Evaluating AI-generated clinical documentation: a narrative review of quality frameworks and the case for a purpose-built instrument

BACKGROUND: Clinical documentation is a cornerstone of safe, high-quality healthcare, yet has become an increasingly burdensome task for clinicians. Artificial Intelligence (AI) scribes have emerged as a promising solution to this problem, ambiently capturing clinical encounters and alleviating administrative load. Rapid deployment has occurred across healthcare settings, with many commercial options entering the market. However, adoption has outpaced development of rigorous evaluation frameworks, with no existing standardised method of evaluating the quality of AI-generated clinical documentation across products. OBJECTIVE: To review the existing literature on AI scribe implementation and clinical documentation quality evaluation, identify limitations of current approaches, and establish the conceptual foundations and domain requirements for a purpose-built evaluation instrument. METHODS: A narrative review was conducted, searching MEDLINE, Embase and Google Scholar in March 2026 with supplementary reference list screening. Records were screened by title, abstract and full text for relevance to AI scribe outcomes and error patterns, clinical documentation functions, and note quality instruments. Approximately 60 records were identified, and 41 references are cited, including background references and publications added during the revision. RESULTS: AI scribes consistently demonstrate reduced documentation burden and improved clinician wellbeing, however there are limited existing evaluative frameworks.Six recurring functions of clinical documentation were identified, and four AI-associated error patterns emerged: hallucination, misattribution, omissions and intent distortion. Instruments designed for human-authored documentation may be insensitive to these. AI-targeted adaptations incompletely address these gaps. Equity and linguistic bias remain largely unaddressed, with stigmatising language present in AI drafts and incompletely removed at clinician sign-off. CONCLUSIONS: There is a clear need for a validated, theoretically grounded evaluation framework designed specifically for AI-generated clinical documentation. Such an instrument should assess fidelity and attribution, faithfulness to clinician intent, transmission of clinical reasoning, downstream interpretability, and decision accountability, whilst assessing equity and context as cross-cutting considerations.