The Challenges and Promise of AI for Clinical Documentation in Pediatric Inpatient Care

The integration of generative artificial intelligence (AI) into health care is occurring rapidly with the potential to transform clinical practice and patient outcomes. Among the most promising applications is the use of AI to improve clinical documentation. Ambient AI scribes, which use speech recognition, natural language processing, and machine learning to convert clinical conversations into clinical notes, offer the potential to streamline documentation,1 but their effects on health care quality and patient safety remain unclear.Most existing research on AI scribes focuses on outpatient care, with limited data on inpatient use, especially in pediatrics.1–3 Hospitalized children receive care rooted in their clinical course, which is focused on achieving multiple goals for discharge. During morning rounds, multidisciplinary teams discuss daily plans with families who influence clinical decisions. All of these need to be concisely but clearly communicated in clinical notes. Given these unique challenges, AI-assisted documentation must be thoughtfully implemented.In this article, we present physician perspectives from our pilot experience implementing an ambient AI clinical documentation system in the pediatric medical intensive care unit and pediatric intermediate care unit at Boston Children’s Hospital, a quaternary referral freestanding children’s hospital and teaching hospital. We reflect on the impact of this technology on clinical documentation quality, physician workflows, and cognitive reasoning and provide recommendations for tailoring AI to fully support documentation for hospitalized children without compromising care quality.We piloted a commercially available ambient AI scribe technology that uses a medically tailored automated speech recognition system to produce a transcript of rounds and natural language processing to transform the transcript into a draft clinical note. The pilot (December 2024–March 2025) aimed to optimize the AI system for inpatient settings, as it was initially developed for outpatient use. Six attending pediatric intensivists used the AI scribe to listen to rounds and generate admission and daily progress notes. Families and clinical staff were informed and given the opportunity to opt out. At that time, the AI platform operated independently of the electronic health record (EHR); thus, physicians copy-pasted generated notes from a secure external program onto the patient’s medical record. Physicians retained full control over final documentation, reviewing and editing the AI-generated note before signing.Ambient AI can capture clinical conversations in real time, meeting or exceeding clinician standards for inclusion of critical data in notes.2 However, AI systems can mishear information and introduce confabulations (factually incorrect information),4,5 posing risks to patient safety.6 AI note accuracy also declines with increasing case complexity and note length,7 raising concerns for children with medical complexity who often have detailed histories and multiple problems. In our pilot, the AI scribe captured most essential information accurately and was able to correctly differentiate team members speaking during multidisciplinary rounds. However, physicians also reported documentation errors. Many were misheard words/phrases easily fixed during editing, but others were confabulated tests or treatments that were not discussed at all or being used. Such serious errors obligate clinicians to review notes at a granular level, undermining expected efficiency gains.8High-quality clinical notes are not merely transcripts—they should reflect clinical reasoning, the patient/family’s values, and the appropriate clinical context.9 One of the known limitations of our pilot was the lack of meaningful integration of AI with the EHR, precluding consideration of the patient’s clinical course and relevant data in writing notes. Thus, during rounds, a short clinical summary needed to be verbalized to ensure that context was considered. Clinical reasoning was also inconsistently captured because this was often not explicitly discussed.EHR documentation burdens contribute greatly to clinician burnout.10 AI-driven tools can potentially address this by improving clinician efficiency.11,12 Early data show promising reductions in cognitive demand and better clinician engagement with patients,1,3,13,14 although benefits to long-term clinician wellness are still unknown. In our pilot, physicians found the user interface intuitive and noted the platform’s remarkable ability to capture and summarize information. However, potential time savings were counteracted by the considerable need to edit notes. Perspectives also varied regarding the technology’s impact on cognitive load, with some experiencing better engagement during rounds without the need to remember minute details, while others felt that line-editing notes and adding clinical reasoning documentation offset initial reductions in cognitive burden. Still, physicians recognized ambient AI’s potential for more efficient note-writing, potentially increasing time for clinical care and teaching and improving work-life balance.As with EHR