Blog

Health systems are deploying AI into team coordination faster than anyone is measuring what it does there Health systems are putting AI into the clinical encounter faster than they can evaluate it. Ambient scribes, algorithmic triage, AI-assisted handoffs, automated patient messaging, nearly all of it lands on the same thing: how a team coordinates around a patient. In 2025, CMS began asking patients whether their care team worked well together. On the first national data, it ranked as the number one predictor of overall hospital rating. Those two facts are on a collision course, and almost nobody is measuring the intersection. What follows is the evidence, from CMS survey files, Medicare claims, and staffing data, for three claims. That "communication" is not one thing but at least three. That the channel AI touches most is the one that matters most and has been invisible longest. And that most AI disclosure to patients is structurally the wrong operation: it widens uncertainty rather than closing it. Hospitals already knew. The payment system didn't. The 2025 revision was the largest change to HCAHPS since 2006, and the teamwork question was among the additions. In an analysis of more than 18,000 Q1 2025 discharge surveys by the Healthcare Experience Foundation with PRC, staff collaboration and team dynamics ranked as the number one predictor of likelihood to recommend a hospital (32%) and the number one predictor of overall hospital rating (38%). Here is the part worth sitting with: this was not news to the hospitals already measuring it. Press Ganey's inpatient instrument has carried a version of this question, staff worked together to care for you, for years. Press Ganey's own research had already identified the patient's perception that the care team worked well together as among the strongest variables driving Likelihood to Recommend, and a companion analysis with Compass One found caregiver teamwork the single strongest predictor of patient loyalty. &nbsp; So, what changed in 2025 was not discovery. It was standardization, public reporting, comparability, and payment. When a patient rates communication highly, more than one thing may be happening. There is warmth, i.e., courtesy, listening, presence. There is resolution of uncertainty. The patient understands what happened, what comes next, what to watch for. And there is now, measured directly for the first time, coordination. Whether the people caring for them acted like a team. At the hospital level these are nearly indistinguishable. In public data the information and warmth composites correlate around 0.93, which is why two decades of HCAHPS analysis produced a slogan rather than a mechanism. But at the patient level, CMS's own table, built from 2.3 million completed surveys, puts nurse communication and discharge information at 0.31. The channels separate only when you stop averaging. The halo is an artifact of aggregation, not a fact about how patients think. Individual patients distinguish a courteous nurse from a useful handout perfectly well. Hospital means erase the distinction. That has a practical consequence worth stating plainly: a health system with respondent-level survey data can do this analysis today. A reader working from Care Compare cannot and shouldn't try. What separates once you look Written information carries independent weight. The item asking whether a patient received information in writingabout symptoms to watch for predicts overall rating even after controlling for every warmth item on the survey, courtesy and listening, from nurses and physicians alike. A printed sheet cannot be relational. The effect is modest, but warmth cannot generate it. Disclosure and resolution are opposite operations. Explaining what a new medicine is for predicts higher ratings. Listing its possible side effects does not, and in some specifications carries a negative independent weight. Same clinician, same conversation, same scale. One collapses the patient's uncertainty about what happens next. The other widens it, it informs them there is more to worry about than they knew. The signal lives after discharge. Split a Medicare episode into spending during the index admission and spending in the 30 days after. Communication measures predict nothing during the stay. Nothing at all. The entire effect is post-discharge, and the information channel specifically predicts lower readmission spending. Whatever these measures capture operates on the handoff, not on the performance while the patient is still being watched. And the deficit is an information deficit. CMS publishes experience scores separately for medical and surgical patients. Surgical patients rate hospitals 5.3 top-box points higher, but the gap concentrates almost entirely in information: doctor communication, discharge information, and care transition lead at roughly 1.2, 1.2, and 0.95 standard deviations, while nurse communication is 0.47 and cleanliness 0.30. Medical patients, with uncertain trajectories and more handoffs, are not treated appreciably less warmly. They are informed dramatically worse. One number deserves to stand alone. Among medical patients, 47.5% strongly