[Healthcare](https://bfcs.ai/industry/healthcare/) · [Faster](https://bfcs.ai/faster/)

# Can AI Improve Patient Triage?

AI now sits at the front door of care, in online symptom checkers, nurse helplines and emergency department triage. The best evidence of speed comes from one vendor-funded emergency department study. Consumer tools still misjudge urgency in bench tests, and the open governance questions are calibration and equity.

Updated **October 2026** · 7 min read · Evidence-graded · 8 sources

## The BFCS verdict

Every article scored on the same four dimensions

| Dimension | Verdict | Why |
| --- | --- | --- |
| Better | Partly · Low to moderate confidence | Decision support helps nurses; consumer tools misjudge urgency · In a vendor-funded US study, the share of critical-care patients identified as high acuity rose from 78.8% to 83.1% after triage decision support was introduced. Symptom checkers gave correct triage advice in 49% of Australian vignette tests, and the median app in a follow-up by German researchers got 55.8% right. |
| Faster | Partly · Low confidence | Faster first contact in one ED study; little else measured · One vendor-funded quality improvement study across three emergency departments measured median time to the first care area falling from 12.0 to 8.0 minutes (abstract only read). We found no measured evidence that consumer symptom checkers shorten time to appropriate care. |
| Cheaper | Unproven · Low confidence | No cost evaluation found · We found no published cost evaluation of AI triage in the sources we read. Re-sorting patients towards lower acuity levels may free resources, but that is our inference, not a measured saving. |
| Safer | Conditional · Governance-gated | Safe only with calibration and equity checks · Symptom checkers missed more than 40% of emergencies in one evaluation. Nine language models changed recommendations for identical emergency cases by sociodemographic label. Ahpra asks practitioners to understand bias in AI, particularly for Aboriginal and Torres Strait Islander patients and patients from diverse backgrounds. |

## The problem, in numbers

Triage decides who is seen first, and the queue is long. In 2024-25, 67% of Australian emergency department patients were seen within the time their triage category allows, down from 71% in 2020-21. One in ten waited longer than 1 hour and 58 minutes. (Source 1)

Human triage is not a perfect baseline either. In a US study of 5.3 million visits at 21 emergency departments, triage using the Emergency Severity Index mistriaged 32.2% of encounters, measured against what care patients went on to need: 3.3% were undertriaged and 28.9% overtriaged. Black patients had an 18.5% greater relative risk of undertriage than White patients after adjustment. (Source 2) That is a US scale, not the Australasian one, but it shows the two errors any AI tool must be calibrated against.

This review asks whether AI moves patients to the right care sooner. Our review of [diagnosis](https://bfcs.ai/articles/can-ai-make-diagnosis-better) covers image reading; this one covers routing.

### The evidence, in four numbers

8.0 minutes · median time to the first care area, down from 12.0, after AI triage support · measured; vendor-funded quality improvement study, 174,648 visits at three US emergency departments, abstract only read (Taylor et al., NEJM AI, 2025)

32.2% · of encounters mistriaged under the US Emergency Severity Index, without AI · measured by an outcome-based algorithm; 3.3% undertriaged and 28.9% overtriaged across 5.3 million visits (Sax et al., JAMA Network Open, 2023)

49% · of triage tests answered correctly by symptom checkers available in Australia · measured on 48 written vignettes across 19 triage tools, not real patients; abstract only read (Hill et al., MJA, 2020)

Under 44.2% · of cases in which people using a language model chose the right next step, no better than control · measured; randomised preregistered study of 1,298 UK participants on ten scenarios; abstract read, with recruitment checked in the full text (Bean et al., Nature Medicine, 2026)

## What improves, and why

AI appears in three places at the front door: consumer symptom checkers, decision support for nurses on telehealth lines, and prediction tools at the emergency department triage desk. In our reading, the speed gain comes from sorting more patients correctly at first contact, so that fewer sit in the uncertain middle category.

