vs proprietary 0.94–0.96 in complex diagnostics).
- AGI capabilities and risks: Agentic systems can deliver high clinical performance in controlled settings (examples: ~89% AUC outcome prediction; 92% cancer screening in cited studies) but introduce novel governance risks—algorithmic bias, opaque autonomous actions, and high administrative error/appeal rates (paper cites ~73% appeal rate in AI insurance denials).
- Economic tradeoffs:
- Implementation cost range: ~$250K–$2M per system; workforce retraining estimates cited at $1.4B.
- ROI (reported from literature): open‑source solutions 180–250% over 3 years vs proprietary 120–180% (paper synthesizes multiple sources).
- Market projection: $29.01B (2024) → $504.17B (2032), CAGR ≈ 37.2%.
- Deployment strategy: Proposes intelligent orchestration ("hybrid architecture") that routes high‑risk clinical decisions to validated proprietary models and lower‑risk tasks to open‑source implementations.
- Policy recommendations: tiered risk management, global explainability/certification standards, federated learning architectures, adaptive policymaking to prevent regulatory fragmentation and to enable safe innovation.
- Quantified operational gains (from cited studies): 30–45% reduction in diagnostic time, 25–40% improvement in administrative efficiency; multi‑agent systems could reduce chronic care costs by up to ~40% in optimistic scenarios.
- Performance and safety metrics summarized: standard diagnostic metrics (accuracy, precision, recall, AUC, F1), economic equations (TCO, ROI, ICER/QALY), agentic metrics (task completion rate, autonomy level), and reliability metrics (MTBF, error rates).
Data & Methods
- Evidence base: Systematic literature survey of ~25 contemporary sources (journals, conferences, technical reports, industry publications) focused on 2023–2025 material; synthesizes peer‑reviewed and industry studies cited throughout the paper.
- Analytical approach:
- Comparative framework across dimensions: performance, security/privacy, customization/adaptability, cost/accessibility, transparency/accountability.
- Quantitative modeling: standard diagnostic statistics (confusion matrix metrics, AUC), economic models (TCO, ROI, CAGR, logistic growth for market share), QALY/ICER cost‑effectiveness analysis.
- Agentic performance metrics added (task completion rate, autonomy level, human intervention frequency) and reliability/safety measures (MTBF, error rates).
- Validation practices reported in cited studies: sample sizes frequently in the 10k–50k case range for model validation, statistical power ~0.9, confidence intervals ±1.5–2% reported for accuracy metrics.
- Visuals and frameworks: Hybrid architecture diagrams, intelligent data‑flow routing, performance radar charts, adoption curves and strategic decision flowcharts (described but figures not reproduced here).
- Limitations noted or implied in the paper:
- Heavy reliance on secondary literature and heterogeneous sources (mix of peer‑reviewed and industry reports).
- Some performance/cost figures are drawn from individual studies or optimistic scenarios; generalizability may be limited across contexts and care settings.
- Regulatory and economic projections (market size, ROI ranges) depend on assumptions about adoption rates, regulatory harmonization, and technology evolution that may change rapidly.
Implications for AI Economics
- Market structure and competition:
- Open‑source gains in parity lower technological barriers and can increase competition, potentially reducing vendor lock‑in and licensing rents. Reported higher ROI for open‑source (180–250%) suggests stronger cost‑efficiency for adopters, especially in resource‑constrained settings.
- Proprietary firms retain advantages in validated, high‑risk clinical segments where regulatory certification and commercial support are valued—supporting a segmented market with coexistence of open and closed models.
- Investment and adoption dynamics:
- High upfront implementation costs ($250K–$2M) and retraining expenditures create entry costs and may slow diffusion; however, high projected returns and operational gains (time savings, throughput improvements) provide strong adoption incentives for larger providers and payers.
- Regulatory fragmentation increases compliance costs and uncertainty, raising the hurdle for smaller entrants; international certification harmonization would lower transaction costs and facilitate cross‑border markets.
- Distributional and equity effects:
- Open‑source adoption can democratize access to advanced AI in lower‑resourced health systems (lower licensing costs, offline deployment for privacy), potentially narrowing health technology gaps.
