Can AI triage chatbots reduce clinician workload without increasing errors?

The best case: AI triage can slash workload without harming outcomes

In high-volume, standardized screening programs, AI triage can dramatically reduce the number of cases that need human review while keeping detection rates steady. A 2021 study of nearly 16,000 breast cancer screening exams found that an AI triage system could cut radiologist reading time by 72.5% compared to double reading of digital breast tomosynthesis (DBT) images — from 568 hours down to 156 hours [1]. At the same time, the AI detected 95 of 113 cancers versus 92 detected by double reading, a difference that was not statistically significant, meaning cancer detection was noninferior [1]. The recall rate (false alarms) also dropped by 16.7% [1]. This is the strongest quantitative evidence among these studies that AI can meaningfully reduce workload without increasing errors — but it applies to a very specific task: triaging mammograms into low-risk (skip human review) and high-risk (send to radiologist) categories.

The catch: AI struggles with diagnostic uncertainty and complex cases

When the clinical picture is ambiguous — the hallmark of many real-world triage decisions — AI chatbots underperform compared to trained clinicians. A 2024 study tested two leading chatbots (GPT-4o and Claude-3) on questions designed to involve diagnostic uncertainty, where symptoms and history don't point to a clear diagnosis. Family medicine residents scored 61-63% correct, while Claude-3 scored 57.7% and GPT-4o scored 53.3% — both significantly worse [2]. Most of GPT-4o's errors were logical errors (62.5%), meaning it made reasoning mistakes rather than missing facts [2]. Similarly, in emergency triage, ChatGPT-4o showed only moderate agreement with expert physicians (kappa 0.695, where 1.0 is perfect agreement) and was least sensitive (50%) for mid-urgency cases (Triage Level 4), meaning it often under-triaged patients who actually needed attention [4]. These results show that AI triage is not yet reliable enough to replace human judgment in complex or borderline cases.

The gap between potential and reality

The studies paint a clear picture: AI triage can excel at narrow, repetitive tasks like reading mammograms or sorting obvious emergencies, but it falters when nuance is required. In ophthalmology triage, ChatGPT-4 listed the correct diagnosis in its top three 93% of the time and gave appropriate triage urgency 98% of the time — comparable to physicians [3]. Yet in the same study, Bing Chat was less accurate (77% correct diagnosis) and tended to overestimate urgency, showing that not all AI chatbots are equal [3]. A 2026 paper on an AI-powered e-triage system deployed in UAE hospitals reported a 45% reduction in documentation time and 82% improvement in prioritization accuracy, but it also acknowledged ongoing challenges with data privacy, algorithmic bias, and integration with existing electronic health records [5]. The takeaway: AI triage can reduce workload in specific, well-defined settings, but deploying it broadly without human oversight risks missing critical cases or misclassifying patients.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2021 to 2026, 3 from 2024 or later, 1 in Q1 journals, collectively cited 247 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 50 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions