Guillaume de Zwirek, CEO, Artera, says health system technology leaders face a simple choice in an increasingly AI-saturated market: master the details or risk falling behind. In a wide-ranging discussion, he argued that the pace of change now rivals the industry’s COVID-era telehealth surge and is forcing CIOs and their teams to make faster, better-informed decisions about where AI belongs and how to deploy it safely.

De Zwirek contends the ambient scribe wave made AI “part of the vernacular” across hospitals, but the next meaningful frontier is automation that handles high-volume, low-acuity operational work across channels—especially the phone. The appeal, he argues, is immediate bottom-line relief from persistent labor pressure, coupled with better patient experience and 24/7 responsiveness. “AI for a while has been a solution in search of a problem,” he said, but call-center workloads offer clear, measurable targets such as appointment verification, rescheduling, and routine information requests that dominate volume.

De Zwirek says agentic AI—task-completing systems stitched from multiple components, including speech-to-text, text-to-speech, and large language models—can manage these flows end-to-end. Health systems, he noted, are already seeing production deployments in specialty areas, and scaling from there. The opportunity is to remove friction at the front door while gathering the oversight documentation that boards expect: transcripts, accuracy scores, handoff rates, and patient-satisfaction markers.



“The technology is there today to allow an artificial voice to complete tasks on behalf of the patient, everything from scheduling to managing prescriptions, doing refills, things of that nature, and resetting portal passwords,” he said. The claim is not that AI eliminates human roles; rather, it concentrates human attention on exceptions and higher-acuity needs while turning routine traffic into consistently executed workflows.

Buy, Build, or Partner: Three Realistic Paths

Asked how a health system should obtain such capabilities, de Zwirek outlined three routes. The first is to build directly “to the metal,” assembling and operating the full stack—from model selection to speech layers to EHR and revenue-cycle integration—with the necessary DevOps, security, and ML engineering. He pegs the floor for a respectable in-house effort at several million dollars, plus ongoing costs to keep pace with infrastructure changes.

The second route is to buy a turnkey wrapper around generic voice-automation middleware. While this can accelerate pilots, he said it often doubles costs without delivering material advantages in latency, reliability, or healthcare tailoring, and it increases dependency on vendors who themselves depend on others.

A third route—contracting with a sector-specific partner that has already gone “to the metal”—can reduce time to value while aligning the stack with healthcare standards (HL7, SIU, payer rules, contact-center metrics).

For executives facing the reality that multiple current vendors now claim similar AI features, de Zwirek recommends prioritizing empirical evidence over marketing. That means requiring live, de-identified transcripts; scorecards that benchmark AI versus human performance for defined tasks; clarity on tenancy models and data isolation; and visibility into how often the vendor refreshes models, prompts, and guardrails as the ecosystem evolves.

Governance at the Speed of Change

A central tension, de Zwirek warned, is that AI infrastructure is shifting faster than many governance cycles. “We are using new forms of AI every two months in this company,” he said, describing a landscape where a single cloud or model upgrade can obsolete custom components overnight. In that environment, paper compliance (for example, HIPAA alone) is insufficient. Leaders should expect modern attestations such as HITRUST and SOC 2,