agreed they understood their care when they left the hospital. Fewer than half. Why teamwork ranking first is the finding Warmth has always been the largest single driver of hospital ratings. It survives nearly every control, case mix, service mix, hospital size, wages, occupancy, and, in our analysis, actual RN paid hours from CMS occupational mix data. Adding real nurse staffing moved the warmth coefficient by essentially zero. That has been an awkward fact for the field, because the standard explanation for poor experience scores is that nurses are stretched too thin. The teamwork result offers a better account. What a patient reports when asked how often nurses treated them with courtesy is not a judgment about a person, it is a report across every nurse they encountered, on every shift, including nights and the weekend. "Always" is a statement about consistency, and consistency is a property of a team, not an individual. Which means warmth may have been partly measuring teamwork all along, with no item on the public instrument able to absorb that variance. Vendors had the item. CMS did not. Now it does, and it ranks first. The intervention implication is different in kind from what most programs do. Training targets the average interaction. Consistency is about the tail, finding and fixing the encounters where it failed. Those require different work, different data, and different management. The AI implication Telling a patient a model was involved in their care is structurally closer to side-effect disclosure than to written discharge instructions. That analogy needs defending, because most process disclosures are inert, nobody's experience degrades because a commercial lab ran their blood panel. Three conditions must hold for a disclosure to raise rather than lower uncertainty: it introduces a consideration the patient wasn't tracking, they have no way to act on it, and they perceive it as consequential. Lab-vendor disclosure fails the third. AI disclosure currently satisfies all three. Picture it. A patient is told an AI flagged something on their imaging. They leave with a follow-up scan in six weeks, a word they don't understand, and nobody to ask what "flagged" meant. They were given information. Their uncertainty went up. They will carry it for six weeks. Most AI transparency language stops at the first move: we used a model, here is what it does, here are its limitations. That is disclosure. It is not resolution. It leaves open the three questions the patient actually needs answered:<ul><li>What just happened?</li><li>What do I do next?</li><li>Who is responsible if something goes wrong? While those stay open, the patient goes home carrying more residual uncertainty than they arrived with. Experience scores and post-discharge outcomes both register the cost. Closing the loop is a design requirement, not a courtesy. The information transfer must be actionable, confirmed, and attached to a specific next step with a named owner. The system cannot treat the moment of disclosure as the end of its responsibility. The second failure mode, which is larger Human-in-the-loop is necessary and insufficient. The more precise test is whether the patient's loop is closed. But there is a bigger risk than an unclosed loop, and the new data names it. If teamwork is the number one driver of how patients evaluate a hospital, then the question to ask of any AI deployment is not only whether it disclosed. It is whether it made the team work better together, or whether it added a participant the team cannot see, cannot question, and cannot hand off to. An AI system can surface an excellent recommendation and still fragment accountability. It can insert a step nobody owns. It can produce a finding that arrives without a person attached. Every one of those degrades coordination while every individual interaction remains perfectly polite. That will show up in the number that now matters most, and no amount of communication training will touch it, because it isn't a training problem. It's an architecture problem. This is the uncomfortable possibility for health systems buying AI right now: the technology can improve clinician cognition, improve documentation, improve throughput, and lower the patient's experience of care, through a channel the organization has only just started measuring. The practical stance Information functions as insurance only when it reduces the uncertainty a patient carries. Much of what currently passes for AI transparency does the opposite, it adds a layer of explanation without collapsing the distribution of what happens next. The organizations that will own this space will do three things. They will treat loop closure as a first-order design requirement and measure whether patients can answer the three questions a week later. They will distinguish resolution from disclosure in their interfaces and their governance standards. And they will evaluate every AI deployment against whether it strengthens or fragments the team the patient experiences. Because the costly failures are not the ones that happen while the patient is still in the room. They are the ones that happen after.</li></ul>

