CHIME Survey Finds AI Use at 93%, With Post-Deployment Monitoring Far Behind

Vitea Newsroom
Editorial team
Nov 16, 2026
6 minutes
Editorial team
Physician with tablet showing AI

CHIME’s Most Wired survey finds AI in use at 93.2% of participating organizations, while only 16.4% of AI users fully document how models are monitored after launch. This week’s roundup also covers the AMA’s governance framework for clinical AI agents, WellSky data on staff demand for AI, a Penn State study on patient trust, and new research on how medical students handle AI errors.

AI use in healthcare organizations climbed again this year. The processes for checking that AI once it is running did not climb with it. CHIME’s annual Digital Health Most Wired benchmark, released Nov. 13, puts numbers on that difference.

The other four stories in this roundup approach the same problem from different positions: a physician association assigning accountability for AI agents, a survey of healthcare professionals who want more AI, and two studies on how people weigh AI output against human judgment.

For health system leaders, the common thread is practical. After an AI tool is approved, what evidence shows it is still performing as intended and being used within policy?

1. CHIME: AI use reaches 93%, but post-deployment monitoring lags

CHIME’s 2026 Digital Health Most Wired National Trends Report found that 93.2% of participating healthcare organizations use AI. Generative AI moved fastest. The share of organizations reporting full generative AI adoption rose to 54.8% from 31.0% in a single year.

Oversight did not keep up with that growth. Only 16.4% of AI users fully document how models are monitored after deployment, which puts governance and infrastructure on this year’s work list. In its announcement of the findings, CHIME CEO Russ Branzell said the lasting value of these technologies depends on the leadership, governance and infrastructure around them.

The report also measured resilience. Organizations that test failover are about twice as likely to restore all critical systems within four hours (45.7% vs. 23.5%). CHIME presents the finding as a correlation.

Why it matters: A model that passed review at go-live can drift as it meets new data, new workflows or a vendor update. Without documented monitoring, those changes tend to surface through complaints or incidents instead of routine checks. The failover finding is a useful comparison. Organizations that rehearse for downtime recover faster, and the same habit applied to AI means checking models on a schedule, against set thresholds, with a record of what was found. Oversight that stops once vetting is done leaves the longest stretch of an AI tool’s life unwatched.

2. AMA sets out how clinical AI agents should earn authority

The American Medical Association has outlined a governance framework for agentic AI systems that take on clinical functions. Its starting point is that clinical authority should be earned through evidence and kept through demonstrated performance, with oversight adjusted as performance, risk or the technology changes.

The framework organizes review of any agentic system around four questions:

  1. What clinical task does the AI perform?  
  1. Can it perform that task, and what happens if it fails?  
  1. What is it permitted to do, and under what oversight?  
  1. Who is responsible for the risks and outcomes?

The AMA is also developing a liability framework that ties responsibility to each party’s authority and ability to prevent harm. Developers would answer for safety and effectiveness within their control, and physicians for the clinical judgment they can meaningfully exercise. Health systems would be accountable for selecting, implementing, governing and monitoring AI in real-world use. The AMA also states that physician oversight cannot substitute for inadequate evidence or poor system design.

Why it matters: The AMA places a large share of responsibility with the organization that deploys the AI. A health system that approves a clinical agent will need to show what the agent was allowed to do, how those limits held in practice and how its performance was tracked after launch. Written permissions only carry weight if they are enforced where the agent is used and logged. The position on physician oversight also bears on policy design. Many AI policies treat clinician review as the final safeguard, and the AMA is saying that review cannot make up for a system that was poorly validated or poorly designed. Organizations that treat AI risk as patient safety risk will need controls that act before an output reaches the clinician.

3. WellSky: 69% of healthcare professionals want their organization to use more AI

WellSky’s fifth annual Evolution of Care report found that 69% of respondents want their organization to use more AI over the next 12 months.

Respondents named four conditions for confidence in AI tools: human oversight, clinical validation, transparency and fit with existing workflows. Training, workflow integration and competing priorities were the most common barriers to adoption. The full report concludes that scaling AI depends as much on governance and trust as on the technology.

