Morning: A Clinic That Is Already Awake
At 7:20 a.m., Dr. A. unlocks the door to her HIV clinic.
The waiting room looks the same as it has had for years: plastic chairs, public health posters curling at the edges, and a coffee machine that works intermittently. What has changed is that the clinic day has already begun. An agentic artificial intelligence (AI) system integrated into the health system’s infrastructure has spent the early morning assembling patient summaries, identifying gaps in care, coordinating outreach tasks, and flagging potential safety concerns.
Dr. A. does not think of this system as a replacement for her work. She thinks of it as an invisible junior colleague, one that prepares but does not decide. This distinction matters.
As AI evolves from single-purpose tools into agentic systems capable of perceiving data, reasoning across context, and acting within workflows, HIV clinicians face a defining question: Will they become passive users of automation, or active orchestrators who define boundaries, supervise decisions, and preserve human judgment?¹
Commentary: From Tools to Orchestration
Earlier generations of clinical AI solutions focused on discrete tasks, such as risk scores, deterministic alerts, and documentation aids. Agentic AI systems represent a structural shift. These systems coordinate multiple models, retrieve real-time data, generate plans, and execute tasks with varying degrees of autonomy.²
For HIV care, this shift is especially consequential. HIV medicine is longitudinal, data-dense, and deeply influenced by social context. As argued in The Infectious Diseases Orchestrator, AI literacy is now a core clinical competency. Clinicians must understand not only what these systems can do, but how to supervise them, where to limit automation, and how to ensure equity and safety.¹
First Patient: Prepared Without Judgment
Dr. A.’s first patient is M. Before the visit, the system has already assembled a longitudinal summary that includes viral load trends, antiretroviral therapy (ART) switches, missed appointments, emergency department visits, housing instability, and recent treatment for injection-related cellulitis. It has flagged inconsistent pharmacy refills and a missed viral load check.
Notably, the system does not label M. as “nonadherent.” That language was deliberately removed. Early predictive models often encoded deficit-based narratives that amplified stigma. In response, the clinic redesigned its AI outputs to highlight barriers rather than blame, using phrases such as “housing insecurity,” “untreated depression,” and “transportation gaps.”
- arrives late and guarded. He does not have a phone. He does not interact with patient portals, apps, or chatbots. Phones are lost, stolen, or broken. Digital access is unstable.
- tells her, almost casually, that he stopped coming to clinic years ago after being told he was “noncompliant.” He remembers the word. He remembers how it felt like a verdict rather than a description. Dr. A. does not correct him. She lets the silence do its work. This is the kind of damage no algorithm measures easily, but one that lingers long after a missed appointment.
Dr. A. listens as AI protected the time and cognitive space required to do motivational interviewing well. AI not only minimized the time Dr. A. spent on chart review and documentation, but it also transformed the traditional static appointment schedule into a dynamic slot, flexibly adjusting based on predicted case complexity and the overall clinic flow.
Commentary: The Digital Divide Is Not a Side Issue
Many people with HIV – particularly those experiencing housing insecurity, untreated mental illness, or substance use disorders – remain excluded from digital health innovations. Any AI strategy that assumes universal smartphone or broadband access will worsen disparities.
Agentic systems, if designed thoughtfully, can account for the digital divide. Instead of requiring M. to log in, the system can support outreach workers, coordinate shelter-based lab draws, and flag care gaps without demanding digital participation.
Second Patient: AI as the New “Dr. Google”
The next visit is different. A young woman on PrEP arrives with pages of notes, summaries generated by an AI chatbot she used the night before. The chatbot never interrupted her. It never appeared rushed. It never seemed skeptical. It reassured her about her symptoms, validated her concerns, and confidently explained potential risks. Some of the information is accurate, some is misleading, and all of it is delivered with confidence.
Dr. A. has learned that modern AI systems are optimized for helpfulness and affirmation. They are trained with reinforcement signals that reward agreement and reassurance.3, 4 That can support patient engagement or dangerously over-validate incorrect assumptions.
Rather than dismissing the chatbot, Dr. A. reframes it. “Let’s talk about which parts apply to you,” she says.
Commentary: Validation Versus Clinical Reasoning
Large language models generate plausible text, not clinical truth.5 They may hallucinate, overgeneralize, or miss rare but critical contraindications.6, 7 In HIV care – where drug–drug interactions, resistance patterns, and psychosocial context matter – this limitation is not theoretical.
Patient-facing AI is now a reality. Clinicians must be prepared to contextualize AI-generated information, explain its limits, and redirect patients toward evidence-based care. When handled well, AI can become a bridge to engagement rather than a competitor to the clinician–patient relationship.¹
Midday: Population Health Without Surveillance
Between visits, Dr. A. reviews the population health dashboard. The system continuously monitors viral suppression rates, overdue labs, PrEP follow-up, vaccination status, and sexually transmitted infection (STI) trends. It detects patterns that most individual clinicians rarely see, such as missed visits linked to a recent shelter closure, rising STI rates in a specific zip code, and ART interruptions tied to a temporary pharmacy supply disruption.