- Conversely, financial and regulatory barriers for AGI implementations could concentrate advanced capabilities among better‑funded institutions, risking uneven diffusion.
- Externalities and public goods:
- Open‑source development produces knowledge spillovers and public‑good benefits (auditability, transparency) that can accelerate innovation but may require governance mechanisms to manage safety risks.
- Proprietary models internalize validation/support costs but may underprovide transparency; policy interventions (e.g., disclosure requirements, certification that includes explainability criteria) can correct these market failures.
- Policy instruments and economic tradeoffs:
- Tiered risk regulation recommended by the paper aligns regulatory stringency with clinical risk—this reduces unnecessary compliance costs for low‑risk innovation while protecting patients in high‑risk domains.
- Federated learning and privacy‑preserving collaboration reduce data‑monopoly rents and enable cross‑institutional learning, but introduce coordination costs and require investments in secure infrastructure.
- International certification and explainability standards lower information asymmetries and transaction costs, supporting more efficient procurement and smoother market entry for compliant products.
- Overall economic outlook:
- If the projected CAGR and performance improvements materialize, healthcare AI could generate large productivity gains and welfare benefits. Realizing these gains depends on policy choices that balance incentives for private investment, openness for innovation diffusion, and robust governance to manage agentic risks.
Notes and caveats: the summary synthesizes claims and quantitative results as reported in the paper, which itself aggregates multiple external studies of varying provenance. Policymakers and economists should treat point estimates (AUCs, ROI ranges, cost figures, appeal rates) as indicative rather than definitive and prioritize independent, context‑specific cost‑benefit and safety evaluations before large‑scale procurement or regulatory commitments.
Assessment
Paper Typereview_meta
Evidence Strengthn/a — The paper is a literature synthesis and conceptual analysis rather than an empirical study designed to estimate causal effects; it aggregates peer-reviewed studies, industry reports, and deployment observations of varying quality rather than presenting new causal identification or primary experimental evidence.
Methods Rigormedium — The authors perform a cross-disciplinary, systematic-style survey and thematic synthesis that draws on diverse sources, but they do not conduct a formal meta-analysis, pre-registered systematic review with strict inclusion criteria, or original empirical testing; findings are therefore subject to heterogeneity in underlying study designs, reporting bias, and rapidly evolving evidence.
SampleAggregated evidence from peer-reviewed clinical and technical studies, industry reports, case studies of deployments, and observed trends in open-source and proprietary healthcare AI systems; no new primary clinical trials or randomized field experiments were conducted.
Themesgovernance adoption productivity inequality org_design
GeneralizabilityRapidly evolving AI model capabilities (especially LLMs/agentic systems) limit the longevity of conclusions, Findings are healthcare-specific and may not generalize to non-health sectors, Bias toward settings and vendors that are documented in the literature (likely higher-resource health systems and major markets), Heterogeneity in study designs, populations, and clinical contexts reduces comparability, Regional regulatory and infrastructure differences mean policy recommendations may not transfer across jurisdictions
Claims (19)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Open-source models provide greater transparency and inspectability, enabling better auditability and explainability. Ai Safety And Ethics | positive | high | transparency / auditability / explainability |
0.04
|
| Open-source models enable customization and local retraining that can align models with institutional workflows and patient populations. Organizational Efficiency | positive | medium | model alignment with local workflows / local performance |
0.02
|
| Open-source deployment options (e.g., on-premises) reduce data-sharing exposure and improve privacy. Ai Safety And Ethics | positive | high | data privacy / data-sharing exposure |
0.04
|
| Open-source models show narrow but growing parity with proprietary models on some diagnostic tasks. Output Quality | mixed | medium | diagnostic performance / accuracy on specific tasks |
0.02
|
| Proprietary systems lead on reliability, maintenance, and validated integrations with clinical systems. Organizational Efficiency | positive | high | system reliability / maintenance burden / integration maturity |
0.04
|
