Ambient artificial intelligence documentation may reduce administrative pressure in emergency care while preserving physician oversight. A 12-month evaluation published in the International Journal of Medical Informatics followed an in-house ambient AI scribe across 48 hospitals in a Spanish healthcare network. The system was used in more than 1 million eligible emergency consultations and was associated with shorter visits, stable transcription performance, better report quality and positive clinician and patient experience. Deployment covered general emergency care before expanding to trauma, gynaecology, obstetrics and paediatrics. Operational data, structured report audits and clinician and patient surveys were used to assess adoption, efficiency and experience.
Adoption Expands Across Emergency Services
The scribe captures clinician-patient conversations through ambient audio and converts them into structured draft notes. Speech recognition and language-processing components organise the content within predefined templates in the electronic health record. Clinicians can review and edit every field before validation, and no generated content enters the medical record without explicit approval. Physician responsibility therefore remains unchanged.
Must Read: AI Scribes Cut Workload but Need Oversight
Use of the system was optional. Patients were informed during the consultation and gave verbal consent before activation, while standard documentation continued when consent was not provided. Data used for the evaluation were aggregated and anonymised and processing followed applicable data protection requirements, including the General Data Protection Regulation.
Deployment began in general emergency care and later extended to four additional specialties. Use increased steadily over the year, despite a temporary fall when the new services were introduced. Overall adoption reached about 45%, rising from below 10% in the first month to almost 60% by the end of the period. More than 2,000 physicians used the tool at least once. The number of regular high-volume users also grew substantially, indicating that use became established among a sizeable group of clinicians as rollout progressed. Monthly active users increased throughout implementation, while new-user onboarding slowed after the early rollout as deployment approached saturation across hospitals.
Consultations Become Shorter While Accuracy Remains Stable
Emergency consultations supported by the scribe were consistently shorter than those completed without it. Average duration was about 46 minutes with the tool and 60 minutes without it, giving a reduction of roughly 15 minutes. Across monthly comparisons, the relative saving averaged just over one fifth and became greater during later stages of implementation. The difference remained visible throughout the year rather than appearing only during the initial rollout.
Transcription performance was stable as use expanded. Average accuracy remained close to 94%, with only limited monthly variation and no significant decline over time. Performance was also not linked to changes in clinical workload. Accuracy was measured by comparing AI-generated fields with the final physician-validated entries after review. The method assessed similarity across structured fields, while categorical mismatches received no credit.
A separate audit compared conventional and scribe-assisted reports from five physicians. Assisted reports achieved higher average quality scores, and improvement was seen across all five clinicians rather than being driven by one individual. The audit covered only 50 reports, making it small compared with the overall scale of deployment. It also used an automated assessment tool, although every result was reviewed by the network’s corporate medical documentation team. The findings therefore suggest better documentation quality but require confirmation in broader samples.
Clinician and Patient Experience Is Generally Positive
Clinician experience was assessed in two survey rounds during implementation. Overall ratings remained stable, with modest improvements in doctor-patient interaction, perceived system learning and transcription quality. The differences between rounds were not statistically significant, so the results do not establish that experience improved over time.
A later comparison grouped respondents according to how frequently they used the scribe. High-use clinicians gave better scores across all seven assessed areas, including patient care, doctor-patient interaction and stress and quality of life. However, the groups were small and the differences did not reach statistical significance. The pattern may indicate that clinicians who use the tool more often have a more favourable experience, but the evaluation cannot confirm that relationship.
Patient experience was measured through monthly Net Promoter Score data. Scores were modestly higher for scribe-assisted consultations than for consultations without the system. The difference remained stable across the year and was not related to the level of adoption.
Several limitations remain. Not all emergency specialties were included for the full period, clinician surveys relied on self-report and subgroup comparisons had limited statistical power. Longer-term effects on professional behaviour, documentation quality, patient outcomes and system efficiency are uncertain. Accent, voice characteristics and speaking rate were not assessed. All authors were employees of the hospital network and two held management roles, although no external funding was reported.
Large-scale deployment of an ambient AI scribe was feasible across a Spanish emergency care network and became increasingly embedded in routine practice. Use was associated with shorter consultations, stable transcription accuracy, stronger audited documentation and generally positive clinician and patient experience. Mandatory physician review kept final responsibility with the clinician before any content entered the medical record. The findings are tempered by small survey and audit samples, incomplete specialty coverage and uncertainty about longer-term outcomes. The network developed the system internally and employed all authors, including two in management roles.
Source: International Journal of Medical Informatics
Image Credit: iStock
References:
Alcazar-Peral JM, Alvaro-de ´ la Parra JA, Ciardo P et al. (2026) Deployment of an ambient AI scribe in emergency care: A 12-month evaluation in a large Spanish hospital network. International Journal of Medical Informatics; 220:106584.
