In the rapidly evolving landscape of digital health, a fundamental debate has emerged regarding how healthcare organizations should deploy agentic artificial intelligence. The conversation, recently sparked by Imprivata CEO Fran Rosch in MedCity News, posits that the safest way to interact with AI agents is through a "transactional and self-limiting" framework. Rosch’s perspective, rooted in the rigorous demands of cybersecurity and identity management, suggests that AI agents should be treated with the same caution as contract labor—constantly monitored, restricted in access, and never fully trusted.
However, this "zero-trust" approach, while technically sound from a firewall-centric perspective, ignores the clinical and relational realities of modern healthcare. As we integrate AI into patient care, the industry faces a critical crossroads: do we treat AI as a transient, untrusted tool, or as a persistent, evolving partner in the clinical journey?
The Imprivata Thesis: Cybersecurity First
To understand the current tension, one must first examine the argument presented by Imprivata. As a leader in identity security, the company’s mandate is to protect the integrity of the healthcare ecosystem. From this vantage point, an AI agent is a potential vector for unauthorized data access or malicious manipulation.
Rosch’s analogy comparing AI agents to contract nurses is illustrative of a risk-averse security culture. By limiting the scope of an agent’s interaction—treating it as a "temp" that handles a single task and is then terminated—organizations can theoretically isolate potential failures. In this model, the AI is never granted "keys to the kingdom." It is a siloed operator, minimizing the blast radius if the agent hallucinates, malfunctions, or becomes a security liability.
While this approach is undeniably effective at mitigating short-term cyber risk, it introduces significant operational friction. If every interaction is a "first meeting" between the patient and the AI, the continuity of care is fractured.
The Case for Continuity: Why "Temp" AI Fails Patients
The primary flaw in the "transactional" model is that it treats healthcare as a series of isolated data exchanges rather than a continuous, longitudinal relationship. In clinical practice, the efficacy of a treatment plan is rarely defined by a single interaction; it is defined by the sustained engagement between the care team and the patient.
The Role of Longitudinal Learning
Just as a human clinician benefits from understanding a patient’s medical history, preferences, and communication style over time, an AI agent’s utility grows through experience. When an AI is allowed to function longitudinally, it can:
- Adapt to Workflows: Learn the nuances of a specific clinic’s EHR documentation style or a patient’s preferred method of communication.
- Identify Patterns: Recognize early indicators of non-adherence or worsening symptoms that a "one-off" interaction would miss.
- Drive ROI: By reducing the need for repetitive onboarding and manual data entry, a persistent agent creates compounding value rather than redundant, low-level tasks.
The Human Element: Trust as a Clinical Metric
In healthcare, trust is not an abstract concept—it is a clinical outcome. Patients are often dealing with vulnerable, high-stress situations. When they interact with a digital tool, they are looking for stability. An "adopt and drop" model forces patients to repeatedly re-introduce themselves to new systems.
Conversely, a virtual healthcare companion that persists over time fosters a sense of familiarity. This is particularly vital for chronic disease management. A patient who interacts with a consistent digital interface that understands their history is significantly more likely to share sensitive information and follow through on complex treatment plans than one who is subjected to a "rotating cast" of AI agents.
Chronology of the Debate: From Efficiency to Empathy
The evolution of this debate follows the broader trajectory of AI integration in medicine:
- Phase I: The Efficiency Era (2020-2023): Healthcare systems focused on AI for basic automation—scheduling appointments, routing calls, and basic triage. In this phase, the transactional, "temp" model was the gold standard because the AI’s capabilities were limited and the primary goal was administrative throughput.
- Phase II: The Agentic Revolution (2024-Present): With the rise of Large Language Models (LLMs) and agentic frameworks, AI has moved from simple automation to complex, reasoning-based support. As AI begins to assist in clinical decision support and personalized patient education, the limitations of the "transactional" model have become increasingly apparent.
- Phase III: The Trust & Security Synthesis (The Future): We are currently entering a phase where the industry must reconcile the security requirements of cybersecurity firms like Imprivata with the clinical necessity for long-term patient engagement.
Implications for Healthcare Strategy
If we move away from the "isolate and terminate" model, how do we build systems that are both secure and clinically effective?
1. From "Isolation" to "Earned Trust"
Rather than keeping AI in a box, organizations should implement a tiered trust model. New agents start with restricted access, but as they demonstrate accuracy, safety, and reliability, their permissions and contextual memory are expanded. This mirrors how a hospital credentials a new employee. We do not treat a resident physician like a temporary day laborer; we supervise them, observe their performance, and grant them more responsibility as they prove their competence.
2. The Cybersecurity/Clinical Paradox
Cybersecurity professionals often see risk in terms of exposure, whereas clinicians see risk in terms of poor outcomes. If an AI is restricted from knowing a patient’s history, the risk of a "bad outcome" (e.g., missed medication, incorrect triage) actually increases. Therefore, the most secure system is not the one with the fewest permissions; it is the one with the most accurate, context-aware, and monitored capabilities.
3. Regulatory and Ethical Considerations
Trust is a two-way street. As AI becomes more integrated into patient care, the ethical responsibility of the developer and the provider grows. Transparency becomes paramount. Patients must know when they are interacting with an AI, what that AI knows about them, and how that information is being protected. Building trust requires that we be open about the AI’s capabilities and its limitations, rather than hiding it behind a veil of transactional anonymity.
Official Responses and Industry Outlook
The industry is clearly divided. Cybersecurity vendors argue that the threat landscape is too volatile to allow for persistent, broad-access AI agents. They point to the rise in data breaches and the inherent difficulty in "patching" an AI that has learned biased or dangerous behaviors.
On the other side, clinical innovators, such as Dr. Gennady Gekht of Hygia, argue that the "transactional" approach is a form of professional stagnation. By prioritizing the lowest common denominator of security, we are failing to leverage the transformative potential of AI to support patients who are chronically ill, isolated, or underserved.
Conclusion: Reframing the Goal
The goal of AI in healthcare should not be to build a "safe" robot that does nothing of consequence, but to build a "trusted" companion that improves clinical outcomes. The challenge we face is not primarily technological; it is philosophical. We must decide whether we value the illusion of safety provided by operational isolation, or the measurable improvement in patient health that comes from consistent, long-term, and secure human-AI collaboration.
The "baby and the bathwater" metaphor is apt. In our rush to secure the healthcare environment against the risks of AI, we must be careful not to discard the very features—continuity, context, and engagement—that make AI a valuable tool for patient care. The future of healthcare innovation lies not in how quickly we can cut an AI agent off, but in how safely we can allow it to earn our trust.
By shifting our focus from operational isolation to verified trustworthiness, we can create a healthcare environment where AI is not just a digital tool, but a reliable, secure, and essential member of the care team.
