CASE STUDIES · PUBLIC-DISCLOSURE COMPOSITES · 9 min
Healthcare AI Case Studies
Three composite case studies assembled from publicly disclosed healthcare AI deployments. Each tells a structured story — decision process, pilot design, measurable results, what the published evidence actually shows, and what the next-quarter scale plan looked like. Plus the on-prem alternative shape that Canadian residency-bound buyers most often end up with. These are not WalledCare customers; the evidence behind each is publicly cited.
Case study 1 — Kaiser Permanente + Abridge (cloud, Epic, ambulatory)
Setting. Kaiser Permanente — 24,000+ clinicians across 40 hospitals and 600+ medical offices in eight states and the District of Columbia. Epic EHR system-wide. Federated cloud-IT governance with mature BAA-based procurement.
Decision. Kaiser selected Abridge in 2024 for what is publicly described as the largest generative-AI deployment in healthcare history. The decision was driven by Epic Pal-tier integration depth, peer-reviewed evidence base, and scale references that matched Kaiser's own footprint.
Pilot shape. Phased rollout starting in defined ambulatory cohorts, expanding across geographies. Two-week clinician implementation cycle per cohort once Epic integration was complete. Ambient documentation only — not coding automation in the first wave.
Published results. Public disclosure has emphasized scale (24,000 clinicians) over a single canonical outcome number. The independent peer-reviewed evidence base behind Abridge — the JAMIA February 2025 University of Kansas Medical Center cohort, the multi-system QI burnout study (PMC, 2025), and the Wisconsin trial — anchors the expected outcome range. KUMC: 7× more likely to call documentation easy, 5× more likely to finish notes before the next patient, 81% workflow ease, 73% less after-hours documentation, 67% lower burnout risk.
What this case demonstrates. The cloud-vendor pattern works at scale when (a) the buyer's primary constraint is speed-to-value with Epic, (b) the BAA-based cloud-processing model is already accepted by procurement, and (c) the deployment is sequenced to manage the cohort-by-cohort change-management load. The Kaiser pattern is not directly applicable to residency-bound Canadian deployments or to non-Epic hospitals; it is the reference for U.S. Epic-first system-wide deployment.
Case study 2 — HCA Healthcare + Commure Ambient (cloud, platform play)
Setting. HCA Healthcare — the largest U.S. for-profit hospital operator, 180+ hospitals across 20 states. Mix of Epic and non-Epic systems across the network. Enterprise procurement aligned to platform-vendor relationships.
Decision. HCA Healthcare named Commure the exclusive ambient AI partner across its network in 2024. The decision was framed as a platform partnership rather than a scribe procurement — scribe is the entry workflow, with planned expansion into coding, RCM, and prior-auth surfaces using the same Commure platform.
Pilot shape. Less public detail than the Kaiser deployment. The platform-partnership framing implies a longer multi-year rollout with phased workflow additions; the operational reference scale matches the size of HCA's network.
Published results. Commure's broader platform metrics (cited in Hospitalogy and Fierce Healthcare coverage) include $25B+ in annual claims processed across the customer base. Specific HCA outcome data has not been publicly disclosed at deployment-cohort granularity.
What this case demonstrates. The platform-vendor pattern works for buyers whose primary constraint is "one strategic AI relationship across multiple workflows." It trades vendor consolidation benefits (single contract, unified telemetry, bundled economics) against the vendor-lock-in and roadmap-dependency that platform partnerships create. The post-Augmedix integration risk is real and not yet fully visible; see the Commure Ambient profile.
Case study 3 — Te Whatu Ora (NZ) + Heidi Health (public-system ED)
Setting. Te Whatu Ora — Health New Zealand — is the unified national public-health authority that absorbed the country's regional district health boards in 2022. Public-system funding model. Mix of legacy and modernized EHR systems across regional clusters.
Decision. Te Whatu Ora began a national rollout of Heidi Health across the 16 largest New Zealand emergency departments after a Hawke's Bay regional trial reported productivity gains. The deployment is one of the largest single-system ambient-scribe rollouts in a public-system context.
Pilot shape. Hawke's Bay productivity trial preceded national expansion. ED-specific deployment — different failure mode profile and integration shape than the U.S. ambulatory-first patterns. Heidi's browser-first integration model fit the heterogeneous NZ EHR landscape better than Epic-embedded vendors would have.
