Artificial intelligence in the medical practice: practical tips

Abstract

The incorporation of generative artificial intelligence (AI) into clinical practice has moved from promise to an everyday decision. Using a computer or phone, clinicians can now transcribe an encounter, summarize a lengthy medical record, draft a report, or retrieve scientific evidence within seconds. Availability, however, does not equal usefulness or guarantee safety. This presentation offers a pragmatic tour of this ecosystem, using the diabetes clinic as its starting point and clinical need —rather than technology— as the criterion for selecting a tool.

Four families of tools with concrete applications in daily work will be discussed. First, ambient documentation assistants can produce draft clinical notes from the patient-clinician conversation. Their use has been associated with lower documentation burden, less after-hours work, and reduced professional burnout, although consent, data protection, and clinician review are required before any text enters the medical record.¹ Second, general-purpose language models can support writing and synthesis, including insurance reports, referral summaries, patient instructions, and educational materials adapted to individual health literacy. Third, environments that query clinician-provided documents can interrogate guidelines, protocols, longitudinal records, glucose sensor reports, or insulin delivery data while keeping answers linked to the supplied sources. Fourth, AI-assisted evidence search tools can accelerate the initial response to focused clinical questions without replacing critical appraisal of the original research.

The session will also provide operational recommendations for structuring an effective instruction, de-identifying information, checking outputs, and recognizing tasks that should not be delegated. The approach adopts a human-in-the-loop model: AI proposes; the professional verifies, decides, and signs. Access to a model alone does not necessarily improve clinical reasoning; benefit depends on training and deliberate integration into the workflow.²

Known limitations will be addressed, including hallucinations, fabricated references, response variability, bias, confidentiality, and professional accountability. Mitigation strategies include restricting answers to selected sources, retrieval-augmented generation, and systematic human verification, while recognizing that none of these measures eliminates risk completely.³

Finally, the presentation will distinguish control algorithms already validated in automated insulin delivery systems from generative applications, whose clinical maturity remains limited.⁴ The objective is for participants to leave with criteria they can apply the following Monday: select a real problem, design a small and repeatable workflow, measure its benefit, and keep clinical accountability in professional hands. Used in this way, AI may reduce friction and recover time for listening, interpreting context, and making decisions together with the person living with diabetes.