AI Breaks What Patients Reward Most

What the new Ramp/Revelio data really tells us about AI and jobs — and the part most people are about to skip.This week you'll see a wave of posts with some version of the same headline: AI isn't killing jobs after all. They'll be pointing to a new study from Ramp and Revelio Labs, and the study deserves the attention. But the headline is the least interesting thing in it.Here's what the researchers did that's worth pausing on. Instead of asking companies how they feel about AI, they linked observed AI spending — actual card and bill-pay data — to workforce records across more than 21,000 U.S. companies. Spending is harder to perform than sentiment. When a measurement is expensive to fake, I pay attention. The finding everyone will quoteFirms that adopted AI most aggressively grew their headcount about 10% over two years. Entry-level roles — the ones we've been told are first on the chopping block — grew even faster, around 12%. And the growth wasn't walled off in one corner of the org chart. It showed up across engineering, sales, administration, customer service, finance, and marketing, with the strongest gains in the information sector.If you stopped reading there, you'd walk away believing AI is a jobs engine. Please don't stop reading there.The fine print that matters more than the headlineThree things in the methodology change the story.The gains went to the committed, not the curious. The lift appeared only for high-intensity adopters — firms spending roughly $30 per employee per month on AI in their first three months — and even for them, not until 6 to 12 months in. The companies that bought a subscription, ran a pilot, and called it a transformation saw no measurable change. AI rewarded conviction, not curiosity.The winners were already winning. The firms that grew were larger, faster-growing, and more engineering-heavy before they spent anything on AI. That's a selection effect, and it's a big one. We may be watching strong companies get stronger, with AI as one more accelerant — not AI lifting the average company.This is a tech-forward slice of the economy, over a short window. The dataset skews toward software and tech-adjacent firms, and it only runs two years. Look at a wider cross-section and the picture darkens: other research has found AI erasing on the order of 16,000 net jobs a month over the past year, with younger and entry-level workers carrying the weight. Both things can be true. The Ramp data describes a specific population under specific conditions. It doesn't repeal the rest of the labor market.<figure id="figure-ilpywvnqmimr0woixx"><img id="ilpywvnqmimr0woixx" src="https://cdn.durable.co/covers/c18GIE1tpydkwWTh92RAjFzVMFI8j8WKlChsuL26LM6kKqMrj5Lei4QeYLoIMZzr.png" alt="" title="" class="object-cover bg-gray-100 rounded-2xl md:rounded-3xl lg:rounded-4xl" style="aspect-ratio: 1/1; object-position: 70% 60%;" data-corners="default" data-aspect="1:1" data-caption="" data-positionx="70%" data-positiony="60%" data-fullwidth="false" data-mce-src="https://cdn.durable.co/covers/c18GIE1tpydkwWTh92RAjFzVMFI8j8WKlChsuL26LM6kKqMrj5Lei4QeYLoIMZzr.png" data-mce-style="aspect-ratio: 1/1; object-position: 70% 60%;"></figure>So what is the real finding?Not that AI creates jobs. The honest reading is narrower and, to me, more useful:AI multiplies the decisions a company was already making.A firm built to scale, with the appetite to reinvest productivity gains into more output, used AI to grow. A firm built to cut, or one that treated AI as a line item rather than a capability, got cutting — or got nothing. The technology didn't decide the direction. It amplified one that was already set.This is the thesis I keep returning to in my own work, and I'll name it plainly: the Human Middle. In any AI system, humans remain the design authority. The model doesn't choose whether AI expands an organization or hollows it out. People do — through strategy, through governance, through what they choose to do with the time and cost the technology gives back. This study is one more dataset confirming that the outcome lives in the middle, with us, not in the tool.Why this matters in healthcareIn health systems, the stakes on that middle are higher than almost anywhere.Deployed one way, AI lowers the cost of documentation, coordination, and clinical busywork — and a system that's wired to reinvest can turn that recovered capacity into more care: more time at the bedside, more roles, more reach into communities that have been waiting. Deployed another way, the same tool becomes a quiet justification for trimming the workforce and absorbing the savings, with the patient experience flat or worse.Same technology. Opposite outcomes. The difference isn't the model. It's whether leadership treats AI as a platform for ambition or a lever for subtraction — and whether the governance exists to make that an intentional choice rather than a default.The Ramp data won't settle the AI-and-jobs debate; no two-year window can. But it does hand us a quieter, more durable point. The tool isn't the story.What we decide to do with it is.https://ramp.com/data/ai-jobs-impact

