As a frontline physician in South Sudan, I see approximately two hundred patients on call day across two public hospitals. Discussions about agentic AI often assume that every autonomous system will operate under the supervision of experienced specialists. That assumption reflects the reality of high-income health systems, but not mine.
In our pediatric referral hospital, the hematology-oncology service is staffed by only two hematologists. In the adult referral hospital where I work, I have never worked alongside a medical oncologist or a radiologist. Patients with advanced cancers are frequently referred directly to general surgeons because there are no oncology services available. Surgery may be performed, but chemotherapy is often unavailable, leaving patients without the comprehensive cancer care that modern medicine considers standard.
Diagnosis itself can become an insurmountable barrier. A biopsy is often too expensive for families who must pay entirely out of pocket. Some patients return to their villages to sell livestock or ask their extended families to contribute money for a single diagnostic procedure. By the time they return to the hospital, weeks or months later, the disease has frequently progressed to an advanced stage. Delayed diagnosis becomes delayed treatment, and delayed treatment becomes avoidable mortality.
These realities expose a critical blind spot in today’s conversation about agentic AI. Most governance frameworks assume that an expert clinician will review, challenge, or override the AI’s recommendations. But what happens when there is no specialist to provide that oversight?
Real physician voices, twice a week
Free, and one click to unsubscribe.
The places that stand to benefit most from autonomous AI are often the least able to supervise it.
Clinical documentation illustrates this challenge. In our hospitals, medical records are still entirely paper-based. Clinicians care for overwhelming numbers of patients while documenting every encounter by hand. Every minute spent writing is a minute taken away from direct patient care. The result is clinician burnout, incomplete documentation, and reduced time with patients.
These experiences inspired me to begin building an offline-first AI clinical documentation assistant. The goal is not to replace clinicians, but to reduce documentation burden, improve the quality of clinical notes, and return valuable time to patient care. Designing the system to work offline is essential because reliable internet access cannot be assumed. Even for unsalaried intern doctors, the cost of mobile data is a significant barrier to using cloud-based AI tools.
The consequences of limited digital infrastructure extend beyond the hospital walls. Many patients travel long distances carrying only paper referral letters. Some lose them during the journey. Others arrive with documents damaged by rain or water. In some communities, cultural beliefs lead families to discourage patients from seeking cancer treatment, believing illness is caused by witchcraft rather than disease. When clinical information exists only on paper, continuity of care becomes fragile.
The supervision gap becomes even more apparent in rural communities. My home area has a population of approximately forty thousand people without a single medical specialist. Health care is delivered primarily by clinical officers with three years of training who often work alone, particularly during nights and weekends, without immediate specialist support.
This raises a question that deserves far greater attention than whether agentic AI can make autonomous decisions: How do we build agentic AI that remains safe when there is no expert available to supervise it?
The future of agentic AI will not be determined solely by increasingly capable models. It will depend on whether we design systems that acknowledge the realities of resource-constrained health systems. Intelligence alone is not enough. Agentic AI must be built for environments where specialists are scarce, connectivity is unreliable, paper records remain the norm, and supervision cannot be taken for granted.
If we fail to address that reality, the greatest blind spot in agentic AI will not be the technology itself. It will be our assumption that someone is always there to supervise it.
This essay is cited in the KevinMD record on artificial intelligence.
Buga Charles George Kenyi is a physician in South Sudan.
