Topic
Health-system copilots
Enterprise-scale AI deployments across health systems — governance, procurement, EHR integration, and the operational reality of running clinical AI at scale.
Health-system copilots covers the operational side of enterprise AI deployment inside U.S. health systems — the governance frameworks, procurement processes, EHR-integration realities, and post-deployment surveillance that decide whether an AI product survives past the pilot budget.
This is the layer where most healthcare-AI reporting is thinnest. The FDA-clearance news gets covered, the vendor funding rounds get covered, but the “we deployed it, here’s what actually happened at scale” story is harder to source and less flashy. It also matters more. HIMSS Analytics’ 2026 AI-maturity benchmark made clear that the median U.S. health system is still early on this maturity curve, and the gap between the top quartile and the median is widening.
What “mature” governance looks like in 2026
Multiple large systems now have well-documented AI-governance programs — Mayo Clinic’s AI Assurance Lab, Kaiser Permanente’s Center for Advanced Analytics, Mass General Brigham’s AI Governance Committee, Duke’s Institute for Health Innovation, and the Stanford AIMI Center all publish some version of their process. The specifics differ; the common structure runs roughly:
- Intake and use-case scoring. Anyone in the system can propose an AI use case. A committee scores it for clinical risk, workflow disruption, and evidence quality.
- Vendor and model due-diligence. Security review, data-flow review, BAA in place, PCCP or model-update commitments understood, dataset representativeness assessed.
- Pilot with defined success criteria. Real KPIs — not “clinicians liked it” — with a pre-specified stop condition.
- Post-deployment surveillance. Model performance monitored against pre-registered thresholds; drift detection; MAUDE/RUAIH reporting on incidents.
- Sunset triggers. Explicit criteria for pulling the tool if performance drifts, evidence changes, or the vendor’s PCCP is not honored.
Systems that skip any of these steps have — repeatedly, publicly — regretted it. Our essay on scaling AI beyond the pilot collects the recurring failure modes.
EHR integration is the whole ballgame
An AI product that lives outside the EHR has, in almost every case, failed to reach enterprise scale. Epic and Oracle Health both opened progressively deeper integration surface areas over 2024–2026. Vendor deployments now differ substantially by whether they are:
- Embedded in the Epic Cosmos / Oracle Fusion Analytics fabric (deepest, tightest, most Epic-/Oracle-friendly).
- Integrated via Epic’s App Orchard / Oracle Health Marketplace (mid-tier).
- Bolted on via SMART-on-FHIR (portable across EHRs, functionally shallower).
- Living in a separate browser tab (rarely survives).
Our reporting on the Epic-vs-Oracle 2026 AI comparison unpacks how this shakes out in practice. The 2026 Mayo-Google platform expansion is an important data point in a slightly different direction: platform-level agent capabilities that sit alongside the EHR rather than inside it, aimed at the orchestration layer that no EHR has yet mastered.
Where the money and attention are going
- Ambient scribes — see the ambient AI scribes topic.
- Inbox management and after-visit summaries.
- AI-augmented clinical decision support embedded in the encounter.
- Sepsis and deterioration prediction — a long-running category with a spotty evidence record; the newer generation is under more scrutiny.
- Prior-authorization automation on the provider side.
- Operational AI — OR scheduling, staffing, supply-chain — which does not touch clinical care directly but pays for many of the pilots that do.
- Workforce-side applications — covered in our AI workforce and clinician-shortage reporting.
The data and infrastructure problem no one wants to own
Every mature governance program eventually runs into the same wall: data. The models work as well as the data they are grounded on, and most health systems’ data is messier, less complete, and more silo’d than vendors assume. Our deep-dive on AI infrastructure and data governance in 2026 traces how the leading systems are investing in a shared data and model-governance layer — a thankless project that pays for itself only when the fifth or sixth AI tool shows up and discovers it already has clean data to run on.
HTI-2 and the transparency ask
ONC’s HTI-2 rule now requires health systems — not just vendors — to display and surface certain predictive-DSI attributes inside the EHR. See the HTI-2 enforcement playbook and the October 2026 enforcement update for what is actually being audited. For most health systems, this has forced AI inventory to become a durable artifact rather than a one-off slide. That is net positive for governance even if the compliance cost is real.
What a 2027-ready copilot program looks like
The health systems best positioned for 2027 share a few traits: a durable AI inventory, a governance committee with clinical and technical representation that actually meets, a data layer that supports more than one vendor at a time, a procurement playbook that treats PCCP and BAA language as first-tier contract terms, and a post-deployment surveillance program with real monitoring — not just dashboards. The gap between a system that has all of these and a system that has none of them is already visible in deployment velocity, and it will only widen.
What we cover
Our reporting in this topic focuses on: enterprise-scale deployment case studies (Mayo, Kaiser, Mass General Brigham, VA), EHR integration announcements, governance-program disclosures, procurement realities, post-deployment surveillance signals, HTI-2 compliance patterns, data-infrastructure and model-governance investments, and the operational AI stack that sits alongside clinical AI. Explore related articles and news below.
Frequently asked questions
- What is a health-system copilot?
- A health-system copilot is an enterprise AI deployment that lives inside a hospital or health system's clinical and operational workflow — ambient scribes, inbox-message triage, chart summarization, prior-authorization automation, scheduling and staffing tools, and AI-augmented clinical decision support embedded in the EHR. The term 'copilot' emphasizes that the clinician or operator remains in the loop and in control.
- What does mature AI governance look like in a health system?
- A mature program runs five steps: intake and use-case scoring, vendor and model due-diligence (security, data flow, BAA, PCCP commitments, dataset representativeness), a defined pilot with pre-registered success criteria and a stop condition, post-deployment surveillance with drift detection, and explicit sunset triggers. Mayo Clinic's AI Assurance Lab, Kaiser Permanente, Mass General Brigham, Duke, and Stanford AIMI all publish variations of this structure.
- Why does EHR integration decide whether an AI product succeeds?
- An AI product that lives outside the EHR has, in almost every documented case, failed to reach enterprise scale. Clinicians will not toggle to a separate browser tab for an AI tool during a visit. The deepest deployments are embedded in Epic Cosmos or Oracle Fusion Analytics; mid-tier deployments go through Epic's App Orchard or the Oracle Health Marketplace; portable SMART-on-FHIR integrations are functionally shallower; standalone tabs rarely survive the pilot.
- How is clinical AI performance monitored after deployment?
- Model performance is tracked against pre-registered thresholds, drift is detected on an ongoing basis, and incidents flow into MAUDE (for FDA-cleared devices) and institutional safety programs including the Joint Commission's RUAIH framework. Owners are paged if the model drifts outside its performance envelope, and mature programs have explicit sunset criteria that force a product off the shelf when drift persists.
- What AI categories are health systems investing in through 2026?
- The highest-dollar categories in 2026 are ambient scribes, inbox-message management and after-visit summaries, AI-augmented clinical decision support embedded in the encounter, sepsis and deterioration prediction (with more scrutiny on the newer generation of models), provider-side prior-authorization automation, and operational AI covering OR scheduling, staffing, and supply chain. Operational AI often pays for the clinical pilots that sit alongside it.