AI Is a Multiplier, Not a Verdict

<h3>What Two New Studies Tell Us About Where Clinical AI Fails — and How to Fix It</h3>Yesterday the Stanford AI Index came out and what its medicine chapter actually tells us: the tools winning in clinical settings aren’t the most sophisticated ones. They’re the ones designed around how clinicians actually work. Construct determines efficacy. Not capability.Two studies published this week, both peer-reviewed, both from leading academic medical centers, explain exactly why that’s true. One shows the failure at the reasoning level. The other shows the architectural fix. Together, they make the most compelling case I’ve seen for why human-centered design isn’t a nice-to-have in clinical AI. It’s the variable that determines whether these tools help or harm.The Failure FirstA team from Harvard Medical School and Mass General Brigham just published a landmark evaluation of 21 frontier LLMs — including GPT-5, Claude 4.5 Opus, Gemini 3.0, and Grok 4 — across the full arc of clinical reasoning. Differential diagnosis, diagnostic testing, final diagnosis, management, and clinical reasoning questions. Every model. Every stage.The finding that should stop every healthcare leader cold: Failure rates exceeded 80% on differential diagnosis across every single model tested — meaning fewer than 1 in 5 differentials were fully correct. GPT-5. Grok 4. Claude 4.5 Opus. All of them. The most capable AI systems ever built fail more than 8 times out of 10 at the task that sits at the beginning of every clinical encounter.At the same time, final diagnosis accuracy was relatively strong — often above 85%. The same models that can’t generate a differential diagnosis reliably can identify the correct final answer when given enough information.This looks paradoxical until you understand what’s actually happening. The authors explain it precisely: clinicians preserve uncertainty and iteratively refine differential diagnoses, whereas LLMs collapse prematurely onto single answers. As corresponding author Marc Succi, MD, stated:“Differential diagnoses are central to clinical reasoning and underlie the art of medicine that AI cannot currently replicate.”Clinicians know what patients feel here: uncertainty is the hardest part of waiting through test results for the diagnosis. The brain loops. Stress stays high. People want an answer — any answer — to make it stop. LLMs are great at giving answers. That's the problem. They give them too early, collapsing possibilities when a human would hold them open. That’s not a data problem. It’s not a benchmark problem. It’s not something that will be solved by the next model release. It’s a reasoning architecture problem that has persisted across every generation of models tested.One important nuance: reasoning-optimized models — GPT-o1, DeepSeek R1, Claude 4.5 Opus — significantly outperformed standard models overall. But even they failed differential diagnosis more than 70% of the time. Better, not safe. The gap narrowed; the problem didn’t go away.LLMs are extraordinarily good at pattern recognition against known information. They are structurally weak at holding multiple competing hypotheses open simultaneously while iterating toward a conclusion under uncertainty. That gap — between what they can do and what early clinical reasoning actually requires — is where the danger lives.The Architecture That Fixes It A second study, from the Windreich Department of AI and Human Health at Mount Sinai, tested a lightweight orchestrator that routes each task to a dedicated worker agent handling one specific job.The results were striking. As workload scaled, as it always does in real clinical environments, single-agent accuracy collapsed from 73% to 16.6%. Multi-agent accuracy held at 65.3% even at the highest load tested.The same models. The same technology. A 48-percentage-point accuracy gap at scale, determined entirely by how the work was structured.The mechanism matters. Each worker agent receives only the tokens relevant to a single decision. Attention isn’t diluted across irrelevant material. The effective context stays within the range the model was designed to handle.Co-author Mahmud Omar put it plainly: “When a single agent handles everything, you can’t trace where it went wrong. With the orchestrator, every step is logged — which tool was called, what it returned, and how the answer was assembled. At 80 simultaneous tasks, the single agent dropped to 16 percent accuracy while burning 65 times more compute — and you’d have no way to figure out why. That kind of transparency isn’t optional in medicine.”What These Two Studies Say Together The Harvard/MGH paper tells you where LLMs break. They break at the beginning — at the open-ended, uncertainty-laden, iterative reasoning that clinical diagnosis actually requires. They perform well when the task is bounded and the answer is deterministic. They fail when the task requires holding ambiguity and reasoning forward.The Mount Sinai paper tells you what to do about it. You don’t ask the AI to do the things it can’t do. You constrain the task. You partition the work. You route each decision to the right agent with the right scope. And you keep the human exactly where clinical judgment is required — not as a checkpoint at the end, but as the authority at every stage where uncertainty is real.To be precise: the Mount Sinai team didn’t test differential diagnosis — they tested bounded tasks like retrieval, extraction, and dosing calculations. Their orchestrator architecture hasn’t been