The strongest measured evidence is from the emergency department. In a quality improvement study across three US emergency departments and 174,648 visits, the share of visits sorted into the middle acuity level fell from 48.8% to 39.7% after an AI triage tool was introduced, and low-acuity visits rose from 23.9% to 35.4%. Median time from arrival to the first care area fell from 12.0 to 8.0 minutes, and time to departure from 311.0 to 292.0 minutes. (Source 4) Patients who needed critical care were identified as high acuity in 83.1% of cases, against 78.8% before. (Source 4) The study was vendor-funded: Beckman Coulter, which sells emergency department triage decision support, paid for it, three authors list affiliations with it, it was not randomised, and we read the abstract only. (Source 4) It builds on an earlier retrospective study by two of the same authors, who called for prospective validation. (Source 4)

In Australia, Healthdirect says a decision support system from the vendor Infermedica, using a probabilistic model, supports its helpline nurses and powers its online Symptom Checker; it was piloted on the helpline in one state in 2022 and scaled nationally in 2023. (Source 7) Healthdirect says it was tested with Macquarie University before deployment. A conference paper by Macquarie researchers describes the method: a bench test on 78 written vignettes, compared with the previous system. (Source 7) We found no published measure of whether it shortens the time to appropriate care.

> Triage speed is only worth having if the sorting is right. A fast wrong answer moves a patient to the wrong queue sooner.

## What the evidence actually says

Consumer tools are the weak link. Symptom checkers available in Australia gave correct triage advice in 49% of tests on 48 written vignettes, and only 30% for non-urgent cases; the authors found them risk-averse. (Source 3) A follow-up by German researchers found median app accuracy of 55.8% in 2020, against 59.1% five years earlier. The apps had become less cautious, missed more than 40% of emergencies, and none outperformed laypeople on both the emergency and the self-care decision. (Source 3) A systematic review of ten studies found triage accuracy ranging from 48.8% to 90.1%. (Source 3)

Language models look better on paper than in use. Researchers at Flinders University found three frontier models of 2024 chose correctly between urgent and non-urgent care in 92.4% of 144 decisions on 48 synthetic vignettes, close to physicians in an earlier study, mostly erring towards urgency. (Source 5) But in a randomised study of 1,298 UK participants, people using the same class of models picked the right next step in fewer than 44.2% of cases, no better than a control group using their usual sources. (Source 5) The authors point to the interaction between person and model as the weak point, not the model's medical knowledge.

Equity evidence is mostly indirect. In the US study of human triage, the gap was widest for Black men, who had a 41.0% greater relative risk of undertriage than White women. (Source 2) For AI, the clearest test we found presented nine language models with 1,000 emergency department cases, each rewritten in 31 sociodemographic versions with identical clinical details. Some LGBTQIA+ subgroups were recommended mental health assessment about six to seven times more often than clinically indicated, and high-income cases were recommended more advanced imaging. (Source 6) We found no published equity analysis of an AI triage tool in Australian use.

## Where it falls short

-   **Undertriage is the dangerous error.** Symptom checkers in one evaluation missed more than 40% of emergency vignettes. (Source 3)
    
-   **Bias can be learned.** Across nine language models and 1,000 emergency cases, identical clinical details labelled Black, unhoused or LGBTQIA+ were more often sent to urgent care, invasive interventions or mental health evaluation. (Source 6)
    
-   **The speed evidence is thin.** The measured gain comes from one vendor-funded, non-randomised study, read here in abstract only. (Source 4)
    
-   **Benchmarks mislead.** Models scored well alone but did not help real people decide better. (Source 5)
    

## Where governance decides the outcome

Calibration and equity are the two questions to settle before deployment. Calibration means choosing, and publishing, the balance between undertriage and overtriage. In our reading, the symptom checker evidence shows both failure modes: risk-averse advice in one evaluation, missed emergencies in another. (Source 3) Equity means testing outcomes by patient group, since human triage already shows gaps by race and models can widen them. (Source 2), (Source 6) Ahpra says practitioners should understand the bias that can exist in the data and algorithms used in AI, linking this to the health and safety of Aboriginal and Torres Strait Islander people and patients from diverse backgrounds. (Source 8) In the US study, nurses who agreed closely with the tool outperformed it alone, which in our reading supports decision support over automation. (Source 4)

## How to think about adoption

Start with decision support for trained triage staff, not unsupervised consumer advice. Measure before and after with outcome-based definitions of undertriage and overtriage, and report them by age, sex, First Nations status and language. Test any language model with real users, not only vignettes. Ask vendors for independent evaluation, since the speed evidence here is vendor-funded. In our reading, AI can move the first contact faster, but the evidence that it moves patients to the right care sooner is still early.

