Can AI Lower the Cost of Care?
Australia spent an estimated $270.5 billion on health in 2023-24. The largest estimate we found of what AI could save is a projection, and it expects proportionally deeper cuts in administration than in clinical care. On the clinical side, the evidence we found shows tools that find more disease can add follow-up work, and a gain seen in trials did not survive one real-world rollout.
On this page
Better: Note and confidence
Moderate confidenceTrial gains do not always survive deploymentIn a pragmatic rollout at Stanford, AI-assisted colonoscopy did not lift the adenoma detection rate (40.1% with the tool against 41.8% before), although randomised trials have reported gains.Faster: Note and confidence
Low confidenceWe found few measured time savings outside documentationIn a randomised study of 52 US primary care physicians, AI-drafted replies to patient messages increased reading time by 21.8% and did not significantly change reply time.Cheaper: Note and confidence
Low confidencePlausible on administration; unproven on clinical costThe largest estimate we found (5 to 10% of US health spending) is a projection by authors including three from McKinsey, and excludes one-time build costs. A review of 117 economic evaluations found only 28% counted implementation costs.Safer: Note and confidence
Governance-gatedOnly where costs, benefits and risks are trackedThe national auditor found the federal health department had not analysed the costs of its Medicare fraud detection system and had only an unclear analysis of its benefits, and recommended the department work with other agencies and the sector to assess AI risks in healthcare.The problem, in numbers
Australia spent an estimated $270.5 billion on health in 2023-24, about 10.1% of national economic activity, and governments funded about 69.6% of it.Source 1 Hospitals took $113.8 billion. The AIHW figures record $7.7 billion as administration, which is about 2.8% of the total by our calculation, but the AIHW page does not say what that line includes.Source 1 In our reading, paperwork done by clinicians and hospital staff is likely to sit inside hospital and clinical budgets rather than in that line.
US research shows how large administration can be when insurer overhead and provider administration are both counted. A peer-reviewed estimate, built from government reports, provider accounting data and surveys, put administration at 34.2% of US national health expenditure in 2017, against 17.0% in Canada.Source 4 The authors note the two countries' accounting categories differ somewhat. In our reading, those are two very different insurance systems, not a benchmark for Australia.
The Productivity Commission estimated in 2024 that better use of digital technology, not AI alone, could save more than $5 billion a year. It also cited research, which the pages do not name, that up to 30% of health workforce tasks could be automated with digital technology and AI.Source 2 Both are estimates.
What improves, and why
Administrative work is repetitive and rule-bound: scheduling, claims, correspondence, coding. In our reading, that makes it easier to automate and easier to check than a clinical decision, because the right answer is usually known.
The largest published estimate we found makes the same split. Researchers, three of them from McKinsey, projected in an NBER paper that wider AI use could save 5 to 10% of US health spending, roughly US$200 billion to US$360 billion a year in 2019 dollars.Source 5 For hospitals, they estimated AI could cut roughly 9 to 19% of administrative costs but only 4 to 8% of medical costs. Because medical costs are four times larger, about 60% of the projected hospital savings still came from the medical side.Source 5 The rates are the authors' estimates "based on our experience", not measured results; they assume adoption at full scale, and they leave out one-time implementation costs that the authors put at 1.0 to 1.5 times a year's net savings.Source 5
Ambient scribes, which listen to a consultation and draft the clinical note, are covered in our review of clinical documentation time.
A tool that finds more disease does not lower the cost of care by itself. Every extra finding needs someone to act on it.
What the evidence actually says
Administrative tools: measured, and mixed. Outside scribes, we found fewer measured results than projections. In a randomised study of 52 US primary care physicians, run from June to August 2023, AI-drafted replies to patient messages increased the time spent reading messages by 21.8%, made replies 17.9% longer, and did not significantly change the time taken to reply.Source 6 The authors report the physicians recognised the drafts' value, but in that study the drafts did not measurably save time.