adoption, ambient AI systems will likely shift not only how clinicians write notes but also how they reason through clinical problems.9 For many physicians, note-writing is a crucial part of synthesizing information and refining diagnoses or care plans. Some worry that delegating this process to AI might reduce critical thinking time.9,15,16 Similarly, physicians in our pilot expressed concern that overreliance on AI could erode the development of clinical reasoning skills for trainees and hinder optimal care for patients with medical complexity whose problems are often clarified during note-writing.Key concerns relevant to AI scribes include data privacy,11 informed consent, and transparency in documentation practices.4 In our pilot, concerns were raised about how protected health information was recorded, stored, and transmitted, even though staff were notified that recordings would be destroyed after 30 days. Some physicians noted that conversations during rounds were more guarded, with staff delaying sensitive discussions or questions until the recording was off. In contrast, families were generally accepting of the new technology, with all those approached allowing AI to be used.Based on our experiences, we offer the following recommendations for AI scribes to meet the specific needs and challenges of pediatric inpatient documentation.Developers must train models to become more adept at general and subspecialty-specific pediatric terminology, including norms for pediatric data. Confabulations must be addressed aggressively, as they threaten documentation fidelity, patient safety, and trust in AI technology. Safeguards for catching inaccuracies should be built into the system rather than relying solely on physician vigilance.8Aside from greater ease of note creation, meaningful EHR integration of ambient AI systems is crucial for incorporating relevant information not discussed on rounds and providing richer context for decisions. While some AI programs are already incorporated into EHRs in that they can add notes directly into the medical record, many still lack the ability to contextualize notes within patients’ overall clinical trajectory using prior documentation and data.In academic settings, resident physician or advanced practice provider notes are usually detailed with systems- or problem-oriented documentation, while attending physician notes provide a broader view of the plan and expected outcomes. Having AI systems that can be adjusted to generate detailed to more summative notes would give clinicians control over how the clinical story is presented and may eventually eliminate the redundancy of separate note-writing efforts. Moreover, AI systems can be trained to support clinical documentation improvement efforts for best billing practices.AI-generated notes must include more comprehensive documentation of clinical reasoning to achieve a shared mental model for clinical staff and patients/families. EHR integration can help by providing the necessary context for clinical decisions. Rounding scripts may need to be altered to include information justifying care plans so these can be included in notes by the AI scribe. Additionally, graduate medical trainees’ use of ambient AI tools should be closely supervised. Trainees’ clinical reasoning skills must be developed and evaluated through other means in addition to documentation.We must ensure the ethical use of AI systems despite lagging legal standards for implementation. Patients/families should be informed when ambient AI systems are in use and given the opportunity to decline. Rigorous standards for Healthcare Insurance Portability and Accountability Act (HIPAA) compliance, robust data encryption, and secure data storage by AI must be clearly established.17 Ethical issues regarding secondary data use for further AI research must be addressed.17 Clinical staff and patients/families should be educated on data security and safety measures to promote comfort in implementation. Increased transparency about AI- vs clinician-generated sections of notes and mapping of note contents to source rounding transcripts can further support trust. Finally, institutional policies must be clear that clinicians hold responsibility for final documentation.To be successful, AI developers, clinicians, and patients/families must collaborate to implement and evaluate these recommendations. Throughout our pilot, we worked closely with the AI system developers, resulting in iterative improvements contributing to the development of an EHR-integrated system supporting inpatient-specific notes.Ambient AI for clinical documentation in pediatric inpatient settings promises to be a transformative technology that can improve documentation, increase workflow efficiency, and support cognitive reasoning. Current limitations allow for ample opportunities to tailor this technology to enhance care. Future improvements will require collaboration across stakeholders and rigorous evaluation to ensure effectiveness without adversely affecting safety and quality.The authors thank Dr Chase Parsons and Ms Yiyao Barnoon for leading the implementation of ambient artificial intelligence documentation systems at Boston Children’s Hospital and for reviewing this manuscript.