Why it matters: When demand from staff runs ahead of what the organization has approved, staff often find their own tools. A clinician using a personal AI account can look much like an approved enterprise tenant on the network, which makes the activity hard to spot, let alone govern. That is how shadow AI usually takes hold. The conditions respondents named for trusting AI are also what governance is meant to supply. An organization that can demonstrate oversight, validation and transparency is in a stronger position to approve more AI, and to approve it sooner.

4. When AI disagrees with the doctor, patients trust the doctor less

A Penn State experiment, reported in Medical Economics, tested how patients react when an AI second opinion is added to a clinical conversation. The 135 participants discussed a stress-management plan with a doctor. In half the sessions, an AI second opinion agreed with the doctor. In the other half, it disagreed.

When the AI agreed, participants trusted the recommendation more. When it disagreed, they doubted the AI and also rated the doctor as “lazier,” even though the doctor’s advice was the same in both conditions. The author reads this as evidence that patients judge the clinician and the AI together, as one recommendation. It is a single experiment with a modest sample, and the policy conclusions are the author’s own.

Why it matters: AI output is reaching patients through portals, visit summaries and second-opinion features. When the AI and the clinician disagree, someone decides whether the patient sees that and how it is framed, even if the decision happens by default. The author argues that this choice determines where liability and lost trust end up, and that today the clinician absorbs both without anyone having decided it should work that way. Patients have said they want a human making the final call. Health systems deploying patient-facing AI can set rules for disagreement before launch, covering what the patient is shown, how it is explained and who answers for the outcome.

5. Medical students who trust AI more catch fewer of its errors

A cross-sectional study of 82 medical students, published in BMC Medical Education and summarized by Bioengineer.org, asked participants to label 10 AI-generated anatomy statements as true or false. Researchers had planted factual errors in 40% of the statements and did not warn the students.

General trust in AI showed a strong negative correlation with accuracy (r = −0.84). Students who rated AI output as generally reliable performed worse. Those who scored higher on the Cognitive Reflection Test, a measure of analytical thinking, performed better. The authors note that the design cannot establish cause and call for longitudinal and experimental follow-up.

Why it matters: Clinician review is a common safeguard in AI policies, particularly for documentation tools that draft notes for a signature. The study involved students, not practicing clinicians, but its results are consistent with research on automation bias, where people who rely on a system become less likely to catch its mistakes. An error that slips past review can become part of the patient’s chart. Read alongside the AMA’s position on physician oversight, the finding supports safeguards that do not depend entirely on the reviewer, such as limits on which tools can be used for which tasks and monitoring that tracks error rates over time.

Final thoughts

Each of this week’s stories concerns what happens after an AI tool is already in use. CHIME’s data shows that most AI users have not fully documented how they monitor models after deployment. The AMA expects health systems to answer for the governance and real-world monitoring of clinical AI agents. Staff want more AI than their organizations currently provide. Two studies suggest that pairing AI with human judgment is less straightforward than many policies assume: patients think less of clinicians when AI disagrees with them, and reviewers who trust AI miss more of its mistakes.

Questions for health system leaders this week:

  • Which AI models in use today have a documented monitoring process, and who reviews the results?
  • For each AI agent with clinical functions, are its permitted actions written down and enforced?
  • Would you know if staff were using personal AI accounts in place of approved tools?
  • When patient-facing AI disagrees with a clinician, what does the patient see?
  • What catches an AI error that the reviewing clinician misses?

Vitea is an AI governance platform built specifically for healthcare. Command enforces AI policy in real time at the point of use, Lens shows which AI tools are in use across the organization, and Pulse monitors AI behavior after go-live. Talk with our team about prompt-level policy enforcement.

Follow Vitea on LinkedIn for each week’s healthcare AI news.

Suggested for You

Inspired by what you’ve recently viewed.

Bring AI under control
without slowing innovation.
We're here to help you innovate and transform
Discover every AI in use, including shadow AI
Enforce 100+ out-of-the-box policies in real time
Stop risky AI activity before sensitive data is exposed
Continuously monitor AI performance and prove governance on demand
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.