I still remember the day the shiny new EHR went live. We were promised instant access to every lab, perfect legibility, and (my favorite buzzphrase) “more time for patient care.” What arrived instead was a blinking cursor waiting for someone to type every blood-pressure reading, family-history nuance, and treatment plan. That “someone” was often the physician. Handwritten scrawl, once dashed off in seconds (illegible but efficient), was replaced by after-hours typing sessions that stretched late into the night.
Soon I tried documenting during the visit, but that only planted my face in the glow of a monitor while a patient described wheezing across the room. Eye contact vanished; conversation felt transactional. The clerical load frayed my evenings and my patience. I flirted with hiring a human scribe: expensive, hard to train, and usually gone to medical school within a year.
Then I watched an early large-language model spin raw speech into structured text almost instantly. In that moment it was obvious: An AI scribe could do what no voice recognition software or short-term scribe ever managed—capture the clinical story faithfully, at the speed of conversation, and give the doctor back to the patient.
Electronic medical records have delivered on some of their early promises: illegible scrawl is gone, decades of labs and imaging are a click away, and a patient’s story no longer lives in a stack of manila folders. Yet those same records are choked with boilerplate and “note bloat,” requiring physicians to navigate through extensive duplicated sections just to find key information. Every data field still has to be typed by someone (usually the doctor), which has turned exam-room clicks into evening “pajama-time” charting. A 2017 study of 142 clinicians found that physicians spend an average of 86 minutes after clinic each day finishing EHR work, including simple actions like renewing prescriptions for ongoing medications, which can involve over thirty clicks. Add the fact that most EHRs still cannot talk seamlessly to one another, and the hidden cost of convenience becomes clear: more hours on keyboards, fewer at the bedside, and a rising undercurrent of burnout that technology was supposed to cure.
Certain specialties feel the EHR burden more acutely than others. Documentation-heavy fields such as primary care, family medicine, internal medicine, and outpatient subspecialties like my own, allergy and immunology, must record meticulous histories, track chronic-disease metrics, and log frequent follow-up visits. Each encounter generates pages of required elements: updated meds, spirometry results, new reactions, and prior-authorization forms for biologics. The sheer volume of notes, multiplied by high-patient throughput, makes real-time charting impossible; work inevitably spills into the evening, when facts blur and details are harder to recall. Confronted with that growing gap between clinic pace and documentation demands, I found myself searching urgently for a better way.
When the first large language models appeared, their raw output felt like science fiction, messy, yes, but bursting with promise. Hallucinations popped up, facts sometimes warped, yet each weekly update was a leap; the curve looked exponential. No clinic-ready, HIPAA-secure scribe existed, so I became an overnight student of AI and built my own. That meant solving for encryption, prompt engineering, and regulatory gray zones while the tech shifted under my feet almost daily. Early drafts could invent family histories or fill in details that did not exist, but careful prompt tweaks (and rapidly improving models) steadily drove the error rate down.
The result was an in-house ambient scribe that listens to the visit, processes speech in a HIPAA-compliant pipeline, and outputs a concise, professional note that reads as if I typed it myself. Details I used to lose in the frenzy now appear automatically, and the evening documentation grind has vanished. I finish every session with charts closed, see one or two extra patients a day, and feel the weight of clerical work lifted. What began as a curiosity became a daily partner in care.
Today the tool is indispensable; achieving the same productivity and job satisfaction without it would be hard to imagine. As physicians become more familiar with these systems and as they become easier to deploy, I expect AI scribes to be regarded as standard medical infrastructure. A 2024 JAMA Network Open study found that virtual scribes cut physicians’ EHR workload by an average of 5.6 minutes per visit: including 1.3 minutes less note-writing and 1.1 minutes less pajama time. Soon every doctor will have a personal, always-available scribe that completes accurate, HIPAA-secure notes, liberating clinicians to focus on patients in busy outpatient settings and, ultimately, across every corner of medicine.
Medicine is already a big-data enterprise: decades of labs, imaging, procedure reports, and now whole-exome or polygenic-risk scores all crowd the modern chart. No human can fully synthesize that torrent, yet an LLM can triage it in seconds, surfacing the three abnormal trends, the family-history red flag buried in a scanned PDF, and the guideline update published last night. The physician remains indispensable for gleaning nuance from a history and deciding what matters, but AI will point us to the right questions, flag drug-gene interactions, and cross-check every plan against the latest society recommendations.
As genomics moves from novelty to routine, discovery will outpace any doctor’s memory; AI will bridge that gap, instantly mapping variants to therapy or clinical trial eligibility. Drug interactions will be screened in real time, prior encounters retrieved without hunting through tabs, and evidence summaries delivered at the bedside.
Will AI replace doctors? No more than autopilot has replaced airline captains. Aircraft have been able to land themselves for decades, yet a human still takes the controls for the final flare. Likewise, algorithms will propose diagnoses and regimens, but a physician will review and sign off, for safety, context, and human connection. By offloading data grind while amplifying insight, AI can let fewer physicians care for more patients without burning out, positioning us as expert interpreters between the patient and a vast, ever-learning medical knowledge base.
AI can give us back what early EHRs quietly stole: time, connection, and clinical clarity. Health-system leaders should stop treating large-language models as side projects and begin funding them as core infrastructure, fully integrated into the EMR. Physicians must insist that every new tool be judged by a critical metric: does it create more face-to-face patient time than it consumes? And we must remain vigilant that AI does not repeat the sins of the first digital wave by substituting fresh clerical clicks for old ones.
For our part, clinicians should start building literacy in prompt crafting, data stewardship, and algorithmic bias, because patients will soon expect (and deserve) doctors who can translate terabytes of longitudinal and genomic data into clear, humane care plans. Medicine is changing fast, but if we deploy these systems wisely, AI will not distance us from patients; it will clear the screen, lift our eyes, and help us practice the kind of attentive, data-driven care we imagined when the digital era first began.
William S. Micka is an allergist-immunologist.
