We conducted a Bayesian budget-impact simulation to estimate AI-driven cost savings across 3 health care sectors in the United Kingdom and Australia. Rather than fitting a traditional hierarchical model, we used a Bayesian simulation framework that synthesizes evidence from heterogeneous implementation studies while explicitly incorporating uncertainty in effectiveness, adoption, and implementation risk. The primary time horizon was 2024 for the cross-sectional analysis and was extended to 2024-2030 using inflation-adjusted temporal projections based on annual inflation rates of 4% for the United Kingdom and 5% for Australia.

We adopted a payer/provider perspective, using the National Health Service (NHS) for the United Kingdom and the Australian Government for Australia. The primary estimand was annual gross cost savings, defined as the reduction in expenditure attributable to AI-enabled efficiency gains rather than cost offsets or net savings after implementation costs. We therefore did not model implementation costs such as software licensing, system integration, staff training, governance, or cybersecurity. Nor did we model health outcomes, quality-adjusted life-years, or incremental cost-effectiveness ratios because our objective was a budget-impact analysis rather than a cost-effectiveness evaluation.

We drew health care expenditure data from official national statistics for the 2022-2023 period. For the United Kingdom, we used the Office for National Statistics (ONS) estimate of £292 billion in total health care expenditure [1]. For Australia, we used the Australian Institute of Health and Welfare estimate of Aus $252.5 billion [2]. We converted all monetary values to US dollars using International Monetary Fund (IMF) December 2023 spot exchange rates of £1=US $1.27 and the Reserve Bank of Australia (RBA) December 2023 spot rate of Aus $1=US $0.68. We used spot rates because they align with the official expenditure reports. Purchasing power parity rates could alternatively be used for welfare comparisons, but they would not materially alter either the direction or the rank ordering of our results. Full expenditure-base calculations are reported in Table S3 in Multimedia Appendix 1.

We derived sector expenditure weights from ONS, the Australian Institute of Health and Welfare (AIHW), and OECD: Organization for Economic Co-operation and Development (OECD) health accounts [1,2,20,21]. Radiology accounted for 0.83% of total health care spending in the United Kingdom and 2.8% in Australia [22-24]. Workflow optimization, defined here as health administration and clinical workflow process costs, accounted for 2% of UK expenditure (approximately £5.84 billion) and 2.77% of Australian expenditure (approximately Aus $7.0 billion) [1,2]. Workforce optimization, capturing compensation of employees, accounted for 30% of UK health care expenditure and, for Australia, was parameterized using the workforce payroll component of recurrent public hospital expenditure, equal to Aus $56.62 billion (representing 52.9% of recurrent public hospital expenditure and an effective model weight of 22.4% of total Australian health care expenditure) [20,21]. These shares imply approximate sector bases of £2.4 billion (US $3.05 billion) for UK radiology, Aus $7 billion (US $4.76 billion) for Australian radiology, £5.84 billion (US $7.42 billion) for UK workflow optimization, Aus $7.0 billion (US $4.76 billion) for Australian workflow optimization, £87.6 billion (US $111.25 billion) for UK workforce optimization, and Aus $56.62 billion (US $38.5 billion) for Australian workforce optimization [1,2,20-24].

As detailed below, our structured review yielded only 3 to 4 studies per domain, although all had sample sizes of at least 100 participants. That relatively sparse evidence base motivated the use of informative priors and underscores the need for future meta-analyses and registry-based evidence.

We modeled annual cost savings (S) for sector i in country j as:

where B_j represents base health care expenditure, w_i,j denotes sector weight, A_i,j is the adoption rate, E_i,j represents AI effectiveness, and R_i,j captures implementation risk. In practical terms, equation 1 multiplies total health care spending by the sector’s share, the fraction of activity that adopts AI, the cost-reducing effect of AI, and the probability that implementation succeeds, yielding each sector’s expected cost saving. The multiplicative specification treats adoption rate A and implementation risk R as statistically independent. Although higher-risk environments may exhibit slower AI adoption in practice (introducing negative correlation), the current literature does not provide joint (A, R) data sufficient to parameterize a bivariate distribution. The independence simplification is therefore a necessary modeling choice whose conservative implications—a modest upward bias in projected savings—are discussed in the Limitations section.