## How we read this evidence

BFCS.ai does not fabricate figures. The only measured speed gain comes from one vendor-funded, non-randomised study read in abstract only. Symptom checker and language model accuracy figures come from written vignettes, not real patients, and should not be read as real-world triage performance. Several sources were read in abstract only, as marked. We found no Australian trial of AI triage outcomes and no published cost evaluation, and the TGA's website could not be reached during drafting, so its guidance is not cited.

## Sources

1.  Australian Institute of Health and Welfare, "Waiting times overview," Emergency department care, 2024-25 data (page last updated 10 December 2025). [aihw.gov.au](https://www.aihw.gov.au/hospitals/topics/emergency-departments/waiting-times)
    
    **What we took from it:** 67% of patients were seen on time for their triage category in 2024-25, the same as the previous year and down from 71% in 2020-21; 50th percentile wait 18 minutes; 90th percentile wait 1 hour and 58 minutes. Triage categories and on-time targets: Resuscitation (immediate), Emergency (within 10 minutes), Urgent (30 minutes), Semi-urgent (60 minutes) and Non-urgent (120 minutes).
    
2.  Sax, D. R., Warton, E. M., Mark, D. G. et al. (Kaiser Permanente CREST Network), "Evaluation of the Emergency Severity Index in US Emergency Departments for the Rate of Mistriage," JAMA Network Open 6(3), e233404, March 2023 (peer-reviewed; full text read via PubMed Central). [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC10024207/)
    
    **What we took from it:** Retrospective cohort of 5,315,176 adult encounters at 21 Kaiser Permanente Northern California emergency departments, 2016 to 2020, using ESI version 4 (nurse triage without AI). Mistriage was measured by an outcome-based algorithm derived through expert review and a modified Delphi process, validated against 200 records with 96% agreement: mistriage 32.2%, undertriage 3.3%, overtriage 28.9%; ESI sensitivity for high-acuity patients 65.9%. In adjusted analyses, Black patients had an 18.5% greater relative risk of undertriage and 4.6% greater relative risk of overtriage than White patients. Black male patients had a 41.0% greater relative risk of undertriage and 9.9% greater relative risk of overtriage than White female patients. Authors received grants from the Kaiser Permanente Lokahi Risk Reduction Program.
    
3.  Symptom checker evaluations. [ro.ecu.edu.au](https://ro.ecu.edu.au/ecuworkspost2013/8431)
    
    Schmieding, M. L., Kopka, M., Schmidt, K. et al., "Triage Accuracy of Symptom Checker Apps: 5-Year Follow-up Evaluation," Journal of Medical Internet Research 24(5), e31810, 10 May 2022 (peer-reviewed; full text read via PubMed Central): [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC9131144/)
    
    Wallace, W., Chan, C., Chidambaram, S. et al., "The diagnostic and triage accuracy of digital and online symptom checker tools: a systematic review," npj Digital Medicine 5, 118, August 2022 (peer-reviewed; abstract only read, via Europe PMC): [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/35977992/)
    
    **What we took from it:** Hill, M. G., Sim, M. and Mills, B., "The quality of diagnosis and triage advice provided by free online symptom checkers and apps in Australia," Medical Journal of Australia 212(11), 2020, pp. 514-519, doi:10.5694/mja2.50600 (peer-reviewed; abstract only read, via the Edith Cowan University repository): 36 symptom checkers tested with 48 vignettes; 19 triage tools gave correct advice in 338 of 688 tests (49%; 95% CI 44% to 54%); correct for 63% of emergency, 56% of urgent, 30% of non-urgent and 40% of self-care vignettes; the authors describe triage advice as risk-averse, often recommending more urgent care than appropriate.
    
    22 apps on 45 vignettes; median triage accuracy 55.8% in 2020 against 59.1% in 2015; the median app identified 51.9% of emergencies, and the abstract says apps missed more than 40% of emergencies; odds of overtriage to undertriage fell from 2.82:1 to 1.11:1; no app outperformed laypeople on both the emergency and the self-care decision. No funding was required; the lead author was employed in 2014 and 2015 by the company now called Ada Health.
    