One alert catches her attention. Over the past two weeks, several patients who rely on the same community shelter have missed appointments and laboratory monitoring. The system does not label them as disengaged. Instead, it links the pattern to the shelter’s sudden relocation across town after a funding lapse, an event invisible in the electronic health record but captured through coordinated public health and community data feeds.
Dr. A. forwards the insight to the outreach team. No automated messages are sent to patients. No risk flags appear in the chart. The response is human: rescheduled labs, transportation vouchers, and a mobile clinic visit coordinated with the shelter’s new location.
Early versions of these systems caused harm. They flagged “high-risk” patients without explaining why. They optimized efficiency over fairness. They quietly reinforced structural inequities by treating social instability as individual failure.
Now, equity checks are embedded by design in the agentic system’s guardrails. Outputs are stratified. Context is surfaced. Disparities trigger human review, not automated action.
Commentary: Public Health Promise and Privacy risk
AI holds enormous promise for HIV population health and public health surveillance. It can support earlier intervention, smarter outreach, and more responsive systems of care. At the same time, it raises serious concerns about privacy, consent, and unintended surveillance, particularly for marginalized communities.8, 9
HIV medicine carries a long memory of harm, from criminalization statutes to breaches of confidentiality to surveillance practices that disproportionately targeted already stigmatized populations. Agentic AI systems, if left unchecked, could easily drift from supportive care coordination into coercive monitoring.
A model that predicts loss to follow-up could, in the wrong hands, become a justification for intrusive tracking. A system designed to detect transmission clusters could unintentionally expose individuals to legal or social risk if governance boundaries fail.
This is why clinician involvement in AI governance is essential. HIV specialists understand trust, stigma, and the historical consequences of surveillance in ways that cannot be encoded into software alone. AI governance is not optional oversight; it is clinical responsibility.¹
Afternoon: When AI Almost Fails
Late in the day, the system flags a potential ART–psychiatric medication interaction and drafts a recommended adjustment. Dr. A. pauses.
The recommendation is subtly wrong. Not catastrophically, but wrong enough to risk destabilizing a patient if accepted blindly. Dr. A. realizes that the real risk is not the error itself, but what would happen if she lacked the confidence and skill to question it. She corrects it. The system logs the correction. It learns. But it never decides alone.
Commentary: Hallucinations and Clinical Safety
Despite rapid advances, AI systems still hallucinate, misinterpret context, omit critical information, and fail in edge cases.5, 6 In clinical medicine, especially in complex HIV care, this remains a major safety concern.
As agentic AI systems mature, the most important safeguard is not better AI models or frameworks, it is preserved human oversight empowered by AI literacy. This requires a shift in mindset: Trust in AI is built through understanding a system’s limitations, while confidence comes from knowing how to safely interpret, question, and act on its outputs rather than defer to them.
In the near term, AI must remain advisory, auditable, and interpretable, with clear accountability for final decisions resting with clinicians. Human-in-the-loop design is not a temporary compromise, it is a foundational requirement for safe, trustworthy AI-enabled care.1, 10
End of Day: What Has Changed
As Dr. A. finishes her notes (drafted by ambient AI, edited by her), she reflects on the day. She spent less time clicking, less time searching, and less time remembering what machines remember better.
She spent more time listening, more time explaining, and more time practicing HIV medicine as a human discipline. AI did not replace her; it rearranged the work.
Conclusion: The Future Is Not Fully Digital
The future HIV clinic is not fully automated. It is not fully virtual. And it is not fully digital. It is digitally augmented.
HIV medicine has always required clinicians to navigate uncertainty, stigma, and rapidly evolving science. From the early days of the epidemic to the rollout of ART and PrEP, progress came not from passive adoption of tools, but from clinician advocacy, community partnership, and leadership grounded in trust. AI is simply the next test of that tradition.
Agentic AI systems can reduce burden, expand capacity, and support equity, but only if HIV clinicians remain active orchestrators rather than passive users.¹ This technology will arrive regardless. The outcome depends on who leads it.
If HIV specialists engage with AI literacy, governance, and design, these systems can support compassionate, equitable, human-centered care. If not, they risk reinforcing the very disparities HIV medicine has spent decades trying to dismantle.
The future is not written by algorithms alone; it is shaped by clinicians who choose to lead.
Dr. John Hanna is Associate Chief Medical Information Officer for Digital Health and Research Analytics at ECU Health and Affiliate Faculty at the Brody School of Medicine at East Carolina University. He completed the first integrated adult infectious diseases and clinical informatics fellowship nationally at the University of Texas Southwestern in 2023. Board certified in internal medicine, infectious diseases, and clinical informatics, he leads health system–wide efforts to integrate AI and digital health tools safely and thoughtfully into clinical care. His work focuses on AI maturity, advanced analytics, and expanding equitable access to telehealth across rural eastern North Carolina. His research and operational work center on advancing AI literacy and responsible AI deployment to improve outcomes while reducing disparities.