AI consultation prep cut perceived workload in a small oncology usability study
A real-world usability study in oncology found that AI-supported first consultation preparation reduced perceived clinician workload and shaved minutes off the front end of a visit. The system was designed to turn unstructured clinical documentation into structured summaries for first consultations, a task that often consumes time before the patient ever enters the room. In this prospective single-group pre-post evaluation, nine clinicians completed a baseline assessment and then used the tool across 33 consultations.
The workload signal moved in the direction clinicians would want in a crowded outpatient clinic. Overall perceived workload, measured with NASA-TLX, was 28.9% lower after use of the system than at baseline, with a Hodges-Lehmann estimate difference of −12.3 and a 95% confidence interval from −21.7 to −3.3. Self-reported average time savings reached 3.26 minutes per case, with a standard deviation of 2.65 minutes. In the mechanics of oncology scheduling, a few minutes may sound modest; in practice, repeated across a full clinic session, those minutes can shape how much time remains for chart review, medication reconciliation, counseling, and the nuanced questions that follow a cancer diagnosis.
Structured summaries helped, but not every case was fully captured
Preparation succeeded in 84.9% of consultations, yet only 63.6% of summaries were complete. Those two figures belong together. A system that often works is not the same as a system that reliably reproduces every clinically relevant detail, and oncology consultation prep is unforgiving when medication histories, pathology results, imaging findings, prior treatments, and referral notes have to be assembled into one coherent view.
The design of the tool speaks to the central bottleneck in modern cancer care: data are already abundant, but they are scattered across notes, reports, and prior encounters. An AI system that can synthesize that material into a usable pre-visit summary may reduce the cognitive load of triage before the clinician speaks to the patient. In that sense, the benefit is less about replacing clinical reasoning and more about clearing the path for it. The consultation begins with less hunting and more interpreting.
A single-sentence reality check is unavoidable here: the summaries were not complete in more than one-third of cases.
That gap matters because first oncology consultations often set the tone for treatment planning, shared decision-making, and discussion of prognosis. Missing or incomplete details can force clinicians to verify the record manually, which erodes the time savings that AI was meant to create. The study therefore reads as an early workflow signal rather than proof that the system is ready for broad deployment in routine care.
Workflow gains came from a small, co-designed real-world cohort
The study used a prospective single-group pre-post design and was co-created with end users, a choice that likely improved practical relevance. Co-design matters in this setting because oncology workflows vary by clinic, subspecialty, and local documentation style. A tool built without clinician input often fails on the details that define usability: where information appears, how it is grouped, and whether it can be trusted at the pace of a live outpatient schedule.
Case numbers varied by clinic scheduling and enrollment timing, which introduces the possibility of survivorship bias. In plain clinical terms, the consultations that were available to be studied may not have represented every case that passed through the clinic, and that can tilt usability estimates toward sessions that were easier to complete or better suited to the tool. The small sample size also limits how far these findings can travel beyond the participating setting.
Still, the direction of effect was consistent with what would be expected from a well-integrated documentation aid. If an AI system can reduce the time spent assembling a first-visit summary, it can redirect clinician attention toward the parts of oncology care that machines do not own: rapport, risk communication, treatment preference, and the difficult alignment of patient goals with disease biology. The technology does not create those conversations, but it may make room for them.
Safety, bias controls, and the limits of an exploratory workflow signal
Several safeguards were in place to reduce bias from commercial involvement. The AI system was version-frozen before the first study session, and no prompt or model changes were made during the study period. Primary outcome data, including NASA-TLX responses, UTAUT-based items, net promoter score, self-reported time savings, and qualitative session notes, were collected and analyzed by non-commercially affiliated authors at the Netherlands Cancer Institute. Participants also identified discrepancies between AI-generated summaries and the underlying source documents during use, and four participant-reported cases underwent secondary review for safety classification.
Those protections strengthen the internal credibility of the usability signal, but they do not transform the study into a proof of clinical effectiveness. This was not a randomized comparison against standard preparation, and it was not powered to evaluate downstream outcomes such as diagnostic accuracy, treatment delay, patient experience, or clinician burnout over time. The report also did not establish whether the observed workload reduction would persist once novelty fades or whether performance would hold across other oncology subspecialties, languages, or documentation environments.
For now, the most defensible reading is narrow and useful. AI-FCP appears to have reduced perceived workload and saved a small amount of time during first consultation preparation, while leaving enough incomplete summaries to keep human verification central. In oncology, where the first visit often carries disproportionate weight, even a modest reduction in administrative friction may matter. The harder question is whether that same workflow gain can be reproduced outside a single center and across a broader range of patients, notes, and clinic realities.