    10 studies included; triage accuracy ranged from 48.8% to 90.1%; the authors judge accuracy variable and low.
    
4.  Taylor, R. A., Chmura, C., Hinson, J. et al., "Impact of Artificial Intelligence-Based Triage Decision Support on Emergency Department Care," NEJM AI 2(3), AIoa2400296, published 27 February 2025, doi:10.1056/AIoa2400296 (peer-reviewed; abstract only read, via Europe PMC). [beckmancoulter.com](https://www.beckmancoulter.com/about-beckman-coulter/newsroom/press-releases/2022/q4/2022-oct-11-beckman-coulter-acquires-artificial-intelligence-company)
    
    Vendor-funded, not randomised. [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/42932561/)
    
    Levin, S., Toerper, M., Hamrock, E. et al., "Machine-Learning-Based Electronic Triage More Accurately Differentiates Patients With Respect to Clinical Outcomes Compared With the Emergency Severity Index," Annals of Emergency Medicine 71(5), 2018, pp. 565-574.e2 (peer-reviewed; abstract only read, via Europe PMC): [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/28888332/)
    
    **What we took from it:** Quality improvement study at three US emergency departments comparing 180 days before and after implementation: 174,648 visits. Low-acuity visits rose from 23.9% to 35.4%, mid-acuity fell from 48.8% to 39.7%, high-acuity fell from 27.3% to 24.9%; identification of critical-care patients as high acuity rose from 78.8% to 83.1% (P<0.001); identification of emergency surgery need did not change significantly (41.2% to 39.2%); median time from arrival to initial care area fell from 12.0 to 8.0 minutes, to disposition from 190.0 to 182.0 minutes and to departure from 311.0 to 292.0 minutes; nurses with high agreement with the tool outperformed the AI alone, and those with lower agreement performed worse. "Funded by Beckman Coulter"; three authors list Beckman Coulter affiliations. The abstract does not name the tool. Beckman Coulter acquired StoCastic, LLC, maker of the TriageGo emergency department triage decision support tool, in October 2022 (Beckman Coulter press release, 11 October 2022,
    
    retrospective study of 172,726 visits at two emergency departments; more than 65% of visits were triaged to ESI level 3, and the model identified more than 10% (14,326) of level 3 patients for up-triage; the authors say further prospective validation is needed. Four authors list affiliation with StoCastic LLC; funded by AHRQ grant R21 HS023641.
    
5.  Sorich, M. J., Mangoni, A. A., Bacchi, S. et al., "The Triage and Diagnostic Accuracy of Frontier Large Language Models: Updated Comparison to Physician Performance," Journal of Medical Internet Research 26, e67409, 6 December 2024 (peer-reviewed research letter; full text read via PubMed Central): [pmc.ncbi.nlm.nih.gov](https://pmc.ncbi.nlm.nih.gov/articles/PMC11662182/)
    
    Bean, A. M., Payne, R. E., Parsons, G. et al., "Reliability of LLMs as medical assistants for the general public: a randomized preregistered study," Nature Medicine 32(2), 2026, pp. 609-615, published online 9 February 2026, doi:10.1038/s41591-025-04074-y (peer-reviewed; abstract read via Europe PMC, full text checked via PubMed Central for recruitment and the correction notice): [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/41663592/)
    
    **What we took from it:** Flinders University; 48 published vignettes with a two-way urgent or non-urgent triage decision; GPT-4o, Claude-3.5-Sonnet and Gemini-1.5-Pro combined 92.4% (133 of 144), against 91.3% for primary care physicians and 74.1% for lay internet users in the earlier Levine et al. study; the most common error was overestimating urgency; the authors name the small number of cases as the main limitation; the vignettes are synthetic.
    
    1,298 participants living in the UK, ten scenarios, assigned to GPT-4o, Llama 3, Command R+ or a source of their choice; tested alone, the models identified conditions in 94.9% of cases and the correct disposition in 56.3% on average; participants using them identified relevant conditions in fewer than 34.5% of cases and the correct disposition in fewer than 44.2%, no better than control. A publisher correction (Nat Med 32(5), 1937, May 2026, doi:10.1038/s41591-026-04404-8) fixed mislabelled y-axis values in Figure 2a; it does not change any figure cited here.
    
