By Alon Weizer, MD
A patient describes his symptoms, and for once I’m listening instead of typing, because an ambient AI tool is quietly drafting the note as we talk. When the visit ends I read it back, fix what it missed, and sign. The tool did the transcription; the judgment stayed mine. That small division of labor is the whole argument for how a health system should approach artificial intelligence, and most of the industry has it backward.
Every health system is being asked whether it’s ready for AI. The real question isn’t whether we’re ready. It’s whether our people are. A vendor will sell you a tool. Teaching an entire organization when to trust that tool, when to question it, and when to set it aside is the harder work, and it’s the work that determines whether AI improves care or just adds a system nobody uses correctly. Those skills aren’t for some future state. Our people need them now.
We state one principle deliberately, because it orders everything else: our work is primarily human-driven, with AI to augment and support it. The common phrase “human in the loop” suggests the system runs and a person checks it. In medicine, it runs the other way. The judgment that protects quality and prevents harm stays human. Technology enables that work. At the bedside, it does not replace it.
That principle shows up in how we deploy. Our implementation of Chart with Art, Epic’s ambient documentation tool, has cut physician charting time by nearly 30 percent per encounter. We extended it to inpatient nursing this year, among the first health systems in the country to do so, and built it to work in English and Spanish from the start, because a tool that only works in English doesn’t reflect the patients or the workforce it serves. Cosmos, Epic’s research network spanning 310 million longitudinal patient records across more than 350 health systems, lets a clinician facing a rare condition connect with doctors who have treated similar patients elsewhere, or see how comparable patients responded to a given treatment. But access isn’t automatic. Clinicians complete coursework first, and identifiable data requires added training and IRB oversight. The literacy requirement lives inside the tool, not bolted on after adoption.
So we build workforce readiness deliberately, through a tiered framework running from every employee up through executive leadership, calibrated to the risk each role carries. It starts with a floor most organizations skip, a shared vocabulary, because a workforce that can’t tell a large language model from an agentic one can’t adopt either safely.
Readiness also means naming what we haven’t solved. Newer models have made real progress on hallucinations, the early fear that AI would invent information. Bias and model drift, where a tool quietly stops producing valid results, have not been solved, and they demand continuous audit rather than a one-time launch.
Then there’s the problem few are naming. A decade ago, electronic health records consolidated dozens of disconnected systems into something coherent. AI is doing the opposite. Every function now arrives with its own tool, and the work re-fragments as fast as we automate it. Managing that sprawl, sometimes resisting it, is part of readiness too. So before we add the next tool, we ask what the return is. It doesn’t have to be financial. But it has to be articulated.
None of this slows us down, and none of it replaces clinical judgment. What readiness gives us is a workforce equipped to use powerful tools without losing sight of the person in front of them. It isn’t a rollout you finish. It’s a capability you build, one tier at a time, with the people doing the work.
Dr. Alon Weizer is Chief Medical Officer, Mount Sinai Medical Center.
