Healthcare providers have deployed AI agents to handle patient communications at scale, but three in four organizations running these agents have already been forced to shut one down or roll one back due to governance failures. The leading cause, in a third of those cases, is personal identifiable information surfacing where it shouldn't-a risk that carries an average breach cost of $6.64 million in healthcare, the highest of any industry for 13 years running.
The findings come from a new Sinch report that surveyed over 470 healthcare leaders globally about their use of AI in patient communications. Fifty-five percent of healthcare organizations now have AI agents in production handling real patient interactions, though the industry's deployment rate sits seven points below the overall average across sectors.
What healthcare wants AI agents to do
Healthcare generates exactly the kind of high-volume, repetitive communication that AI agents were built to handle. Providers send appointment reminders, prescription pickup notifications, test result alerts, and follow-ups after missed visits-work that easily outpaces what a call center can manage.
Twenty-two percent of healthcare leaders see outbound voice agents for appointments, reminders, and proactive outreach as the biggest voice AI opportunity, five points above the overall average. Identity verification is another priority: 41% of respondents put it near the top of the list for AI agents to take on, since patients must prove who they are before accessing test results or confirming prescriptions.
Healthcare is among the most likely industries to plan investment increases of 50% or more in AI agent-powered customer communications this year, at 16% against 13% overall. The ambition points at patient engagement at scale, but getting into production has been slower here than elsewhere.
Why healthcare moved slower into production
Patient consent requirements, clinical data protections, and privacy laws raise the bar higher than most industries face. Only 37% of healthcare organizations report clear guidance on AI disclosure, one of the lowest rates in the study. That uncertainty shows in confidence levels: just 42% of healthcare leaders call themselves very confident about deploying at scale.
Yet the research finds that once a healthcare organization is in production, it tracks with every other industry across all 19 compliance and privacy measures studied. The failure rate is only one point higher than the overall average. Compliance pressure slowed the path to production, but once live, the industry's disadvantage becomes a nonissue. Something else is triggering the failures.
The disconnect between leadership and builders
One problem the research uncovered is that technical leaders consistently report rollbacks at a higher rate than business leaders within the same organizations. In healthcare, 36% of C-suite leaders report fully mature guardrails, against 18% of the directors building them. A problem that looks smaller from the top gets funded like it's smaller, so the work that would prevent the next rollback doesn't happen.
A similar visibility gap applies to cost expectations: 34% of C-suite executives in healthcare expect AI to cut costs by more than half, a number that drops to 12% among the directors responsible for delivering it.
"As a C-level leader, you're often not involved in every detail of making technology, but that's precisely the point," said Stefan Wenzel, CPO, SAP Engagement Cloud. "Executives need to operate at a different altitude, with a longer horizon. You see how the technology behaves, and more importantly, you can anticipate where it's heading and what it will unlock next."
Infrastructure is what's actually breaking
The Sinch study found that the strongest predictor of AI deployment success across industries is the communications infrastructure running underneath the agent. In healthcare, that's exactly where the failures trace back to. A third of healthcare's rollbacks were triggered by data leakage or PII exposure-test results sent to the wrong patient, or a medication record pulled into a message that was only supposed to confirm an appointment time.
These are infrastructure failures. The data surfaced because the platform didn't stop it. PII exposure originates a layer below the rule book, in the infrastructure the agent runs on. The fix healthcare providers are betting on reflects this diagnosis: trust, security, and compliance top their investment priorities at 75%, and communications infrastructure follows at 64%-three points above average and six points above AI agent development itself.
What healthcare respondents value most from a partner all points to the platform underneath the agent: reliability, compliance, and accountability. This is a specification for better infrastructure. From the patient's side, it means the right messages get delivered when they should, and sensitive information stays protected. For organizations exploring AI for Healthcare, the lesson is that the agent is only as secure as the pipes it runs on.
Why this matters for healthcare professionals
The report makes clear that AI governance failures in patient communications are not rare edge cases-three-quarters of organizations running AI agents have hit one serious enough to require a rollback. For clinical and operational leaders, the immediate takeaway is that production readiness depends less on the AI model itself and more on the communications infrastructure controlling what data reaches which patient. When infrastructure gaps let PII leak into the wrong message, the cost is measured in breach penalties and broken patient trust. The organizations funding infrastructure over agent development are betting on the right layer.
Your membership also unlocks:
