Physician practices are drawing a hard line between administrative automation and clinical judgment
Almost 70% of consumers will abandon an AI customer service tool after two failures, and in health care that tolerance is even thinner when the workflow touches symptoms, scheduling, consent, or billing accuracy. Juan Jaysingh, CEO of Zingtree, argued that the central question for physician practices is not whether AI should be used, but where it should stop. In his view, the safest applications are the repetitive, low-risk tasks that consume staff time without requiring clinical judgment: summarizing a call, gathering basic context, or routing routine requests.
The dangerous work begins when a chatbot or voice agent moves from convenience to decision-making. A patient with urgent symptoms placed into a routine appointment slot can face real harm. A minor scheduled for physical therapy without a guardian’s approval creates a compliance problem before a clinician ever enters the room. Those examples reveal why health care automation cannot be treated like retail support or travel booking; the consequences of a wrong answer are immediate, and in some cases life-threatening.
Where AI fits, and where it should stop
Jaysingh framed the division in practical terms: AI can collect, organize, and summarize, but the final handling of mission-critical workflows still belongs to a human. That means a practice can use automation to identify why a patient called, capture intake details, or prepare a structured summary for staff. It should not allow AI to independently decide whether a symptom cluster belongs in a routine slot, whether a referral must be escalated, or whether a consent-related issue can be waved through.
The logic is straightforward. Systems that manage front-end communication are only useful if they preserve the integrity of the clinical workflow behind them. In ambulatory medicine, a bad routing decision can delay diagnosis, create frustrated patients, and generate downstream rework for staff who must correct the error. Even in nonclinical interactions, such as insurance questions or billing follow-up, AI should be limited to the pieces it can perform with consistency and fidelity. Once the task crosses into areas where exactness matters every time, the machine should yield.
A brief sentence captures the boundary: if an error could alter care, AI should not be left alone.
Hand-offs fail when the context disappears
The most common breakdown, Jaysingh said, happens at the moment of transfer from AI to a human agent. Patients often repeat the same information after the handoff, which destroys trust and makes the tool feel like an obstacle rather than an aid. In a health care setting, that repetition is more than an annoyance. It can mean a patient relives a symptom story, recites insurance details twice, or rehashes a scheduling problem that should have been settled in one pass.
The handoff problem is not only about convenience; it is about continuity. If the AI gathered the intent of the call, the symptoms described, and the patient’s immediate need, all of that context should follow the patient to the human staff member. A “human on the loop” model, as Jaysingh described it, keeps a person available to step in when accuracy becomes essential while preserving the work AI has already done. That structure lets the human correct edge cases and approve sensitive decisions without forcing the patient to start over.
For front desk teams, the design standard is simple: the patient should never feel like they have been reset to the beginning of the conversation.
Compliance, audit trails, and the responsibility gap
Health care organizations cannot outsource accountability to the vendor if an AI system gives the wrong guidance. Jaysingh placed responsibility squarely on the organization that chooses to deploy the tool, whether it is built internally or purchased from outside. That position reflects the operational reality of medicine: practices own the workflow, the rules, and the clinical and regulatory consequences when those rules are broken.
Before a practice lets AI speak to patients, it should demand HIPAA and SOC 2 alignment from the vendor, along with hard-coded business rules and an audit trail that shows how the system behaved in live use. Certification alone is not enough. The rules that govern scheduling, consent, patient routing, and follow-up must be embedded into daily operations, especially when a workflow requires accuracy every single time. After deployment, the system has to be monitored to confirm that it continues to follow the intended process.
That requirement is especially important because AI errors may not look dramatic in the moment. A scheduling miss can appear routine until it sends the patient to the wrong clinic. A consent problem may remain hidden until staff discover missing authorization paperwork. An audit trail gives the organization a way to see where the system is failing and correct it before those mistakes become routine.
Consumer patience is shrinking fast
The survey data in the interview point to a narrow margin for error. Once an AI tool fails twice, most users give up on it. In health care, that reaction is likely to be more intense because patients often arrive stressed, confused, or worried about symptoms. A tool that loops, misroutes, or ignores context does not just frustrate the user; it can push the patient toward a human workaround before the system has a chance to prove itself.
That has direct operational implications for physician practices. AI adoption will stall if patients experience the same dead-end behavior they already associate with clunky phone trees and unhelpful chatbots. Practices that want to preserve trust have to choose narrow, structured use cases first and reserve human involvement for the situations where the clinical or compliance stakes are high. The first call is the proving ground. If the workflow cannot get that right, it will struggle to earn any broader role in the practice.
Scheduling, symptoms, and other high-risk decisions
Scheduling may look simple from the outside, but in medicine it can become a triage decision. If a patient reports symptoms that require urgent evaluation, the system has to route them correctly the first time. If it is determining whether a patient belongs in a specific clinic, or whether a guardian must be involved before a procedure is booked, the margin for error disappears. Those workflows are not ideal candidates for unsupervised AI because even a small inconsistency can carry a large clinical cost.
The same applies to intake and post-visit communication. AI can help collect history, organize questions, and manage routine billing or insurance interactions. It should not be allowed to improvise when the answer depends on medical history, symptom severity, age-related consent rules, or clinic-specific protocols. The safest deployments are the ones that make the human staff faster without making them less necessary.
Practices evaluating these systems will need to ask a blunt question: can this tool fail safely, every time, in the exact workflow we are asking it to perform? If the answer is no, the task still belongs with a person.