| Vendor support, warranties, and service-level agreements (SLAs) are important for clinical adoption and liability management. Adoption Rate | positive | high | clinical adoption / liability mitigation |
0.04
|
| Centralized updates and monitoring by vendors can reduce operational burden for healthcare providers. Organizational Efficiency | positive | medium | operational burden / maintenance effort |
0.02
|
| Both open-source and proprietary approaches carry risks of algorithmic bias and fairness violations, especially when models are uncontrolled or poorly validated across populations. Ai Safety And Ethics | negative | high | bias / fairness metrics / differential performance across populations |
0.04
|
| Regulatory fragmentation and lack of harmonized standards increase compliance complexity for healthcare AI deployments. Regulatory Compliance | negative | high | regulatory compliance complexity / administrative burden |
0.04
|
| Emerging agentic/AGI capabilities introduce new failure modes and governance challenges that standard ML oversight may not cover. Ai Safety And Ethics | negative | speculative | governance risk / novel failure modes |
0.0
|
| There is evidence of problematic patterns in automated decision appeals and workflow interactions when AI is integrated into clinical processes. Organizational Efficiency | negative | medium | workflow burden / frequency of appeals / process failures |
0.02
|
| Significant financial and implementation barriers (infrastructure, staff, validation) risk worsening access inequities between well-resourced and low-resource providers. Inequality | negative | high | access / equity disparities / adoption gap by resource level |
0.04
|
| Open-source lowers licensing fees but can shift costs toward in-house engineering, governance, and validation. Firm Productivity | mixed | medium | total cost of ownership / cost allocation between licensing and internal expenses |
0.02
|
| Proprietary models concentrate costs into vendor payments and can potentially lower internal operational burden for providers. Firm Productivity | mixed | medium | vendor payments / internal operational burden |
0.02
|
| Federated learning and privacy-preserving collaboration can combine data advantages without centralizing sensitive records and may reduce duplicated validation costs over time. Ai Safety And Ethics | positive | medium | data centralization risk / validation costs / privacy-preserving data utility |
0.02
|
| A tiered risk-management framework that allocates governance intensity to interventions by clinical criticality and autonomy is recommended to maximize benefits while containing harms. Governance And Regulation | positive | medium | governance effectiveness / risk mitigation by intervention tier |
0.02
|
| International certification protocols tied to explainability and safety standards would influence investment incentives and market structure. Market Structure | positive | medium | investment incentives / market concentration / compliance-driven market effects |
0.02
|
| Reliable, well-integrated AI may raise clinical productivity and shift labor toward higher-value tasks, but misaligned deployments risk increased administrative burden (e.g., appeals, oversight). Organizational Efficiency | mixed | medium | clinical productivity / labor allocation / administrative burden |
0.02
|
| Economic outcomes of healthcare AI depend critically on governance design: policies and technical architectures (e.g., federated learning, certification standards, tiered risk management) will determine whether mixed open/proprietary ecosystems yield broad welfare gains or entrench inequities and concentrated market power. Inequality | mixed | medium | welfare distribution / market concentration / equity outcomes |
0.02
|
Entities
Open-source AI models (ai_tool)
Proprietary AI systems (ai_tool)
Large Language Models (LLMs) (ai_tool)
Tiered risk-management framework (method)
Systematic literature survey (method)
Diagnostic performance (outcome)
Algorithmic bias / Fairness violations (outcome)
Access and equity (outcome)
Patients / patient populations (population)
Agentic / AGI capabilities (ai_tool)
Federated learning (method)
Local retraining / customization (method)
International certification protocols (method)
Adaptive, iterative policymaking (method)
Cross-disciplinary literature synthesis (method)
Comparative analysis (method)
Transparency / Inspectability (outcome)
Explainability (outcome)
Reliability and integration (outcome)
Financial and implementation barriers (outcome)
Clinical productivity and labor effects (outcome)
Market concentration risks (outcome)
Healthcare providers (population)
Certification standards and explainability requirements (method)
Privacy-preserving collaboration (method)
On-premises deployment (method)
Thematic extraction (method)
Large health systems (population)
Smaller providers (population)
Low-resource settings (population)
Open-source communities (institution)
Proprietary vendors (institution)