We assumed that observed cost savings followed a normal distribution centered on the value implied by equation 1. We treated the SD, σ, as an estimable parameter and assigned it a sector-country-specific exponential prior with a mean equal to 15% of the observed savings estimate for that sector-country pair. This heteroscedastic specification scales absolute estimation noise to the magnitude of projected savings, so sectors with larger projected savings also carry larger absolute uncertainty. The implied prior mean values for ranged from US $12.0 million for Australian radiology to US $87.2 million for UK workforce optimization, reflecting the variation in observed savings across sectors.

The savings equation links 3 sector-specific latent parameters—adoption rate (A), effectiveness (E), and implementation risk (R)—to 1 or more constructed savings observations per sector-country pair, derived from the eligible literature and converted to comparable annual monetary values. In the current implementation, radiology uses 2 constructed savings anchors per country, whereas workflow optimization and workforce optimization each use 1. To improve identification, we augmented the savings likelihood with sector-specific auxiliary observations for effectiveness and implementation failure so that the joint log-likelihood is the sum of the savings term, the effectiveness-observation term, and the implementation-failure term. The augmented log-likelihood is given by equation 2:

where y^S_ij, y^E_ij, and y^R_ij denote the observed savings, effectiveness, and implementation-failure anchors, respectively, with corresponding measurement uncertainties σ^S_i,j, σ^E_i,j, and σ^R_i,j derived from the literature. This specification anchors E to literature-based cost-reduction estimates and R to empirical deployment-failure rates, while adoption A is estimated conditional on those anchored quantities. The corresponding observed inputs and measurement uncertainties are described in the Prior Distributions section, and Table S1 in Multimedia Appendix 2 reports the observed cost-savings values used for the United Kingdom and Australia. We note that with 3 latent parameters (A, E, R) per sector-country pair—updated against 2 constructed savings observations for radiology and 1 each for workflow and workforce optimization—the model is formally underidentified. Equation 2 partially addresses this by anchoring E and R with independent observations, but residual sensitivity to prior specifications remains and cannot be fully eliminated without sector-specific multioutcome datasets. A prior sensitivity analysis reported in Multimedia Appendix 3 demonstrates that sector rank orderings and order-of-magnitude savings estimates are robust to ±20% perturbations in individual prior means and to alternative σ scaling factors.

We estimated posterior distributions using sequential Monte Carlo (SMC) with 1000 particles [25,26]. We chose SMC over conventional Markov chain Monte Carlo (MCMC) because it requires relatively little tuning and is well suited to analyzing multiple datasets. The SMC algorithm samples through a sequence of intermediate distributions that gradually transform the prior into the posterior by annealing the likelihood. At each iteration, we selected the next temperature so that the effective sample size remained close to 500. We then resampled to eliminate low-weight particles and replicate high-weight particles. After resampling, we applied an MCMC kernel to perturb duplicated particles and reduce particle degeneracy. The number of MCMC iterations was set adaptively using the estimated acceptance rate. Throughout the algorithm, the number of unique particles remained above 900, indicating stable sampler performance. This process is described in Table S3 in Multimedia Appendix 2.

We constructed prior distributions from a structured review of major health care AI studies with sample sizes of at least 100 participants and quantifiable effectiveness estimates. Complete parameter values are reported in Table S1 in Multimedia Appendix 1, and Table S2 in Multimedia Appendix 1 maps those values to the underlying literature.

The effectiveness parameter E_i,j is defined domain-specifically and is not directly commensurate across sectors. In radiology, E represents the proportional reduction in reading cost per screening episode attributable to AI triage. In workflow optimization, E represents the proportional reduction in cost per treatment course or workflow episode. In workforce optimization, E represents the proportional reduction in overtime and excess scheduling costs. These cost categories are mutually exclusive by design: radiology savings affect reading cost per episode; workflow savings affect process cost per course; workforce savings affect labor allocation costs. Adoption rate (A) represents the proportion of eligible activities within the sector that are subject to AI-assisted delivery. Implementation risk (R) represents the probability that a given AI deployment fails to achieve its projected savings within the planned implementation horizon.

