Recently, I cared for a woman at a Pittsburgh shelter clinic whose blood pressure was dangerously elevated. She told me she had insurance through a regional health plan but had not seen a primary care physician in quite some time. I prescribed losartan, a common blood pressure medication, and e-prescribed it to a downtown CVS pharmacy. A week later, she returned. She had not received the medication.

The pharmacy had told her they could not fill the prescription because it was “too old.” That explanation made little sense, so I did what physicians have done for decades when a medication problem arises: I called the pharmacy. Or at least I tried to.

Instead of reaching a pharmacist, I encountered an artificial intelligence (AI) phone system that required me to answer a series of verbal prompts. It asked me to verify a phone number associated with the clinic. Because I was calling from my personal cellphone while volunteering, the system would not recognize me. I tried again, using the clinic’s phone number. Still no success. After multiple attempts, the automated voice instructed me to leave a callback number and disconnected the call.

Unable to get a human being on the phone, I walked to the pharmacy with the patient. There, the real problem quickly became clear. The prescription had not expired. CVS simply was not in network for her insurance plan. She had one plan offered by her insurer, not another. One plan contracted with CVS; the other did not.

No one had explained that distinction to her. The patient was told that she should have received a notice in the mail. Back at the clinic, I called her insurer seeking guidance. I was told that patients are expected to know which pharmacies participate in their plans.

But many of the patients I care for often change addresses, experience housing instability, mental illness, language barriers, or limited health literacy. They may not receive mailed notices. They may not understand dense insurance documents. Even the most diligent patient can struggle because pharmacy networks and insurer contracts change from year to year.

In theory, technology is supposed to make health care more efficient. In practice, it often transfers work and confusion onto patients. Technology in health care assumes that the average patient has high literacy and a smartphone, can troubleshoot technology challenges, and can navigate apps and websites.

AI and automated customer-service systems are becoming ubiquitous in health care. We are told they will save time, reduce costs, and improve access. Sometimes they do. But too often, they are deployed in ways that save organizations money while making it harder for patients and clinicians to solve simple problems.

My patient’s challenge was not medical. We had an effective medication. We had a willing clinician. We had a nearby pharmacy. What we lacked was a straightforward way to identify a bureaucratic obstacle and speak with someone who could help remove it.

The irony is hard to ignore. Health care organizations frequently champion innovation, digital transformation, and patient-centered care. Yet many patients and clinicians increasingly find themselves trapped in endless phone trees, automated chat systems, and AI gatekeepers designed to prevent access to human assistance.

For patients with stable housing, reliable internet access, and ample time, these obstacles may be annoying. For vulnerable patients already struggling to navigate a fragmented health care system, they can become the difference between receiving treatment and going without it.

Technology should support human care, not replace it where human judgment is essential. We should absolutely use AI when it improves efficiency and expands access. But we should also recognize its limits. No algorithm would have known that my patient never received important insurance information, that she was obtaining care at a shelter clinic, or that a 1-minute conversation with a pharmacist could have resolved the problem immediately.

The question is not whether AI belongs in health care. It already does. The question is whether we are willing to preserve meaningful paths to human help when technology fails. For the sake of our most vulnerable patients, we must.

In the interest of full disclosure, AI helped polish the prose of this essay; it did not generate the experiences, observations, or opinions expressed within it. There lies a bit of irony.

Thuy D. Bui is an internal medicine physician.