A MedCity News commentary from Hygia Chief Clinical Officer Gennady Gekht argues that a transactional, self-limiting model for healthcare AI may be too narrow for the way patients and providers actually use technology. The piece responds to a recently published MedCity News article in which Imprivata CEO Fran Rosch said AI agents should be handled with regular reminders that they are not to be trusted and compared them with contract nurses in the healthcare sector.
Gekht does not reject the security logic behind that position. He writes that minimizing AI access to critical patient information can reduce potential risk. His argument is that the same design choice also limits what healthcare organizations want from AI in practice: longitudinal learning, continuity, sustained patient engagement, and better return on investment.
The Strategic Disagreement
The core divide in the piece is whether safety in healthcare AI comes mainly from restricting the agent or from allowing the agent to earn trust over time under monitoring. Gekht argues that if agentic AI is compared to a nurse, the relevant lesson is not only limiting access but also allowing learning. In his view, an AI that works on a problem longitudinally can adapt to client-specific needs, learn behaviors and workflows, create more sustained engagement, and improve processes.
That matters particularly for small practices, which he says play a critical role in healthcare delivery but cannot keep absorbing rising administrative demands without consequences. It also matters for organizations that value continuity over a rotating cast of temporary tools. In this framing, a new AI for every task is not just a user-experience issue; it creates repeated introductions, permissions, and workflow changes that can each become a point of failure.
Trust As An Operating Requirement
Gekht’s article places patient trust at the center of the adoption case. He writes that a virtual healthcare companion needs to interact with patients over time, and that patients must develop trust not only in the healthcare organization but also in the technology itself. He adds that this can be especially difficult for certain vulnerable patient demographics.
The distinction he draws is between a familiar AI companion that understands a patient’s history, preferences, needs, and patterns, and a system that presents a new agent whenever a task appears. In his account, those two designs lead to very different relationships with technology, and therefore different odds that patients will engage, share information, follow recommendations, and remain connected to care.
For companies building clinical or administrative AI products, that argument shifts the commercial test. The question becomes not simply whether the system can complete a task safely in isolation, but whether it can support continuity strongly enough to become part of the care experience rather than a disposable interface layer.
What The Piece Signals
Because the article appears through the MedCity Influencers program, it is a perspective rather than reported consensus. Still, it captures an active fault line in healthcare AI strategy. One side treats narrow permissions and self-contained interactions as the safest default. The other argues that long-term value depends on monitored persistence, gradual expansion of responsibility, and a trust model that strengthens with use instead of resetting at every encounter.
Gekht’s bottom line is that isolating AI operationally may make it safer in a narrow sense, but it risks giving up the larger upside of a trainable and trusted agent. For healthcare decision-makers, the signal is that security architecture alone may not determine which AI tools stick; products that cannot build continuity with patients and workflows may struggle to justify deployment even if their access model is conservative.