For effectiveness, we used truncated normal (TN) priors that reflect both the empirical literature and domain-specific constraints. In radiology, studies by Lång et al [27] and Marinovich et al [28] show that AI-assisted screening can improve diagnostic accuracy and workflow efficiency, although formal modeling of cost-effectiveness or combined deployment impact remains limited [29]. We therefore specified UK radiology effectiveness as TN(0.459, 0.128; lower=0.01, upper=0.6). Workflow optimization used TN(0.55, 0.12; lower=0.01, upper=0.8), anchored primarily to the radiation therapy economic analysis by Natesan et al [30]. Workforce optimization used TN(0.30, 0.10; lower=0.01, upper=0.6), informed by scheduling, overtime, and workforce-cost studies [31,32]. For Australia, we used TN(0.38, 0.06) for radiology, TN(0.50, 0.15) for workflow optimization, and TN(0.30, 0.12) for workforce optimization, with the same lower and upper bounds.

We modeled adoption rates with TN priors bounded between 0.01 and 0.35. UK adoption followed TN(0.12, 0.09), and Australian adoption followed TN(0.15, 0.10), informed by national digital health strategy reports and implementation statistics [33-35]. The upper bound of 35% reflects the maximum penetration we considered plausible under existing national AI strategies for 2024; exceeding that level would likely require infrastructure and workforce investments beyond current plans. We also capped effectiveness at 0.6 in radiology and workforce optimization, where human oversight remains substantial, and at 0.8 in workflow optimization, where the automation potential is greater and clinical risk thresholds are lower [33-35].

We specified implementation risk using sector-specific beta distributions. Radiology risks followed Beta(2.3, 1.7), representing a 57% mean risk; workflow optimization used Beta(2.8, 2.2), representing a 56% mean risk; and workforce optimization used Beta(3.5, 3.5), representing a 50% mean risk. These parameters were derived from a systematic review of AI implementation studies, particularly the analysis of deployment failures by Beede et al [17] and an evaluation of clinical integration challenges by Sendak et al [36]. We then further constrained these priors using direct implementation-failure observations in the augmented likelihood. For UK radiology, a study of the NHS AI chest diagnostics program by Ramsay et al [37] reported that 23 of 66 funded NHS trusts had not achieved clinical use at the time of reporting, yielding a failure-rate estimate of 0.348 with measurement uncertainty of 0.059 [37]. Because all 66 trusts had already been funded and committed to deployment, we interpret this as a conditional implementation-failure rate rather than a general market-adoption rate. The Royal College of Radiologists’ 2023 Clinical Radiology Workforce Census reported that 46% of NHS trusts were not using AI tools in radiology practice, which is broadly consistent with that estimate [38]. For workflow optimization and workforce optimization in both countries, and for Australian radiology, where equivalent national deployment surveys were unavailable, we used a pooled estimate. A systematic review of 39 AI randomized controlled trials by Lam et al [39] found that 18 studies, or 46.2%, failed to demonstrate clinically meaningful outcomes. Further, a systematic review of 14 real-world AI implementations by Yin et al [40] found that 6, or 42.9%, did not improve patient outcomes. Inverse-variance pooling of those estimates yielded an implementation-failure anchor (Y_R) of 0.45 with SE (σ_R) of 0.07.

We analyzed 3 health care sectors chosen for their differing levels of AI maturity and potential cost impact. Radiology included diagnostic imaging services, with savings anchors derived from breast-screening studies [27-29]. Workflow optimization covered administrative and clinical workflow processes, but in the present implementation, its quantitative savings anchor is the radiation therapy course-level cost reduction reported by Natesan et al [30]. Workforce optimization covered staff scheduling, workload distribution, and broader human resource management, with prior evidence drawn from scheduling, overtime, and workforce-cost studies [31,32].

In each sector, we used country-specific parameters to reflect institutional differences between the United Kingdom’s centralized NHS structure and Australia’s more federated health care system.