6.  Omar, M., Soffer, S., Agbareia, R. et al., "Sociodemographic biases in medical decision making by large language models," Nature Medicine 31(6), 2025, pp. 1873-1881, published online 7 April 2025, doi:10.1038/s41591-025-03626-6 (peer-reviewed; abstract only read, via Europe PMC). [pubmed.ncbi.nlm.nih.gov](https://pubmed.ncbi.nlm.nih.gov/40195448/)
    
    **What we took from it:** Nine language models, more than 1.7 million outputs from 1,000 emergency department cases (500 real, 500 synthetic), each presented in 31 sociodemographic versions and one control with identical clinical details. Cases labelled Black, unhoused or LGBTQIA+ were more often directed towards urgent care, invasive interventions or mental health evaluations; some LGBTQIA+ subgroups received mental health assessments about six to seven times more often than clinically indicated; high-income cases received more advanced imaging recommendations. The pattern held across proprietary and open-source models.
    
7.  Healthdirect Australia, "Quality and safety of healthdirect's triage service statement" (last reviewed July 2025). [healthdirect.gov.au](https://www.healthdirect.gov.au/quality-and-safety-of-healthdirects-triage-service-statement)
    
    Luckraj, N., Strazzari, R., Coiera, E. and Magrabi, F., "Assessing the safety of a new clinical decision support system for a national helpline," MEDINFO 2023 proceedings, Studies in Health Technology and Informatics 310, IOS Press, published 25 January 2024, pp. 514-518, doi:10.3233/SHTI231018 (conference paper; abstract and method read via Macquarie University research portal): [researchers.mq.edu.au](https://researchers.mq.edu.au/en/publications/assessing-the-safety-of-a-new-clinical-decision-support-system-fo/)
    
    **What we took from it:** Healthdirect is a clinical triage service by phone or the online Symptom Checker; the helpline is staffed by nurses; a clinical decision support system from the European vendor Infermedica supports the helpline nurses and powers the Symptom Checker; it uses a probabilistic model and is monitored by qualified professionals at Healthdirect and Infermedica; the system was piloted on the helpline in one state in 2022 and scaled nationally in 2023; pre-deployment testing was done with the Australian Institute of Health Innovation at Macquarie University, and an independent review is planned two years after implementation. Self-description by the service operator.
    
    method only is used here; the system used by nurses on the national consumer helpline was tested on 78 standardised vignettes (48 published, 30 proprietary) in two cycles using the vendor's online evaluation tool, and compared with the existing system on the same 47 vignettes.
    
8.  Australian Health Practitioner Regulation Agency and National Boards, "Meeting your professional obligations when using Artificial Intelligence in healthcare" (page reviewed 22 August 2024). [ahpra.gov.au](https://www.ahpra.gov.au/Resources/Artificial-Intelligence-in-healthcare.aspx)
    
    **What we took from it:** Used here only for its statement on bias: practitioners should understand the inherent bias that can exist within data and algorithms used in AI applications, which the page links to the health and safety of Aboriginal and Torres Strait Islander people and patients from diverse backgrounds.
    

Putting this to work

## Working through this in your own organisation?

BFCS.ai is published by Inhouse CX. We help small and mid-sized organisations decide where AI earns its place, and put the governance in place that makes it safe to use.

[Talk to Inhouse CX →](https://inhouse.cx/engage/)

Continue the thread

- [Can AI Make Diagnosis Better?](https://bfcs.ai/articles/can-ai-make-diagnosis-better/): More in Healthcare · HEALTHCARE · BETTER · 3 min · Read →
- [Can AI Cut Clinical Documentation Time?](https://bfcs.ai/articles/can-ai-cut-clinical-documentation-time/): More in Healthcare · HEALTHCARE · FASTER · 7 min · Read →
- [Can AI Speed Up Loan Decisions?](https://bfcs.ai/articles/can-ai-speed-up-loan-decisions/): Also Faster · FINANCE · FASTER · 7 min · Read →

---

Canonical page: https://bfcs.ai/articles/can-ai-improve-patient-triage/
This is the Markdown version of the page, generated from its HTML at build time.
Site guide for agents: https://bfcs.ai/llms.txt · Every page: https://bfcs.ai/sitemap.xml