Clinical tools: more detection, more downstream work. Computer-aided polyp detection in colonoscopy shows how detection can raise cost. Applying published detection performance to 3,735 patients, a simulation found 19.1% would be due for an earlier repeat colonoscopy under US guidelines, or 1.9% under European ones. The authors concluded wider use "would inevitably increase the demand" for surveillance colonoscopies.Source 7
Finding more early disease can still pay. A US microsimulation, a model rather than a measured result, found that if the tool raised the adenoma detection rate from 26% to 37% at US$75 a procedure, it would mean fewer cancers and a lower lifetime cost. But it stayed cost-effective only if detection rose to at least 30%.Source 7 When Stanford put the tool into routine use with a deliberately minimal rollout, the measured detection rate was 40.1% with it against 41.8% before, with no significant difference. The lead author discloses consulting for Medtronic, which makes the system studied.Source 7 In our reading, the modelled saving rests on a gain that this real-world rollout did not deliver.
Economic evaluations: optimistic, and incomplete. A 2026 peer-reviewed review of 117 health economic evaluations of medical AI found most reported savings or cost-effectiveness, but only 28% counted implementation costs, and 63% assessed tools at readiness level 4 to 5 on a 9-point scale. The authors judge the evidence limited by suboptimal reporting, bias and immature technologies.Source 5
Share of evaluations, from the review abstract. Most reported savings, yet fewer than a third counted the cost of putting the tool in place.
Where it falls short
Projections are not savings. The largest estimate we found excludes one-time build costs and rests on the authors' own judgement of savings rates.Source 5
Detection creates work. More findings can mean more follow-up procedures, and the cost case depends on gains that one real-world rollout did not reproduce.Source 7
Flags need people. The federal health department stopped using an AI fraud model when staff resourcing for manual assessment could not keep up with the volume it generated, and switched to a simpler model that, it said, gave direct control over that volume.Source 8
Prices follow costs slowly. The 2026-27 national efficient price of $7,418 per weighted unit is built on 2023-24 hospital cost data.Source 3 In our reading, any hospital saving from AI would reach that price only years later.
Where governance decides the outcome
The audit evidence shows how little is tracked. The national auditor found that the federal health department ran an AI-enabled model in its Medicare fraud detection system from July 2024 to December 2025. A preliminary analysis found it accurately identified matters 22% of the time (18 of the 82 flagged matters staff reviewed), which the auditor said showed the model was efficacious. The department then replaced it with a non-AI model, without assessing the benefits and costs of that decision, and had not analysed the cost of building the system.Source 8 The auditor also recommended that the department work with others to assess the risks of AI in healthcare, including providers using AI in Medicare claiming, and the department agreed.Source 8 In our reading, a health service that cannot say what a tool costs cannot say whether it is cheaper, or whether it is safe.
How to think about adoption
Start with administration, where tasks are repeated and the right answer is known, and measure the cost per task before and after rather than relying on projections. Count the build, the licence, the training and the staff time spent checking output, since most published evaluations leave some of these out.Source 5 For clinical tools, model the downstream work as well as the detection gain: every extra finding needs someone to act on it.Source 7,Source 8 Plan the rollout, not just the purchase, because a tool that worked in a trial may not work the same way on your wards.Source 7 And treat a clinical cost saving as a hypothesis until a local evaluation shows it.
BFCS.ai does not fabricate figures. Spending figures here are agency estimates, the US savings figure is a projection by consultancy-affiliated authors, the surveillance figure is a simulation, and the cost-effectiveness result is a model. We found no measured, independent study showing that AI lowered the total cost of care in an Australian health service, and we say so where it matters.
Sources
- 1
Australian Institute of Health and Welfare, "Health expenditure Australia 2023-24," summary (last updated 29 October 2025): aihw.gov.au
Same report, "Other services": aihw.gov.au
Back to textWhat we took from it
total health spending an estimated $270.5 billion, or $10,037 per person, around 10.1% of the nation's total economic activity; real growth of 1.1% from 2022-23; hospitals $113.8 billion; governments funded about 69.6% ($106.2 billion Australian Government, $82.0 billion states and territories).
$7.7 billion on administration in 2023-24, within $20.0 billion of other services (with patient transport and aids and appliances). The page does not define what the administration line includes. The shares of total spending for hospitals (about 42%) and administration (about 2.8%) are our calculation from these figures.
- 2
Productivity Commission, "Leveraging digital technology in healthcare," research paper and media release, 7 May 2024: pc.gov.au
and pc.gov.au
Back to textWhat we took from it
better integrating digital technology into everyday practice could save more than $5 billion a year; better use of electronic medical record data could save up to $5.4 billion a year by reducing hospital stays; fewer duplicated tests up to $355 million in the public hospital system; telehealth consumer travel gains around $895 million a year; research shows up to 30% of tasks currently undertaken by the healthcare workforce could be automated using digital technology and AI (the pages do not name that research). These are estimates of digital technology broadly, not AI alone.
- 3
Independent Health and Aged Care Pricing Authority, "Understanding the NEP and NEC Determinations 2026-27" (March 2026): ihacpa.gov.au
Back to textWhat we took from it
the national efficient price for 2026-27 is $7,418 per national weighted activity unit (NWAU); the NEP is based on the average cost of an admitted acute episode of care in public hospitals, and the "average" hospital service is worth one NWAU; the determination used 2023-24 activity and cost data.
- 4
Himmelstein, D. U., Campbell, T. & Woolhandler, S., "Health Care Administrative Costs in the United States and Canada, 2017," Annals of Internal Medicine 172(2), 2020, pp. 134-142 (peer-reviewed; abstract read via the Europe PMC record of PubMed 31905376). pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
Based on government reports, accounting data providers file with regulators, physician surveys and census employment data. US insurers and providers spent $812 billion on administration in 2017, $2,497 per capita, 34.2% of national health expenditures; Canada $551 per capita, 17.0%. The authors note accounting categories for the two countries differ somewhat. A correction was published (Annals of Internal Medicine 173(5): 415, 1 September 2020); the publisher refused automated access, so we could not read what it changed.
- 5
Sahni, N., Stein, G., Zemmel, R. & Cutler, D. M., "The Potential Impact of Artificial Intelligence on Healthcare Spending," NBER Working Paper 30857, January 2023, revised October 2023; also a chapter in The Economics of Artificial Intelligence: nber.org
Godoy Junior, C. A., Boverhof, B.-J., Rutten-van Mölken, M. P. M. H. et al., "Technological Maturity and Cost-Effectiveness of Medical Artificial Intelligence: A Systematic Review of Health Economic Evaluations," Value in Health 29(5), 2026 (peer-reviewed; abstract only read). ispor.org
Back to textWhat we took from it
Health Care Challenges (2024). Sahni, Stein and Zemmel list McKinsey & Company affiliations. Estimate: wider AI adoption within five years using current technology could save 5 to 10% of US healthcare spending, roughly $200 billion to $360 billion a year in 2019 dollars; gross savings rates per domain are estimated "based on our experience", the authors note there are few experimental studies of AI's impact on costs to draw on, and the estimates assume adoption reaches full scale. For hospitals (about $1,096 billion in costs, 20% administrative), net savings of $60 billion to $120 billion, about 40% from administrative costs (roughly 9 to 19% of that category) and 60% from medical costs (roughly 4 to 8% of that category). The estimates do not include one-time implementation costs, which "in our experience are 1.0 to 1.5 times the annual net savings", nor new core technology or last-mile change management.
117 studies; the majority reported cost savings or cost-effectiveness; 28% included implementation costs and 57% operational costs; 63% evaluated tools at Clinical Machine Learning Readiness Level 4 to 5 (on a scale of 1 to 9); the authors conclude current evaluations are limited by suboptimal reporting quality, risk of bias and immature technologies. Bar widths in the figure are scaled to the largest value (63% = full width); our calculation.
- 6
Tai-Seale, M. et al., "AI-Generated Draft Replies Integrated Into Health Records and Physicians' Electronic Communication," JAMA Network Open 7(4): pubmed.ncbi.nlm.nih.gov
Back to textWhat we took from it
e246565, 2024 (peer-reviewed; abstract read via PubMed). Randomised waiting-list quality improvement study at a US academic health system, June to August 2023: 52 primary care physicians (25 immediate, 27 delayed activation) and a contemporary control group of 70. Drafts were associated with a 21.8% increase in read time (95% CI 5.2% to 41.0%), a 5.9% decrease in reply time that was not statistically significant (95% CI -16.6% to 6.2%), and a 17.9% increase in reply length (95% CI 10.1% to 26.2%); participants "recognized GenAI's value" and suggested improvements.