We examined 3 main scenarios to test the sensitivity of our results to implementation conditions, with full results reported in Table S1 in Multimedia Appendix 3. We applied scenario multipliers to 1000 bootstrap samples drawn from approximate posterior distributions, using normal approximations parameterized by posterior means and SDs. The resulting credible intervals (CrIs) therefore represent the 2.5th to 97.5th percentiles of the bootstrap distributions rather than simple transformations of posterior means. Multimedia Appendix 3 also reports a separate 6-scenario temporal analysis—baseline, high inflation, low inflation, currency volatility, technology acceleration, and regulatory constraint—applied to the multiyear projection model. Note that because the interval estimates rely on a normal approximation to the marginal posteriors rather than the raw particle set, this approximation can impose symmetry on skewed distributions and produce boundary effects at the parameter caps.

The baseline scenario used the core posterior distributions. The optimistic scenario assumed enhanced implementation success with reduced risk parameters, for example, Beta(3.3, 2.7) for radiology, representing a 55% mean risk; increased adoption rates at the 75th percentile of the baseline distributions; and effectiveness parameters at the upper quartile. The pessimistic scenario incorporated higher implementation risks, for example, Beta(1.3, 0.7) for radiology, representing a 65% mean risk; reduced adoption rates at the 25th percentile; and lower effectiveness parameters. The conservative scenario increased implementation risk by 10%, reduced adoption by 10%, and reduced effectiveness by 5% relative to baseline. Together, these scenarios span plausible implementation environments while maintaining internal consistency.

For projections from 2024 to 2030, we defined and extended equation 1 as:

Adoption followed a baseline-preserving logistic learning curve, capped at 35%, with faster uptake assumed in the United Kingdom than in Australia (, where ν is the maximum proportional gain in adoption above the 2024 baseline and τ denotes years since 2024 at the inflection point). Implementation risk decayed geometrically as with δ=0.05, reflecting organizational learning. Effectiveness increased geometrically as with γ=0.03, using caps of 0.60 for radiology and workforce optimization and 0.80 for workflow optimization. Country-specific inflation rates were set at 4% for the United Kingdom and 5% for Australia. Uncertainty was propagated using 1000 Monte Carlo/bootstrap draws from normal approximations to the SMC-derived posterior distributions, parameterized by posterior means and SDs for adoption, implementation risk, effectiveness, and baseline savings. Multimedia Appendix 3 reports the resulting annual and cumulative scenario projections.

External validation against independent national AI cost-savings benchmarks was not feasible because no published assessment matched the sector-level granularity of our model. In place of formal validation, we conducted face-validity checks by comparing posterior sector savings with analogous published estimates. UK workforce optimization savings of US $591.9 million represent roughly 0.53% of the UK workforce compensation base of about US $111 billion, a magnitude broadly consistent with overtime-reduction studies reporting efficiency gains of 2% to 8% in scheduling-optimization settings [31]. UK radiology savings of US $143.9 million against a radiology base of roughly £2.4 billion (about $3.1 billion) represent about 4.6% of sector spending, which is broadly consistent with the 4% to 7% cost-per-study reductions reported in AI-assisted reading studies [11,27]. These checks support the order-of-magnitude plausibility of our projections, although they do not substitute for external validation against detailed cost-accounting data. Expanding the framework to settings with administrative cost-accounting data would enable formal validation against observed cost trajectories; this is the single most important priority for strengthening these estimates.

All analyses were conducted in MATLAB R2023a using custom SMC functions. A complete summary of Bayesian parameter estimates comparing prior and posterior distributions is provided in Table S2 in Multimedia Appendix 2. The complete MATLAB scripts and SMC implementation code are publicly available (see Data Availability section). All data sources are fully cited and traceable to public repositories (ONS, AIHW, and OECD).

This study used only publicly available, aggregated health care expenditure data, published peer-reviewed literature, and national health statistics. We did not use patient-level data, personally identifiable information, or clinical records at any stage. Accordingly, no institutional ethics board or Human Research Ethics Committee review was needed. In the United Kingdom, this study falls outside the scope of the UK Policy Framework for Health and Social Care Research [41], which restricts oversight to research involving NHS patients, their tissue, or their identifiable data; computational modeling based on publicly available national statistics does not constitute such research. In Australia, the study was conducted in accordance with the National Statement on Ethical Conduct in Human Research [42], which exempts from Human Research Ethics Committee review research that relies solely on publicly available information with no potential for harm to individuals. No human participants were recruited, no clinical interventions were administered, and no informed consent was required.

Our Bayesian analysis revealed substantial potential for AI-driven gross cost savings across both health care systems, with total projected annual gross cost savings of US $948.8 million for the United Kingdom and US $737.1 million for Australia. Table 1 presents the comprehensive posterior estimates for all sectors and countries. The prior-to-posterior evolution of each component for each sector and country is shown in Multimedia Appendix 4.

In the United Kingdom, workforce optimization generated the largest projected savings at US $591.9 million annually (95% CrI US $398.8-US $854.8 million), accounting for 62.4% of total projected UK savings. This result reflects the considerable potential for AI-driven efficiency gains in staff scheduling, workload distribution, and human resource management within the NHS. Workflow optimization contributed US $213.1 million (95% CrI US $133.1-US $300.6 million), representing savings from administrative and clinical process improvement. Radiology contributed US $143.9 million (95% CrI US $111.4-US $169.3 million), despite accounting for a smaller share of total health care expenditure.

Australia showed the same sector ordering but lower absolute savings overall. Workforce optimization generated US $561.4 million annually (95% CrI US $331.9-US $815.3 million), accounting for 76.2% of total projected Australian savings. Workflow optimization contributed US $95.2 million (95% CrI US $62.4-US $135.8 million), and radiology contributed US $80.4 million (95% CrI US $65.3-US $94.9 million). The shared ranking—workforce, workflow, then radiology—suggests that the main cross-country difference lies in scale rather than in the location of savings. At the same time, the CrIs overlap across several sector estimates, so those differences should still be interpreted probabilistically rather than definitively.

Bayesian updating shifted several parameters in meaningful ways relative to their literature-based priors. Figures 1 and 2 illustrate the movement from prior to posterior distributions for the United Kingdom and Australia, respectively, and Table S2 in Multimedia Appendix 2 reports the corresponding numerical estimates.

The clearest change appears in implementation risk. Posterior implementation risk fell below the prior mean in every sector-country pair, indicating that the external failure-rate anchors were more favorable than the literature-derived beta priors. In UK radiology, implementation risk dropped from a prior mean of 57.1% to a posterior mean of 35.5% (95% CrI 24.5%-46.5%), pulled strongly toward the NHS deployment-failure anchor of 0.348 [37]. The remaining 5 sector-country pairs also moved downward, though more modestly: UK workflow fell from 54.5% to 46.8%, UK workforce from 49.8% to 46.4%, Australian radiology from 57.2% to 47.3%, Australian workflow from 55.8% to 47.4%, and Australian workforce from 49.1% to 46.7%. In those cases, the posterior was pulled toward the pooled failure-rate estimate of 0.45 [39,40]. Overall, posterior implementation-risk estimates ranged from 35.5% to 47.4%, uniformly below the prior range of 49.1% to 57.2%. This pattern suggests that the empirical deployment evidence was somewhat more favorable than our conservative prior assumptions, even though implementation risk remained substantial in all sectors and continued to constrain realized savings.

Adoption followed a different pattern. UK radiology adoption rose modestly from 13.6% to 17.0%, consistent with growing uptake of AI tools in chest imaging and mammography. In every other sector-country pair, however, posterior adoption fell below its prior mean. The declines were especially pronounced in workforce optimization, where posterior adoption fell from 13.8% to 3.9% in the United Kingdom and from 16.1% to 10.7% in Australia. These low estimates suggest that organizational barriers to implementation in human resource management are much stronger than national strategy documents alone would imply.

Effectiveness estimates also moved, mainly through shrinkage toward the auxiliary observation values. UK radiology effectiveness was anchored at 43% relative to a prior mean of 43.1%, workflow optimization at 52% relative to 55%, and workforce optimization at 26.8% relative to 30.1%. Australian radiology shifted upward from 36.8% to 42.6%, reflecting the influence of a randomized controlled trial-derived effectiveness anchor. Posterior SDs for effectiveness were also much narrower than prior SDs—for example, in radiology, the SD fell from 10.3% to 2%—showing that the augmented likelihood sharpened those parameter estimates as intended.

Annual scenario analysis showed that projected savings are highly sensitive to implementation conditions. Across scenarios, total projected savings ranged from $356.6 million to $1.845 billion in the UK and from $267.2 million to $1.454 billion in Australia. Table 2 summarizes the sector-by-sector results. Under the optimistic scenario, projected annual savings rose to $1.845 billion in the UK, a 95% increase over the baseline estimate of about $0.949 billion. Australia reached $1.454 billion, a 97% increase over its baseline estimate of about $0.737 billion. These outcomes assumed lower implementation risk, stronger adoption, and effectiveness at the upper quartile of the posterior distributions. In both countries, workforce optimization remained the dominant contributor under these favorable conditions. The optimistic case also makes clear that gains are driven less by any single technology than by system-level conditions—especially adoption and implementation success—that lift all sectors simultaneously.

Under the pessimistic scenario, projected savings fell sharply to US $356.6 million in the United Kingdom and US $267.2 million in Australia. This case assumed implementation risk above 65% across sectors, adoption below 8%, and effectiveness at the lower quartile. The contrast between optimistic and pessimistic scenarios highlights the importance of organizational readiness and implementation success. In both the centralized UK system and the federated Australian system, higher risk and weaker adoption compress projected savings dramatically.

The conservative scenario yielded US $744.5 million for the United Kingdom and US $574.6 million for Australia. These estimates may offer the most plausible short-run planning values because they reflect moderate implementation friction rather than best-case or worst-case assumptions. Even so, the corresponding 95% CrIs remained wide, so they are better treated as planning ranges than as precise forecasts—particularly when comparing United Kingdom and Australian totals. Figure 3 illustrates the resulting sectoral distributions, and Tables S1 and S2 in Multimedia Appendix 3 report the annual scenario results together with the separate 2024-2030 temporal scenario analysis.

Emerging studies suggest that integrating AI across multiple domains—such as radiology, administrative workflows, and workforce planning—may generate meaningful synergies [18,19]. Although our savings function could accommodate such cross-sector effects, we were unable to quantify them because the empirical evidence remains too limited. Even so, the sectoral pattern offers a clear strategic signal. Workforce optimization dominates projected cost savings across scenarios in both countries, contributing 62.4% of UK total cost savings and 76.2% of Australian total cost savings in the baseline case. That pattern suggests that human resource management may be the highest-value target for early AI investment, even though current policy and research attention remain more heavily concentrated on clinical AI applications.

Radiology remains strategically important despite its smaller aggregate savings. In the United Kingdom, it showed the largest prior-to-posterior decline in implementation risk because we were able to anchor the model to direct NHS deployment evidence [37,38]. It may therefore serve as a practical entry point for implementation learning and provide institutional experience that can support later expansion. Workflow optimization also remained economically important in both countries, ranking second in projected savings despite the organizational complexity of administrative and clinical process redesign.

Our analysis reveals that AI implementation in the United Kingdom and Australian health care systems has substantial economic potential, with projected annual gross cost savings of US $949 million and US $737 million, respectively. However, 3 critical findings challenge conventional wisdom about health care AI economics. First, workforce optimization dominates potential cost savings across both countries, contributing 62.4% (United Kingdom) and 76.2% (Australia) of total projected benefits despite receiving less attention in AI implementation literature compared to clinical applications. Second, posterior implementation risk falls below prior expectations in every sector (posterior range 35.5%-47.4% vs prior means 49.1%-57.2%) because empirical deployment failure rates are more favorable than the conservative literature-derived beta priors; nevertheless, implementation risk remains the binding constraint on realizing these savings. Third, the substantial variation between pessimistic (US $0.357 and US $0.267 billion for the United Kingdom and Australia, respectively) and optimistic (US $1.845 and US $1.454 billion for the United Kingdom and Australia, respectively) scenarios demonstrates that organizational and implementation factors, rather than technological capabilities, primarily determine AI’s economic impact.