Published results. NZ Herald and Te Whatu Ora communications have referenced productivity improvements and clinician adoption signal without yet publishing a single canonical study-grade outcome number. Heidi's broader operational metrics (8M clinician hours per year, 2M consults per week globally) provide the scale envelope.
What this case demonstrates. Public-system deployment differs from U.S. enterprise procurement in two ways: integration constraints favor browser-first vendors over Epic-embedded, and outcome reporting follows public-system accountability cycles rather than vendor-marketing-cycle narratives. The pattern is directly relevant to Canadian public-system deployments where similar EHR heterogeneity applies; see the Canadian compliance hub for the regulatory overlay.
Case study 4 — The on-prem alternative shape (Canadian community hospital, composite)
Where the three preceding cases describe specific publicly-disclosed cloud deployments, this fourth case is a composite reference shape based on multiple Canadian community-hospital pilots Moneli Automation has scoped. No single hospital is referenced; the pattern recurs across residency-bound Canadian buyers.
Setting. Mid-sized Canadian community hospital, 200-400 beds. Meditech or Oscar mix depending on region. Operating under PHIPA (Ontario) or HIA (Alberta) or Law 25 (Quebec) with strict residency interpretation. Existing IT capability moderate; no Epic deployment.
Constraint. The U.S. ambient-scribe vendors at the top of the WalledCare directory are operationally credible but residency-bound. Cloud-region commitments to Canadian Azure / AWS regions are not contractually firm for many of them at the procurement timeline this hospital is on. The Kaiser pattern is not transferable; the HCA pattern requires a platform spend that doesn't fit the budget; the Te Whatu Ora pattern is closer but Heidi's Canadian-region contractual commitment is still in negotiation.
Decision. Hospital chooses an on-prem deployment using the WalledCare reference architecture — Llama 3.x via vLLM, Whisper for transcription, Qdrant for retrieval, Haystack for orchestration. Department-scale build at the ~$30K capex point (2× A100 80GB). Browser-first clinician surface, FHIR / HL7 sidecar to the Meditech or Oscar EHR.
Pilot shape. 30/60/90-day framework per the WalledCare pilot playbook. Shadow mode in weeks 3-4, clinician-supervised review weeks 5-8, expanded cohort weeks 9-12. Same audit framework as the cloud cases; differential audit log captured on hospital infrastructure.
Expected outcome envelope. Per the published evidence range — UCLA NEJM AI 9.5% time-in-note, Mass General Brigham 13.4 min/day — and the safety-rate envelope of the npj Digital Medicine framework (1.47% / 3.45%). The hospital owns its outcome data; the data feeds the next-quarter expansion decision.
What this case demonstrates. The on-prem alternative produces evidence the hospital owns and can reuse in subsequent vendor conversations. Compliance posture is "describe the data path on our own infrastructure" rather than "negotiate contractual residency choreography." For residency-bound Canadian buyers, the on-prem path is often the easier shape — see the Canadian procurement 2026 article.
What these case studies do not include
Three honest caveats. None of these case studies should be read as either a vendor endorsement or a guarantee that the buyer's deployment will produce equivalent outcomes:
- closeNone are WalledCare customer references. The first three are composites assembled from public disclosures (vendor press releases, journal articles, news coverage). The fourth is a composite reference shape. None should be cited as a Moneli Automation client deployment.
- closeOperational details are partial. Public disclosures emphasize scale and headline numbers. The operational details — clinician training programs, change-management cadence, audit-finding follow-ups — are usually unpublished. Comparisons should account for the missing detail.
- closeOutcomes are aspirational anchors. The peer-reviewed evidence range (UCLA NEJM AI, Mass General Brigham JAMA, multi-system QI study) anchors the realistic expectation. Vendor-curated outcomes often exceed the peer-reviewed range; treat the peer-reviewed numbers as the floor.
Where this fits in the WalledCare directory
Use these case studies as the procurement-conversation reference points — "is our shape closer to Kaiser, HCA, Te Whatu Ora, or the on-prem composite?" Pair with the vendor side-by-side matrix to filter by deployment shape, the Canadian compliance hub for the regulatory overlay, and the blog for the procurement-narrative companion pieces. Moneli Automation will publish named WalledCare customer case studies as deployments mature and customer disclosure permits.