proven to fix the 80% differential diagnosis failure rate. But their architectural logic — constrain scope, isolate context, make every step auditable — is exactly what the Harvard paper suggests as the only responsible way to deploy LLMs clinically today.This is why the ambient AI scribe works. It doesn’t diagnose. It listens and transcribes. Bounded task. Reviewable output. Clinician authority intact. This is why the sepsis prediction tool works. One input, one output, one threshold. The AI surfaces the alert. The clinician decides.And this is why general-purpose clinical reasoning deployments fail. They ask the AI to do the one thing the Harvard paper proves it cannot do — hold diagnostic uncertainty, generate meaningful differentials, reason iteratively toward a conclusion.There is now direct empirical evidence for what happens when you put the human back in. A separate 2025 study of hybrid human-AI collectives — where physician diagnoses and LLM outputs are combined — found that these hybrid groups outperform both human-only and AI-only groups on differential diagnosis across multiple medical specialties. The reason is precise: when LLMs fail, physicians often provide the correct diagnosis. The human isn’t decorative oversight. The human is an active accuracy contributor — not a supervisor, but a participant — at exactly the point where AI is structurally weakest.The Design Principle Underneath Both Papers Neither research team frames their findings as a design argument. But that’s what they are.The tools that work started with the human — with how clinicians actually move through a patient encounter, where their cognitive load peaks, what decision authority they need to retain, what burden they’d give anything to put down. The technology was shaped around that reality.The tools that fail started with the capability — with what the model can theoretically do, and the assumption that clinical value would follow from technical performance. It doesn’t.The Mount Sinai paper’s conclusion is a design specification: constrain the task, partition the work, keep the context bounded, make every step auditable. That’s not an AI architecture document. That’s a human-centered design document written in the language of LLM infrastructure.The Harvard paper’s conclusion is a warning and a prescription in the same sentence: their most responsible role today is targeted, clinician-supervised use in low-uncertainty tasks. The researchers who just tested every major frontier model are telling us these systems should only be deployed on tasks where uncertainty is low and a clinician is supervising.That’s a narrow runway. Most of what vendors are currently marketing doesn’t fit inside it.What This Means for Healthcare Leaders Right Now The question isn’t whether AI belongs in clinical settings. It does, and the evidence for specific, well-designed tools is now robust. The question is whether the tools you’re evaluating were designed with these constraints in mind or in spite of them. What is the specific task this tool is performing? The more bounded and deterministic, the stronger the evidence base. The more open-ended and uncertainty-laden, the more caution is warranted, regardless of benchmark performance.Where does the human retain decision authority? Not as a final reviewer, but as an active participant at every stage where clinical uncertainty is real. Tools designed for this produce 112% ROI. Tools that aren’t produce fewer than 1 in 5 correct differentials.Can every intermediate step be audited? The Mount Sinai architecture produces a logged, replayable audit trail by design. Most single-agent deployments don’t. In a regulatory environment where AI incidents rose 55% in a single year, that difference is not academic.The technology is extraordinary. The failures are predictable. The gap between them is design. Sources: <a data-mce-href="https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2847679?utm_campaign=articlePDF&amp;utm_medium=articlePDFlink&amp;utm_source=articlePDF&amp;utm_content=jamanetworkopen.2026.4003" href="https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2847679?utm_campaign=articlePDF&amp;utm_medium=articlePDFlink&amp;utm_source=articlePDF&amp;utm_content=jamanetworkopen.2026.4003" title="Large Language Model Performance and Clinical Reasoning Tasks" target="_blank" rel="noopener">Rao AS et al.</a> Large Language Model Performance and Clinical Reasoning Tasks. JAMA Network Open. 2026;9(4):e264003. April 13, 2026.<a data-mce-href="https://www.nature.com/articles/s44401-026-00077-0" href="https://www.nature.com/articles/s44401-026-00077-0" title="Orchestrated multi agents sustain accuracy under clinical-scale workloads compared to a single agent" target="_blank" rel="noopener">Klang E, Omar M et al</a>. Orchestrated Multi Agents Sustain Accuracy Under Clinical-Scale Workloads Compared to a Single Agent. npj Health Systems. 2026;3:23.<a data-mce-href="https://arxiv.org/abs/2406.14981" href="https://arxiv.org/abs/2406.14981" title="Human-AI collectives produce the most accurate differential diagnoses" target="_blank" rel="noopener">Zöller N et al.</a> Human-AI Collectives Produce the Most Accurate Differential Diagnoses. arXiv. 2406.14981. 2025.<a data-mce-href="https://hai.stanford.edu/ai-index/2026-ai-index-report" href="https://hai.stanford.edu/ai-index/2026-ai-index-report" title="Stanford HAI 2026 AI Report" target="_blank" rel="noopener">2026 AI Index Report</a>, Stanford Institute for Human-Centered Artificial Intelligence (HAI). April 13, 2026.

