Where does multimodal AI cut workload the most?
The biggest workload reductions come from automating repetitive, high-volume tasks like image interpretation and report drafting. In breast cancer screening, an AI triaging system that automatically read the least suspicious mammograms and tomosynthesis exams reduced radiologist reading time by up to 72.5% — from 568 hours to 156 hours for 15,987 exams — while maintaining non-inferior cancer detection (95 of 113 cancers detected vs. 92 of 113 with double reading) and even lowering the recall rate by 16.7% [2]. Similarly, an AI system for upper gastrointestinal endoscopy (Report-Angel) automatically generated draft reports in 1.5 seconds per lesion, achieving 79-83% clinically acceptable reports and 89-92% lesion-level accuracy across prospective datasets [1]. These are not marginal gains: they free up hours of clinician time per day for the same clinical volume.
In lymphoma imaging, an AI model for segmenting metabolic tumor volume on PET/CT scans matched expert readers in 83% of cases (50 out of 60) on a public benchmark dataset deliberately enriched with challenging cases [5]. The AI produced these segmentations without any user interaction, which could dramatically reduce the manual workload of nuclear medicine physicians who currently spend significant time contouring tumors slice by slice. However, the same study notes that human supervision is still required to minimize errors in the remaining cases [5].
When does AI still need a human in the loop?
Even the best-performing AI systems here do not eliminate errors entirely, and they fail in predictable ways that require human oversight. In the endoscopy study, Report-Angel achieved 79-83% clinically acceptable reports, meaning roughly one in five reports still needed correction by a clinician [1]. In breast cancer screening, the AI triaging strategy reduced workload by 72.5% but still missed 18 of 113 cancers (16%) — though this was statistically non-inferior to the double-reading standard, which missed 21 [2]. The lymphoma segmentation AI matched experts in 83% of cases, but in the remaining 17%, the AI's results deviated beyond acceptable limits, and in 4 of those 10 cases, the AI was only partially concordant with at least one expert reader [5].
The type of error also matters. In a study of AI for assessing dental crown preparations, one model (Claude-3.7-Sonnet-Reasoning) showed excellent agreement with human experts (intraclass correlation coefficient of 0.89), but other models ranged from weak to no agreement (GPT-4o: 0.06; o3: -0.03), and two DeepSeek models failed to complete the task at all [3]. This highlights that not all multimodal AI systems are equally reliable — performance varies dramatically by model architecture and task. The authors of the dementia care framework paper argue that hybrid AI systems — which pair statistical learning with explicit clinical knowledge and clinician oversight — are essential to bridge the gap between benchmark performance and real-world safety [4]. Across all five studies, the consistent message is that AI reduces workload best when used as a tool to augment, not replace, human judgment.
What don't we know yet about AI and workload?
The current evidence is promising but limited in scope. Most studies are retrospective or prospective single-center validations, not large-scale randomized controlled trials across diverse hospitals and patient populations. The breast cancer screening study [2] and the endoscopy study [1] are the largest here, but both were conducted in specific screening programs (Córdoba, Spain, and Chinese hospitals, respectively) — results may differ in settings with different equipment, patient demographics, or radiologist experience levels. The lymphoma segmentation study [5] used a public benchmark dataset of only 60 cases, which is too small to generalize broadly. None of the studies directly measured clinician burnout, time saved per day, or cost-effectiveness — only proxy metrics like reading time or report generation speed. The dementia care framework paper [4] explicitly calls for more pragmatic evaluations that prioritize adoption, safety, equity, workload, and patient outcomes, rather than just benchmark accuracy. Until such studies are done, the answer to 'can AI reduce workload without increasing errors' remains a qualified yes for specific tasks, not a blanket endorsement for all clinical settings.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2021 to 2026, 4 from 2024 or later, 2 in Q1 journals, collectively cited 169 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 63 papers retrieved from a database of over 500 million.
Sources used in this answer
Domain specific multimodal large language model for automated endoscopy reporting with multicenter prospective validation.
In a multicenter prospective validation, the Report-Angel multimodal AI system for upper GI endoscopy generated draft reports with 79-83% clinical acceptability, 89-92% lesion-level accuracy, and an average processing time of 1.5 seconds per lesion, suggesting potential to reduce endoscopist workload.
AI-based Strategies to Reduce Workload in Breast Cancer Screening with Mammography and Tomosynthesis: A Retrospective Evaluation
In a retrospective evaluation of 15,987 breast cancer screening exams, an AI triaging strategy reduced radiologist workload by up to 72.5% (from 568 to 156 hours) while maintaining non-inferior cancer detection (95 vs. 92 of 113 cancers) and lowering recall rates by 16.7%.
Reliability of Multimodal AI for Assessing Preclinical Stainless Steel Crown Preparations: A Comparative Study With Human Experts.
In a cross-sectional study of 133 dental crown preparations, one multimodal AI model (Claude-3.7-Sonnet-Reasoning) showed excellent agreement with human experts (ICC=0.89), but other models ranged from weak to no agreement, and two DeepSeek models failed to complete the task.
Beyond black-box AI: Interpretable hybrid systems for dementia care
This perspective paper argues that hybrid AI systems pairing statistical learning with explicit clinical knowledge and clinician oversight are needed to bridge the interpretability and reliability gaps that currently limit adoption of foundation models in dementia care.
Validation of an AI Method for Automated Lymphoma Metabolic Tumor Volume Segmentation Using a Public Benchmark PET/CT Dataset.
In a validation on a public benchmark dataset of 60 lymphoma PET/CT scans, an AI segmentation model matched expert readers in 83% of cases (50/60) and was partially concordant in 4 more, but the authors emphasize that human supervision is still required to minimize errors.