AI

Houston Methodist finds a sharp divide in how much AI documentation clinicians must check

Physicians working with an ambient scribe and nurses completing flow sheets face different review burdens. That gap is shaping what Houston Methodist asks its technology partners to do next.

Author
HDM Staff
Save
Share

    A physician reviewing an AI-assisted note and a nurse checking a flow sheet can work in the same hospital and have very different experiences with artificial intelligence. Dr. Roberta Schwartz, executive vice president of Houston Methodist Hospital and the health system’s chief innovation officer, put the contrast bluntly. “Our nurses are doing 70% of checking. Our doctors are doing 10% of checking.” Those were her estimates of the review burden for the products she described, a distinction that can disappear when hospital leaders hear a single figure for how much documentation AI can handle.

    The difference matters to Schwartz because her organization already pays people to do this work. A new product may produce an accurate note and still give a nurse another screen to reconcile. A hospital deciding whether to expand an AI contract needs to know whose work actually changes, how much checking remains and whether the department can use the time it gains. As Schwartz asked, “what are we replacing of our operations?”

    John Beadle, co-founder and managing partner of Aegis Ventures, led the November 17 Full Stack, Full Stakes discussion in Las Vegas. Joining Schwartz were Mike Ng, co-founder and president of Ambience Healthcare, and Siobhan Nolan Mangini, a partner at Venrock and a former public company chief financial officer. Beadle asked how a startup that had earned its place by doing one job well could justify taking on the next one. Houston Methodist and Ambience had already worked together for about three years.

    What a scribe knows before a visit

    Ng traced the expansion back to what a physician needs before opening the exam room door. A medication list may conflict with a previous note. The chart may be silent about how a particular specialist works or which local resources are available for a patient. An AI tool preparing the physician for the visit needs to decide what information to surface. Producing a fluent summary does little to resolve a discrepancy between the two sources.

    Pulling that information together also created an infrastructure problem for Ambience. Ng described an institution where a pre-visit preparation tool, deployed to just 10 percent of clinicians, generated more electronic record activity than the rest of the institution’s clinicians and software combined. He did not identify the institution or give an absolute transaction count. What the company learned from that deployment was that a second or third tool repeatedly retrieving the same information would make the problem worse. Ambience built a common way for its products to retrieve and use patient and clinical context across a visit.

    For Ng, “utilization is the backbone” of the investment case. If only one encounter in five involves the product, even a large improvement in those encounters has a ceiling. He put Ambience’s encounter-level use around 70 to 80 percent across its implementations and described work across roughly 100 specialties and subspecialties. He credited the attention paid to how individual specialties document care. The surgeon and the primary care physician may use the same electronic record while expecting different things of a useful note.

    Schwartz saw another opportunity in the people downstream of that note. Coders, registrars and clinical documentation staff spend time searching records or returning to clinicians for clarification. If relevant information is already available when the physician documents the encounter, some of that work can move forward. Before a hospital can count a saving, though, it has to decide what happens to the people and the time involved. Schwartz described reassigning employees or reducing positions over time through attrition as possibilities.

    Which outcomes belong in the contract?

    Ambience has begun putting fees at risk against measures agreed with the customer, Ng said. One health system may prioritize coding accuracy. Another may care more about keeping experienced physicians. The parties need a baseline before rollout, an agreed method of measuring change and a decision about which result will trigger payment. A busy hospital can experience a shift in clinician turnover for several reasons during the same period it introduces a scribe.

    Ng cited John Muir Health’s experience, where the partners reported a 44 percent reduction in primary care provider turnover. He also described Ardent Health’s calculation that every dollar spent on Ambience returned about three dollars in the first year. Ardent has separately reported on the deployment, including documentation time saved. Ng’s argument was about how a hospital could finance expansion. When its finance team accepts the return on the first set of licenses, the next group of clinicians has a source of funding.

    Nolan Mangini saw room for performance contracts while questioning how much upside a hospital should surrender. “You just have to have the data,” she said. Her model was a multiyear agreement in which the parties begin with a limited range of possible gains and losses, then assume greater risk as the evidence accumulates. In her words, “you ease into these things.” A narrowly defined measure gives the finance team something it can audit and the startup a basis for taking responsibility.

    Her example from pharmacy benefits showed how much of a business can develop after the first sale. SmithRx initially attracted employers by passing through rebates, she said. Employers then had reason to ask about lower-cost drugs, biosimilars and how claims were processed. Each problem stood close enough to the original one that the customer could understand why the same company should address it. A free trial of a novel product offers an investor much less evidence of that relationship than a customer deciding to move more of its spending.

    The work patients still do

    Ng recalled having to work out when to begin physical therapy after a back injury. He had received care, but questions that arose at home remained disconnected from the people treating him. He expects patients to consult any number of tools between visits. “Patients are going to choose their own front door,” he said. The harder task for the health system is getting the question, the answer and the treating clinician back onto the same care plan.

    Schwartz described the patient who depends on seven or eight specialists, or the family trying to follow an intensive care admission involving as many as 21 physicians. Each specialist can produce a better note while the family continues to carry information from one professional to another. For Houston Methodist, a consequential next use of AI would make it easier to see who needs to act on that information. The test is whether a patient gets to the right person sooner and whether the nurse or physician can tell that the handoff actually happened.

    Houston Methodist, Ambience Healthcare and Venrock join moderator John Beadle for the Startup Stage discussion. Photo credit: HLTH USA 2026.