CHIME25: How Healthcare IT Leaders Can Drive Human-Centered AI Adoption
AI can reshape clinician workflows and improve outcomes, but only if it delivers clear value with strong guardrails. The leaders at CHIME25 made one point obvious: success looks like transparency, security, scalability and reliability - not hype.
Healthcare may lag in some tech, but it's among the fastest adopters of AI. Adoption alone isn't the target. Working, trusted systems that reduce burden and improve care are.
Prove the Value of AI in Healthcare
The University of Kansas Health System tackled documentation burnout with an ambient dictation tool. The first product missed the mark, so they switched midproject. Starting with 25 providers, they measured a two-hour daily time savings per physician and then scaled. That freed up time to see more patients - or to simply reduce cognitive load.
San Joaquin General Hospital uses an AI algorithm to assess salvageable brain tissue for stroke patients arriving in the six-to-eight-hour window. That insight led to an updated protocol and opened access to function-restoring treatment for more patients. Their team reports favorable outcomes for those in the extended window.
On the payer side, Corewell Health uses generative AI to label and sort prior authorizations. The model augments staff rather than making decisions, speeding up member support and redirecting $500,000 in labor savings.
AI can also clean up handoffs. San Joaquin General found that one AI scribe worked for the ED but produced overly verbose notes for inpatient teams. Involving representatives from each department unlocked constructive feedback, and the vendor added customization that improved handoffs across settings.
Implementation Principles That Actually Work
Make it safe to test and move on. Corewell is on its fourth pilot for agent-based service support. They'd rather fail fast than roll out a tool that slows clinicians down.
Trust is as important as accuracy. The University of Kansas Health System is piloting Microsoft Copilot across IT and informatics but knows trust isn't automatic. Superuser-led, peer-to-peer coaching helps show how tools work in real workflows - and reminds people that "perfect" isn't the standard for humans either.
Generative vs. Agentic AI: Be Honest About Readiness
Generative AI thrives with human verification. Agentic AI acts, which raises the bar for governance and risk tolerance. Leaders warned against letting agents make decisions without a human in the loop, especially in clinical care.
Cost is a real barrier. Large systems can fund agentic solutions; smaller organizations often can't. Push on pricing, ask for proof of value and seek options that scale down without losing quality.
Keep Humans Centered
The goal isn't more tech. It's better time allocation. Clinicians need time for patients - and for their lives outside the hospital.
Intermountain Health uses AI to draft Epic in-basket replies based on the chart and the patient's message. Clinicians review before sending. That simple assist saves 20-30 seconds per message and cuts "pajama time."
As one leader put it: "At the center of healthcare is a human, a patient." AI should support human judgment, not replace it.
Governance, Risk and Measurement
Start with governance, not the shiny tool. Create a cross-functional committee to evaluate internal and external products. Require risk assessments and documentation up front.
Watch for algorithm drift and bias. Maintain an inventory of AI tools with visibility into how each model was tested and validated. Monitor performance continuously and retire what no longer performs.
Security and privacy aren't optional. Ensure data quality, clear audit trails and safe deployment patterns. Hallucinations and false positives happen - make sure they're caught before they hit patients.
Be intentional with rollouts. Intermountain moved to Epic and is turning on AI features one by one, only when they can be monitored with confidence.
A Practical Playbook for Healthcare Leaders
- Pick a real problem with measurable pain (burnout, handoffs, prior auth delays).
- Define success up front (time saved, turnaround time, outcome lift, cost impact).
- Pilot small with opt-in superusers; be ready to switch tools if outcomes lag.
- Keep humans in the loop for clinical use; allow more autonomy in IT operations.
- Stand up governance: risk assessment, documentation, data controls and clear escalation paths.
- Build an AI inventory with vendor validation details and audit logs.
- Measure continuously: performance, safety, bias, user trust and workflow friction.
- Negotiate hard on price; request usage-based or tiered models for smaller sites.
- Invest in enablement: peer coaching, quick-start guides, office hours and feedback cycles.
- Scale only after the metrics hold across departments and shifts.
KPIs That Matter
- Time saved per clinician per day (e.g., documentation, inbox)
- Documentation quality and handoff clarity by department
- Prior authorization throughput and member satisfaction
- Clinical outcomes for targeted pathways (e.g., stroke extended window)
- Clinician satisfaction, burnout indicators and "pajama time"
- Cost per encounter and total labor savings
- Model precision/recall, error rates and drift indicators
- Equity metrics: performance by demographic and site
What to Ask Your Vendors
- Show clinical validation, not just demos. What metrics improved, and where?
- How do you detect and respond to drift and bias in production?
- What audit logs, guardrails and human-in-the-loop controls are available?
- Can outputs be customized by specialty and setting without rework?
- What is the total cost of ownership, and how do we scale down for smaller facilities?
- How is PHI handled, stored and segregated? Can we use enterprise keys/tenants?
- What's the exit strategy and data portability plan?
If you need a quick reference on broader AI adoption trends, see the McKinsey State of AI. For EHR-native features and deployment guidance, review Epic's official resources.
Make AI earn its place. Start small, measure hard, keep humans in charge, and scale only after the data says it's working. That's how you get safer care, lighter workloads and stronger margins - at the same time.
Building skills across teams? Explore role-based upskilling options at Complete AI Training.
Your membership also unlocks:
