Artificial intelligence (AI) has captured the mainstream imagination in the last year with splashy projects such as ChatGPT. Most industries have touted the potential gains and wrestled with the likely costs that the incorporation of the technology will carry as it advances. In medicine, AI stands to loom large in the arena of clinical documentation. Documentation consumes immense time and drains vast energy from a burned-out workforce, with financial costs of documentation estimated at $140 billion annually.^1,2 Appropriately incorporating AI into clinical documentation is one of the few opportunities available to reduce dissatisfaction and inefficiencies stemming from serving the electronic medical record (EMR).

The challenges in modern medicine’s digital landscape are multifold. Extracting patient history can prove mundane when repetitively posing lines of questionings visit after visit. After the extraction, the provider must digest, organize, translate, and transcribe the conversation into the medical record, all while keeping clinical details distinct between patients. An automated, asynchronous, and accurate means of extracting and compiling clinical information ahead of visits is needed. A feasible approach is to use a comprehensive rules-based machine learning construct of layman’s questions posed to the patient that converts answers into medically sophisticated language to be used in a clinical document—much as a conversation is translated and transcribed by the provider in the current paradigm. Early efforts to automate subjective clinical documentation will approximate digital waiting room intake questionnaires. Achieving the goal of producing concise, accurate, and automated medical documentation from such questionnaires will require a significant investment in training raw patient responses to finalized provider-edited notes. Future iterations of history extraction can use physician-specific avatars to conversationally elicit the patient narrative, providing a more personalized digital previsit experience.

AI also holds promise in pulling together and parsing from multiple sources objective clinical data such as labs, radiology studies, and pathology reports.³ Note templates in the current state do this reasonably well, so long as the data are available in the native EMR. AI can vastly improve upon this if allowed to access, organize, and display clinical data in clear and bright ways. Creating conduits to go out and get pertinent clinical results improves efficiency for the back office and providers, while also potentially reducing redundant testing.

AI’s ability to procure and present a significant proportion of the subjective and objective data in a standardized format prior to a visit must then be wedded to ambient scribe technology and near real-time decision support to bring the full promise of AI to bear on clinical documentation. Numerous AI-assisted scribe services currently exist but require further refinement to integrate previsit AI inputs and patient-provider conversation into coherent, comprehensive, and human-like medical documentation. This project is likely to require training on thousands of hours of patient-provider conversation and construction of knowledge networks to drive the real-time decision support incorporating previsit data, visit-based ambient natural language processing, and clinical guidelines.

Figure. Components of a comprehensive artificial intelligence (AI)–assisted clinical documentation system. NLP indicates natural language processing.

The underlying technology to achieve these goals exists; the bigger challenge will be creating the relationships between technologists, physicians, and health systems to create operational products and overcome the inertia of the status quo. Patient data protection, sourcing training data, implementation costs, and vigorous validation of AI results will be crucial in successfully developing AI-assisted clinical documentation. As outlined in the Figure, the project will require many interoperable domains. Simpler domains such as an EMR-integrated patient interface for inputting history and torrents for ingesting clinical data must be designed to function in the context of the more complex systems such as the AI engine and knowledge network.

Alleviating the burden of clinical documentation is fundamentally aimed at improving the patient-physician relationship. The task at hand, with highs levels of repetition and fidelity, is well suited to be solved by AI. Automating clinical documentation is not asking technology to think for providers, but rather to clear the provider’s mind to do the critical thinking and form the relationships with patients that drew them to medicine in the first place.

Conflict of Interest Disclosures: K.G. and J.A.N. are board members and partial owners of OneLine Health, a health care software company.