The Art of Medicine is Still Human

Stanford’s 2026 AI Index was published this morning. All 423 pages of it. And while much of the coverage over the next few days will focus on the U.S. - China model gap closing and the $285 billion in U.S. private investment, I want to talk about what the medicine chapter actually tells us — because it validates something I’ve been arguing for years, with data that most healthcare leaders haven’t seen yet.The tools winning in clinical settings are not the most sophisticated ones. They’re the ones built around how clinicians actually work.That distinction sounds simple. It isn’t. &nbsp; Three Tools. Documented Results at Scale. Stanford identifies three categories of clinical AI that moved from pilot to enterprise-scale in 2025 with measurable outcomes: ambient AI documentation, sepsis prediction, and generative AI embedded directly in EHR workflows. Ambient AI scribes saw the broadest adoption of any clinical AI category. Abridge expanded from approximately 100 to over 150 health systems in a single year, reaching 63% adoption among hospitals running on Epic. The outcomes were consistent across institutions in ways that are rare in clinical AI research:• &nbsp; &nbsp; &nbsp; Sharp HealthCare: 83% reduction in note-writing effort • &nbsp; &nbsp; &nbsp; University of Chicago Medicine: 47% reduction in cognitive load, &nbsp; &nbsp; 58% increase in undivided patient attention • &nbsp; &nbsp; &nbsp; MaineHealth: 23% reduction in note time, used in 70.3% of encounters • &nbsp; &nbsp; &nbsp; Northwestern Medicine: 11.3 additional patients seen per month, 24% reduction in documentation time, 112% return on investment • &nbsp; &nbsp; &nbsp; Stanford Health Care: Statistically significant reductions in burnout, median 20-minute savings per half day of clinic That kind of consistency across different institutions, different patient populations, and different market segments is not an accident. It’s a signal about construct. Sepsis prediction produced the most clinically striking outcomes. The TREWS system, deployed across 13 Cleveland Clinic hospitals, reported an 18.7% relative reduction in sepsis mortality and an 89% clinician adoption rate. UC San Diego’s COMPOSER system reported a 17% reduction in sepsis mortality across more than 6,000 admissions, with an estimated 50 lives saved annually. These aren’t small studies. These are enterprise deployments with documented mortality impact. &nbsp; The Same Technology. Radically Different Outcomes. Here’s where it gets important. The Stanford-Harvard ARISE Network reviewed over 500 clinical AI studies and found that nearly half relied on exam-style questions rather than real patient data. Only 5% used real clinical data. Separately, the NOHARM benchmark found that leading LLMs — the same underlying models powering many of these tools — produced between 11.8 and 14.6 severely harmful recommendations per 100 clinical cases, with 76.6% of those errors being errors of omission: failing to recommend a critical test or intervention. Same models. Catastrophically different results. Stanford is explicit about why: those harmful recommendation rates apply to general-purpose LLMs deployed on open-ended clinical reasoning tasks. The tools driving real adoption — ambient scribes and sepsis alerts — operate within constrained workflows with clinician oversight. The technology doesn’t change. The construct does. This is not a subtle point. It is the entire argument for how healthcare organizations should be evaluating AI investments right now. &nbsp; Construct Determines Efficacy. Not Capability. The tools that are winning share three design characteristics that have nothing to do with model sophistication: 1. They operate inside existing workflows, not alongside them. The ambient scribe doesn’t ask a clinician to change anything about how they conduct a patient visit. It listens. It drafts. The human reviews and signs. The workflow is preserved; the burden is removed. 