Evaluating Patient and Professional Satisfaction and Documentation Time Reduction Through AI-Driven Automatic Clinical Note Generation in Primary Care: Proof-of-Concept Study.

Background: The workload that stems from writing clinical histories is one of the main sources of stress and overload for primary care professionals, accounting for up to 43% of the working day. The introduction of technology, specifically artificial intelligence (AI), in the field of health could significantly reduce the time spent writing clinical reports without compromising the quality of care. Objective: The objective of this study is to evaluate the impact of implementing an AI solution for the automatic transcription of consultations in several primary care centers in Catalonia. Methods: A proof of concept of a multicenter study was carried out with alternating assignment of consultations to the intervention group (use of an AI assistant that automatically generates consultation notes) or control group (usual clinical practice). The impact was evaluated through the recorded documentation time and the initial quality of the transcription measured with the Levenshtein distance expressed as corrected words per minute, complemented by a qualitative categorization of clinician-reported errors and the perceived satisfaction of patients and professionals through questionnaires evaluated through a Likert scale. Results: For the intervention group, the average processing time was 6.63%, while the review time by the professional amounted to 15.2%. Because documentation-time data were not available for the control group, no direct between-group comparison of time savings was possible; time-related findings are therefore exploratory and limited to intervention-group process and review metrics. Levenshtein-based estimates showed that in most cases, the review was <24 words per minute and 26% of drafts required no edits, indicating a high-quality initial transcription. A qualitative analysis of clinician feedback showed that context or meaning errors were the most frequent, while unsupported additions or hallucinations were uncommon. The satisfaction surveys were answered by 289 patients and 213 professionals. Patient satisfaction was high (≥4/5), with no statistically significant differences between the control and intervention groups. The professionals rated the audio quality at 9.06 out of 10 (SD 1.18; medicine) and 7.62 out of 10 (SD 1.58; nursing) and the transcription at 8.14 out of 10 (SD 1.74) and 6.93 out of 10 (SD 1.52), respectively. Conclusions: The implementation of an AI tool was feasible in routine primary care, was well accepted by clinicians, and did not negatively affect patient satisfaction, with a generally low transcription review burden. However, this proof-of-concept study does not allow conclusions about comparative time savings, and adequately powered randomized studies are needed to confirm benefits for care quality and efficiency.