AI summaries cut perceived oncology workload—but 12 of 33 were incomplete
A prospective Dutch usability study found lower task-load scores and self-reported time savings. It involved nine staff participants and no patient outcomes, while four consultations contained clinically relevant discrepancies.
The short version
- What happened
- Nine staff participants completed the baseline assessment; the prospective single-group study included complete data for 33 consultations handled by six clinicians.
- Who it affects
- Clinicians may gain preparation time and a clearer overview, while patients need assurance that compression does not erase allergies, diagnoses or other details that change care. Human review remains central in this evidence.
- What remains unproven
- Nine baseline participants; 33 complete consultation records from one department; no randomised or concurrent comparator; incomplete inclusion from 72 eligible consultations; self-reported time saving; no patient-safety or outcome endpoint; commercial affiliations.
Keep in mind
Nine baseline participants; 33 complete consultation records from one department; no randomised or concurrent comparator; incomplete inclusion from 72 eligible consultations; self-reported time saving; no patient-safety or outcome endpoint; commercial affiliations.
Source analysis · 1 linked source. Primary analysed source date: 17 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Staff reported less effort during 33 AI-supported preparations
Researchers at the Netherlands Cancer Institute tested an AI tool that turns unstructured clinical notes into a structured summary before complex head-and-neck oncology consultations. Nine physicians, nurse specialists and supervisors completed baseline measures; six clinicians then used the system in 33 consultations with complete questionnaire data. Mean NASA Task Load Index score fell from 42.1 to 29.9, a 28.9% reduction. The estimated difference was minus 12.3 points (95% confidence interval minus 21.7 to minus 3.3; p=0.008). For 31 consultations with time estimates, staff reported saving 3.26 minutes on average.[1]
The workflow signal came with visible completeness and safety gaps
Participants judged preparation successful in 28 of 33 consultations, while only 21 of 33 generated summaries were complete. They reported four clinically relevant discrepancies. Two arose downstream when missing allergy information was represented in the electronic record as 'no known allergies'; the authors classified those as moderate risk because absent documentation is not evidence that no allergy exists. One discrepancy was a model misclassification, while another correctly surfaced information that had been hidden in the record. The allergy representation was changed after the study, so the revised version was not evaluated in these results.
For clinicians, a few minutes saved before repeated complex appointments could be valuable, and a structured summary may reduce the effort of searching long records. But a concise summary can also hide omissions or make an uncertain field look definitive. The study therefore supports continued supervised testing, not autonomous preparation or a claim that safety improved.[1]
The distinction that matters
What the small prospective study actually tested
The experiment measured staff experience and summary quality during consultation preparation—not patient outcomes.
Records
Convert unstructured notes into a structured summary
Prepare information before an oncology consultation
Workflow
Collect complete data from 33 consultations
28 successful preparations; 21 complete summaries
Safety
Review reported discrepancies
Four clinically relevant issues, including two allergy-field errors
A small before-and-after study cannot separate tool effects from study effects
This was a single-department, single-group pre/post usability study, not a randomised comparison. Of 72 eligible consultations, 43 were intended for AI-supported use and 33 produced complete data. The fixed order means practice, expectations or the novelty of the system could contribute to lower workload scores. Time savings were clinicians' estimates rather than observed timestamps, and the study did not measure diagnostic decisions, treatment changes, patient experience, adverse events or health outcomes.
Health~Holland funded the work. Three authors are kaiko.ai employees and the company may have a commercial interest; the paper says the model was frozen before use and non-commercially affiliated institute researchers collected and analysed the primary outcomes. Those safeguards are relevant but do not remove the need for independent replication. A stronger assessment would use a randomised crossover design, direct time measurement, blinded completeness review, prespecified safety events and multi-centre evidence that any efficiency gain does not trade away clinically important context.[1]