2. They constrain the AI’s decision surface. Sepsis prediction doesn’t ask the AI to reason across all of medicine. It monitors a specific set of variables and surfaces an alert at a defined threshold. The AI does one thing. The clinician decides what to do about it. 3. The human retains decision authority at every step. This isn’t “human in the loop” in the checkbox sense. It’s the human as the through-line — initiating, conferring, implementing, refining. The AI doesn’t replace clinical judgment; it protects the time and cognitive space required to exercise it. When these three conditions are met, the results are consistent. When they aren’t — when AI is deployed broadly on complex reasoning tasks without constrained scope or human oversight — the results are dangerous. &nbsp; The Governance Gap Is Wider Than Most Leaders Realize The FDA authorized 258 AI-enabled medical devices in 2025. But a peer-reviewed analysis of all authorizations through December 2024 found that only 2.4% of devices with clinical studies were supported by randomized controlled trial data. The vast majority entered the market through pathways that rely on existing evidence rather than new clinical trials. Healthcare leaders are being asked to make AI deployment decisions with an almost entirely absent real-world evidence base. Most clinical AI research still doesn’t use real patient data. Most cleared devices haven’t been validated with rigorous trial methodology. Multi-agent diagnostic systems are beginning to show extraordinary benchmark results — Microsoft’s AI Diagnostic Orchestrator scored 85.5% on complex NEJM cases versus 20% for unaided physicians. But benchmark results are not clinical outcomes, and the gap between authorization and deployment readiness is still wide. &nbsp; What This Means for Your Organization The question I hear most often from health system leaders right now is some version of: “Where do we start?” The Stanford data gives a clear answer: start with tools that reduce burden on clinicians within workflows they already own. The evidence base is there. The ROI is documented. The adoption rates are high because clinicians don’t experience these tools as disruption — they experience them as relief. The tools that are failing — or more precisely, the tools producing dangerous results — are the ones deployed with misplaced ambition. Asking a general-purpose AI to reason across an entire patient case is a different design challenge than asking it to transcribe a conversation or alert on a specific physiological threshold. There’s one more finding worth sitting with. Stanford documents that patient trust in AI is not driven by the technology itself — it’s clinician-mediated. Provider endorsement is the key determinant of patient acceptance. This means that reducing burden on clinicians isn’t just an operational win. It directly produces patient trust in AI tools across the care experience. The human isn’t a checkpoint in this system. The human is the through-line that makes the whole thing work.Read Stanford's full report <a data-mce-href="https://hai.stanford.edu/ai-index/2026-ai-index-report" href="https://hai.stanford.edu/ai-index/2026-ai-index-report" title="Stanford AI Index Report" target="_blank" rel="noopener">here</a>.