The dominance of workforce optimization cost savings challenges the prevailing focus on clinical AI applications in health care technology assessment. While diagnostic AI in radiology has received extensive research attention and regulatory development, our findings suggest that administrative and human resource applications may offer superior return on investment. This pattern likely reflects substantial inefficiencies in current health care workforce management, where manual scheduling, suboptimal workload distribution, and reactive staffing decisions create opportunities for AI-driven optimization that exceed the efficiency gains available in already streamlined clinical processes.

Several limitations shape how these findings should be interpreted. First, although we defined sector cost categories to be mutually exclusive in order to minimize double counting, some residual overlap cannot be ruled out without more detailed cost-accounting data from individual health systems. Savings estimates should therefore be interpreted as upper bounds on nonoverlapping achievable gains. Second, the sector expenditure weights and observed savings values are constructed from heterogeneous published studies rather than direct measurements of national AI-attributable cost reductions. They represent informed approximations and should not be treated as empirical observations in the classical sense. A deeper consequence of this construction is that 3 latent parameters (A, E, R) are updated from a limited number of constructed savings anchors per sector-country pair—2 in radiology and 1 in workflow and workforce—so the model remains only partially identified and posterior distributions retain sensitivity to prior specifications. A prior sensitivity analysis (Tables S3 and S4 in Multimedia Appendix 3) demonstrates that sector rank orderings and order-of-magnitude estimates are robust to ±20% perturbations of individual prior means and to alternative σ scaling factors, but the fundamental identifiability constraint cannot be resolved without larger, sector-specific implementation datasets.

Third, we excluded implementation costs such as software licensing, integration, training, governance, change management, and cybersecurity. Health systems should deduct these costs when translating our projections into business cases. Fourth, most model parameters were derived from published studies conducted in academic medical centers and large hospital networks, which may not represent community hospitals, rural facilities, or lower-resourced settings; transferability across institutions and geographies therefore remains uncertain. Fifth, the model treats adoption rate and implementation risk as statistically independent multiplicative factors. In practice, higher-risk organizations may also adopt AI more slowly, creating a negative correlation that would further reduce savings in high-risk settings. This independence assumption likely introduces a modest upward bias into projected savings. Sixth, several data quality caveats apply collectively: sector expenditure shares are derived from aggregate national health accounts and remain subject to classification uncertainty; cross-country US $ comparisons depend on spot exchange rates; and estimation noise is modeled via sector-specific heteroscedastic priors. Table S3 in Multimedia Appendix 3 reports the sensitivity of posterior savings to the σ scaling factor. Seventh, the model does not capture downstream clinical costs that AI deployment may trigger. For example, AI-assisted radiology screening may increase recall rates or biopsy referrals before sensitivity and specificity stabilize in operational settings, generating downstream costs not reflected in our savings estimates. Likewise, workflow AI that accelerates throughput may shift costs to downstream services. Quantifying such second-order effects will require health system-level cost-accounting data linked to AI deployment records.

Future work could strengthen the framework in several ways. Longitudinal implementation datasets would allow better calibration of adoption curves and risk decay. Natural experiments and richer cost-accounting data could help quantify cross-sector synergies and validate realized savings. The Bayesian framework can be adapted as a decision-support tool for health system planners by parameterizing priors from organization-specific readiness assessments (eg, AI maturity scores, existing digital infrastructure, and workforce capacity) and updating with local deployment data as they accumulate. This would allow progressive narrowing of CrIs over time, transforming the current national-level projections into institution-specific investment guidance. Finally, linking economic outcomes to patient outcomes would provide a fuller assessment of value.

AI could generate substantial expenditure reductions in both the United Kingdom and Australian health care systems, with projected annual gross cost savings of US $949 million and US $737 million, respectively. In both countries, workforce optimization offers the largest potential gains, while implementation risk remains the main obstacle to realizing them. These findings suggest that policymakers should place as much weight on implementation science, governance, and organizational change as on technology acquisition itself. Health systems may also need to rebalance attention toward workforce optimization, develop stronger implementation-risk assessment frameworks, and create mechanisms for shared learning across centralized and federated systems. The Bayesian framework we present is best understood as a decision-support tool that provides probabilistic planning ranges under uncertainty rather than deterministic forecasts.