AI summaries eased oncology consultation prep, but completeness remained uneven
Artificial intelligence is moving into oncology practice not as a diagnosis engine, but as a clerk-like tool aimed at one of the most time-consuming parts of care: making sense of sprawling clinical documentation before the first visit. In this real-world usability study, AI-supported First Consultation Preparation, or AI-FCP, was designed to turn unstructured notes into structured summaries that clinicians could use when seeing new patients. The setting matters. First consultations in oncology often require rapid synthesis of pathology, imaging, prior treatment history, and referral material, all while a clinician is trying to establish rapport and frame decisions that may alter a patient’s treatment path.
Nine clinicians completed a baseline assessment and then used AI-FCP across 33 consultations in a prospective single-group pre-post design. The study focused on workflow performance rather than cancer outcomes, which makes the signal both narrower and more practical. Usability was assessed with validated questionnaires and study-specific tools, capturing how the software affected workload and preparation time in day-to-day practice. Because case numbers varied by clinic scheduling and enrollment timing, survivorship bias could have influenced which encounters were captured. Even so, the central pattern was straightforward: clinicians perceived less strain when the system helped assemble the consultation package.
Perceived workload fell by 28.9% after AI-FCP use, with a Hodges-Lehmann estimate difference of −12.3 and a 95% confidence interval from −21.7 to −3.3. Self-reported time savings averaged 3.26 minutes per case, with a standard deviation of 2.65 minutes. In a busy oncology clinic, minutes are not cosmetic. They determine whether the clinician enters a consultation with a coherent narrative or spends the opening portion of the appointment reconstructing it from scattered records. A tool that trims that burden may not change the biology of cancer, but it can alter the cognitive load at the moment when decisions are made.
The software did not perform uniformly across all cases. Preparation succeeded in 84.9% of consultations, yet only 63.6% of the generated summaries were complete. That gap is the most clinically relevant detail in the report. A summary can be usable without being comprehensive, but missing elements can matter when a consultation depends on prior staging details, treatment sequencing, symptom trajectory, or other pieces that determine whether an oncologist recommends further imaging, additional pathology review, or immediate treatment planning. In other words, the system appears to help, but it still leaves enough gaps to require human verification.
Workflow gains came from information organization, not from clinical replacement
The appeal of large language models in oncology is easy to understand. They can scan free-text notes, clinic letters, and other unstructured material faster than a human can, then reorganize those details into a form that supports decision-making. That capability is especially relevant in cancer care, where the first visit often functions as an information triage point. A structured pre-visit summary can reduce the mental overhead of searching across documents and may shorten the time between patient arrival and meaningful discussion of treatment options.
Still, the study’s design keeps the findings in the usability domain rather than the effectiveness domain. A single-group pre-post format cannot separate the software effect from workflow acclimatization, case mix shifts, or the simple fact that clinicians may become more efficient with repeated exposure. The variable case counts across the study period add another layer of uncertainty, because the consultations captured may not fully reflect the full range of patients seen in routine head and neck oncology practice. Those limitations do not erase the observed workload reduction, but they do define how far the result can be extended.
The user-centered design is one of the more important features here. AI-FCP was co-created with end users, which generally improves fit with clinical work because the interface and output format are built around how clinicians actually think and document. That matters in oncology, where the wrong kind of automation can slow rather than speed care. A summary tool that is too generic, too verbose, or too detached from local workflow may create new work by forcing clinicians to check every line. The present study suggests the opposite can happen when the tool is aligned with consultation preparation needs, although the incomplete-summary rate shows how fragile that balance remains.
There is also a broader clinical implication in the direction of the workload findings. Lower cognitive burden at the start of a consultation may preserve attention for the elements that machines cannot manage well: patient concerns, tradeoffs in treatment intensity, emotional context, and the subtle details that shape shared decision-making. Oncology consultations are not just data aggregation exercises; they are conversations in which timing, uncertainty, and prognosis converge. A tool that clears away some of the administrative fog could make that interaction more focused, even if the final interpretation still belongs entirely to the clinician.
One sentence in the data deserves emphasis: completeness was not equivalent to success. The system often prepared the consultation, but not always fully, and that difference will determine whether AI-FCP becomes a dependable assistant or a partial aid that still demands substantial human cross-checking. For now, the study supports usability and efficiency, not autonomy. Generalizability remains unproven, and the next step will require broader testing across larger and more diverse oncology settings to see whether the workflow gains persist when the tool faces more heterogeneous documentation and more complex case mixes.