At 7:42 p.m. in a Nashville clinic, the last patient has gone home, the exam rooms are quiet, and a physician is still parked in front of the electronic health record, typing notes like a jazz drummer trapped in a spreadsheet solo.

That is the scene behind Rosenbloom, Holmgren, and Steitz’s review, Building Toward a Future for Electronic Health Record Systems (PMID: 42090453, DOI: 10.1146/annurev-biodatasci-092724-050949). The paper is not selling us a shiny robot doctor wearing a tiny white coat. Thank goodness. Instead, it asks a better question: how did the EHR become both the backbone of modern medicine and the keyboard-shaped anvil dropped on clinicians’ evenings?

The answer has layers. Like a good Miles Davis tune, the melody sounds simple until you notice everyone is playing in a different time signature.

The Chart Is Digital, But the Work Got Louder

Electronic health records were supposed to make medicine smoother: searchable histories, medication lists, lab results, quality measures, better handoffs. In theory, the EHR is the bandstand where every part of care gets coordinated.

In practice, it often became the place where clinical care, billing, regulation, documentation, safety alerts, insurance requirements, and patient messages all showed up demanding a solo.

Rosenbloom and colleagues trace how EHR adoption in the United States grew through incentives, rules, reimbursement systems, and quality measurement programs. That matters because the software did not evolve as a pure clinical instrument. It evolved as a billing saxophone, regulatory trombone, population-health bassline, and note-taking kazoo all bolted together. No wonder the groove gets weird.

The review’s main point is blunt: health IT is everywhere, but its benefits have landed unevenly. Some patients gain access through portals and remote monitoring. Some clinicians get better data. Some organizations improve measurement. Meanwhile, clinicians inherit a growing stack of clicks, inbox messages, and documentation demands, because apparently medicine needed more homework.

The Inbox Learned to Scat

Patient portals are one of the clearest examples of this tradeoff. They can help patients see results, ask questions, and stay engaged. That is good. Also, the inbox never sleeps. It has the energy of a trumpet player who just discovered espresso.

A 2023 observational study of Cleveland Clinic primary care clinicians found that incoming medical advice messages rose sharply from late 2019 to late 2021, from an average of 340 to 695 quarterly messages per clinician. Each additional 10 messages was associated with about 12 more minutes per quarter in the EHR outside scheduled hours (DOI: 10.1007/s11606-023-08577-7, PMCID: PMC10973312).

That does not sound huge until you remember this is on top of visits, refills, forms, labs, prior authorizations, and the ancient medical ritual known as “finding the thing in the chart that everybody swears is definitely in there.”

This is where AI enters, not as the headliner, but maybe as a competent rhythm guitarist. Large language models can draft replies, summarize visits, capture conversations, and organize notes. A Stanford quality-improvement study using GPT-4 to draft patient inbox replies found improvements in workload measures and work exhaustion, though clinicians still reviewed and edited the drafts (DOI: 10.1001/jamanetworkopen.2024.3201). Another UCHealth study embedded GPT-4 draft replies into the EHR across nine clinics and reported practical utility for patient-message workflows (DOI: 10.1001/jamanetworkopen.2024.38573).

The key phrase is “draft.” The AI should hand the clinician a clean chart, not grab the microphone and announce lab results like it owns the club.

Ambient Scribes: The Microphone in the Room

The flashier riff is ambient clinical documentation. These systems listen during a visit, transcribe the conversation, identify who said what, pull out clinical concepts, and draft a note for review. It is like having a medical scribe who never steals your snacks, though it may occasionally misunderstand the saxophone.

Early evidence is promising but still young. A Stanford pilot of an ambient AI scribe across 45 physicians and 17,428 encounters found the tool was used in 55.25% of encounters and was associated with reductions in note time, after-hours documentation, and total daily EHR time (DOI: 10.1093/jamia/ocae304). A recent rapid review of real-world digital scribes found generally reduced documentation time, but also emphasized the limited number of studies and variation across settings (PMCID: PMC12513689).

That variation is the whole song. A tool that helps a dermatologist may flop in psychiatry. A model that drafts a decent note may still miss nuance, introduce errors, or make the patient-clinician relationship feel like a podcast being secretly produced by a toaster.

Better Tools, Same Old Humans

The review is strongest when it refuses the easy hype. AI will not fix healthcare’s administrative burden if we simply pour it into broken workflows. That would be like adding a brilliant pianist to a band where nobody agrees on the tempo.

The better future is deeper integration: systems that fit into clinical work, reduce duplicate documentation, support decisions at the right moment, and make patient engagement easier without quietly mailing the bill to clinicians’ nights and weekends.

There is also a privacy and usability angle. If the future EHR produces cleaner summaries, safer document handling matters too. For patients and clinics juggling records, tools like pdfb2.io point toward a useful direction: private, browser-based document workflows where sensitive files do not need to wander off into the cloud like tourists without maps.

The hard part is not inventing more AI features. The hard part is composing them into care without creating more noise. Medicine does not need another soloist. It needs a tighter ensemble.

References

Rosenbloom ST, Holmgren AJ, Steitz BD. Building Toward a Future for Electronic Health Record Systems. Annual Review of Biomedical Data Science. PMID: 42090453. DOI: 10.1146/annurev-biodatasci-092724-050949.

Martinez KA, Schulte R, Rothberg MB, et al. Patient Portal Message Volume and Time Spent on the EHR: an Observational Study of Primary Care Clinicians. Journal of General Internal Medicine. 2024;39:566-572. DOI: 10.1007/s11606-023-08577-7. PMCID: PMC10973312.

Tai-Seale M, et al. Artificial Intelligence-Generated Draft Replies to Patient Inbox Messages. JAMA Network Open. 2024;7(3):e243201. DOI: 10.1001/jamanetworkopen.2024.3201.

English E, Laughlin J, Sippel J, DeCamp M, Lin CT. Utility of Artificial Intelligence-Generative Draft Replies to Patient Messages. JAMA Network Open. 2024;7(10):e2438573. DOI: 10.1001/jamanetworkopen.2024.38573.

Ma SP, Liang AS, Shah SJ, et al. Ambient Artificial Intelligence Scribes: Utilization and Impact on Documentation Time. Journal of the American Medical Informatics Association. 2025;32(2):381-385. DOI: 10.1093/jamia/ocae304.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.