The Stanford AI Index:  The Winning Tools in Healthcare

Discernment and Gratitude exists simultaneouslyThe Complaint and the Gratitude Are the Same Voice.We hear the complaint first — but that's not because the gratitude comes later. It's because we stopped listening.Both emotions are present at the same moment. The person flagging the inconsistent handoff and the person saying you saved my life are the same person, in the same sentence, feeling both things fully and at once. Complaint and gratitude are not a sequence. They are a duality. What we do with that determines everything.What's actually happening is sensemaking.Patients in a hospital are already in a vulnerable, often frightened state. They are watching — carefully. When the third physician enters and washes their hands differently, addresses the family differently, or doesn't return when promised, it registers. Not as a personal failing of that physician. As a pattern break. And pattern breaks in uncertain environments feel like signals of something worse.That heightened state doesn't produce complaint or gratitude. It produces both — simultaneously, at full intensity. The fear and the trust. The frustration and the relief. Holding two things at once is not contradiction. It is the very human experience of being cared for when something is at stake.The staff receiving that feedback often hears only one half of it.They gave years to this work. The complaint lands as not good enough — when the patient is rarely saying that at all. When we only receive the complaint, we miss the gratitude that was always attached to it. And when staff carry only that half, they carry a wound that was never the patient's intention to leave.Learning to hear both at once — to hold the duality rather than collapse it into a single signal — is one of the most underestimated skills in care delivery.Most of the time, it's not about the person. It's about the process.When a patient starts to complain, something predictable happens. Staff pull back. Rounding slows. The instinct is to create distance from someone seen as difficult — to wait it out, document it, route it to patient relations.But that patient is not difficult. They are available.A patient willing to voice a concern in the moment is offering something most patients never will — real-time, unfiltered experience before it hardens into a survey score or a formal grievance. They are not afraid to talk. They are actually inviting you in. The instinct should be to move toward them, not away. That conversation — the one most staff are quietly avoiding — is where the gratitude is also sitting, waiting to come out alongside the concern.Instead what patients report is that they see fewer staff. Except for someone appearing to "check on the complaint." Which they notice. And which confirms every fear they already had about whether they are being cared for or managed.Behavior and process are not the same thing. When we look closely, the source of friction is usually the handoff, the protocol gap, the inconsistency baked into how care is delivered — not the human executing it. Addressing the person without addressing the process doesn't hold. The duality in what patients express is often their way of pointing directly at that gap — praising the people, flagging the system.AI is amplifying this problem in real time.As automation enters clinical workflows, the pattern breaks multiply. Processes that were never clearly defined — never documented from the lived experience of staff and patients — are now being handed to tools that will execute them at scale. The inconsistency gets faster. The blame starts landing on the AI, when the root issue was always upstream.The fix isn't better AI. It's better design before the AI arrives.Understanding what is actually happening — from the staff doing the work, from the patients living the experience — before a single workflow is automated is what determines whether the tool helps or amplifies the harm. Co-creation isn't a values statement. It's a design requirement.Build the process right first. Then build the tool on top of it.Because the patients watching your new AI implementation are doing exactly what they've always done — holding gratitude and concern at the same time. The question is whether the system you built was designed to hear both.

The Duality of Patient Feedback

The April Effect: Why Your Satisfaction Scores Are About to Dip (And Why It's Not Your Fault)It's April 1st, and if you lead patient experience, employee engagement, or any function measured by human satisfaction — your scores are likely sliding. They'll bottom out this month, recover through summer, and dip again (less severely) in fall.This isn't about your service recovery. It's not about staffing. It's biology.Years ago, a reporter called me at NYU Langone asking if I was worried about summer satisfaction dips from new residents. I couldn't answer — and that question launched a research journey I'm still on. What I found: the dip isn't summer. It's spring. And it's not residents. It's neurochemistry.Press Ganey confirmed the pattern across thousands of hospitals. I've since validated it in systems across the US and Middle East. Then I found it in the Google Misery Index — consumer search volumes for sadness, depression, anxiety, and stress peak in the same window. Then employee engagement data. Then financial markets. Then legal decisions. Then movie sentiment.The pattern is everywhere humans rate their experience of anything.The mechanism? Serotonin transporter binding follows a seasonal cycle — it peaks in spring, creating a window of neurochemical vulnerability that affects how we perceive and report satisfaction. April isn't when care gets worse. It's when our ability to feel satisfied gets harder.What this means for you:Your patients need more warmth right now — not because something's broken, but because their biology is working against contentment. Your staff shouldn't carry the burden of "failing" a score that's fighting physics.It matters if you're running a hospital, physician practice, building an app or measuring anything humans rate.How does your data look? Is your dip over or just beginning?

The April Effect

RoarHX

RoarHX helps organizations design the future of human-AI experience.
Founded by former healthcare Chief Experience Officer Dr. Joan Kelly, RoarHX provides frameworks, tools, and advisory services that help leaders identify where AI can transform workflows, improve patient and consumer experience, and unlock new operational value.
Through persona modeling, workflow mapping, and AI readiness assessments, RoarHX helps organizations move from AI curiosity to real-world implementation.

Insights

AI Breaks What Patients Reward Most

AI Is a Multiplier, Not a Verdict

The Art of Medicine is Still Human

The Stanford AI Index: The Winning Tools in Healthcare

The Duality of Patient Feedback

The April Effect