JMIR Publications

ABSTRACT

Background:

Nursing documentation supports patient safety and continuity of care but consumes substantial nursing time and contributes to burnout. Artificial intelligence (AI) tools such as speech recognition and large language models (LLMs) are increasingly proposed to automate documentation, yet nurse-derived specifications of where automation fits, how it must be bounded, and what safeguards it requires are scarce. Because nurses are accountable for the record and judge what may be spoken aloud, these boundaries cannot be inferred from system data.

Objective:

This study aimed to explore nurses' experiences of documentation burden and their requirements for and boundaries around automated documentation, and to translate these into human-centered design requirements.

Methods:

We conducted a single-center qualitative descriptive study with an embedded quantitative prioritization component at a national cancer center in Korea between July and August 2026, comprising 4 focus groups with 15 nurses who also completed a study-specific questionnaire. Deidentified transcripts were segmented into 462 meaning units, provisionally coded by 3 LLMs against a 49-code codebook, and fully adjudicated by 2 human researchers. Framework analysis, descriptive questionnaire summaries with item-specific denominators, and joint-display integration yielded design requirements, a task-automation fit matrix, and a risk- and sensitivity-adapted human-in-the-loop workflow.

Results:

Six themes were identified: documentation competed with direct care, producing deferred charting and memory reconstruction; fragmented systems required repeated entry across interdependent records; nurses sought event-time anchoring with flexible completion; voice capture was acceptable only within context-sensitive boundaries; propagation required verification to prevent passive trust; and records underrepresented relational and contextual nursing work. Repetitive input was the highest-ranked documentation problem (weighted score 26), followed by too many input items (17) and difficulty documenting in real time (16); nursing logs (12/15, 80%) and initial assessments (11/15, 73.3%) were rated most time-consuming. Voice-based automatic input was both the most anticipated and most concerning function (5/15, 33.3% each) and was judged difficult before patients or families (7/15, 46.7%) and in shared or noisy rooms (6/15, 40%). Five joint-display comparisons (3 convergence, 1 complementarity, and 1 expansion) informed 7 design requirements: (1) rapid interim capture; (2) concurrent event drafts with transparent time correction; (3) governed single-entry propagation; (4) bounded, clinician-controlled voice-to-draft capture; (5) risk-tiered verification before filing; (6) concise relational-care documentation; and (7) explicit ownership with auditability and staged implementation. These converged on a 5-stage human-in-the-loop workflow in which capture mode and review intensity scale with clinical risk; legal and privacy sensitivity operates as a cross-cutting constraint.

Conclusions:

Nurses' needs for automated documentation were specific and bounded. Automation should prioritize rapid event-time capture and verified, governed propagation of structured data, keep voice capture clinician-controlled and context-limited, and preserve nurse authorship and accountability. These requirements and workflow provide a nurse-derived specification for designing and evaluating AI-assisted documentation.