- 7
Three peer-reviewed studies of computer-aided polyp detection (CADe) in colonoscopy. pubmed.ncbi.nlm.nih.gov
Thiruvengadam, N. R. et al., "An Evaluation of Critical Factors for the Cost-Effectiveness of Real-Time Computer-Aided Detection: Sensitivity and Threshold Analyses Using a Microsimulation Model," Gastroenterology 164(6), 2023, pp. 906-920 (abstract read via the Europe PMC record of PubMed 36736437): pubmed.ncbi.nlm.nih.gov
Ladabaum, U. et al., "Computer-aided Detection of Polyps Does Not Improve Colonoscopist Performance in a Pragmatic Implementation Trial," Gastroenterology 164(3), 2023, pp. 481-483.e6 (short report, full text via PubMed Central; all authors Stanford University or Stanford Health Care; the lead author discloses advisory and consulting roles, including consulting for Medtronic, maker of the GI Genius system studied). pmc.ncbi.nlm.nih.gov
Back to textWhat we took from it
Lui, T. K. L. et al., "The Impacts of Computer-Aided Detection of Colorectal Polyps on Subsequent Colonoscopy Surveillance Intervals: Simulation Study," Journal of Medical Internet Research 25: e42665, 2023 (abstract read via the Europe PMC record of PubMed 36763451): 3,735 patients colonoscoped 2016 to 2020; applying published CADe performance, 19.1% (n=714) would have shorter surveillance intervals under US (USMSTF) guidelines and 1.9% (n=71) under European (ESGE) guidelines; "widespread adoption of CADe would inevitably increase the demand for surveillance colonoscopies."
modelled CAD at USD 75 a procedure raising the adenoma detection rate from 26% to 37%, giving 79 fewer lifetime colorectal cancers and 34 fewer deaths per 10,000 people and a cost saving of USD 143 (dominant); CAD must raise the rate from 26% to at least 30%, or cost less than USD 579 a colonoscopy, to be cost-effective at a willingness-to-pay threshold of USD 100,000 per quality-adjusted life year.
Retrospective pragmatic trial with historical and concurrent controls, three-month implementation, February to May 2022: CADe used in 1,008 of 1,037 eligible colonoscopies; adenoma detection rate 40.1% during implementation against 41.8% before (P = .44); difference-in-difference odds ratio 1.14 (95% CI 0.83 to 1.56). The authors say the results contrast sharply with randomised trials, offer several possible explanations (trial endoscopists knew the hypotheses; prompts may have been dismissed; mucosal exposure may have degraded), note they used a minimalist deployment strategy by design, and say such a strategy may not ensure success.
- 8
Australian National Audit Office, "Artificial Intelligence and Medicare Benefits Integrity," Auditor-General Report No. 5 of 2026-27, Department of Health, Disability and Ageing, letter of transmittal dated 21 September 2026. anao.gov.au
(report PDF: anao.gov.au
Back to textWhat we took from it
Arrangements to manage AI at enterprise level largely effective; arrangements for managing AI use to support health provider compliance partly effective. An AI-enabled model within the MBS fraud detection system was deployed from July 2024 to December 2025; a preliminary August 2025 analysis showed it accurately identified matters 22% of the time (paragraph 5.11; Table 5.1: 18 true positives of 82 potential matters reviewed, out of 597 identified; the department said true positive rates should not be compared across models); the auditor describes the ad hoc preliminary analysis as showing the model was efficacious (paragraph 19); eight potential non-compliance or fraud matters were referred using, in part, the model. The model was not used after December 2025, when it was determined that resourcing for manual assessment could not keep up with the volume of potential matters it generated (paragraph 19); the department advised it was replaced by a non-AI logistic regression model that also performed effectively and allowed direct control over the volume selected for manual assessment (paragraph 5.13), with no assessment of the benefits and costs of that decision. There has been no analysis of the costs of developing the fraud detection system, and an unclear analysis of benefits and net returns (paragraph 5.6); policies do not require cost-benefit analysis for AI use cases. Recommendation 1 (agreed): identify and assess the risks of AI in healthcare settings, including providers using AI in MBS claiming.