Last week OpenAI published a field study with Penda Health, a primary-care network in Nairobi, exploring an "AI clinical copilot."* The system—built on GPT-4o—quietly double-checks diagnoses and treatments in real time, surfacing suggestions only when it spots a possible error. Penda reports fewer diagnostic and treatment mistakes when the copilot is active and, just as important, solid adoption among clinicians who helped design the workflow.
We can treat those numbers as early signals, not final verdicts. But they raise a set of questions that feel worth asking—especially for teams, like ours at heva, that serve busy healthcare providers in emerging markets.
What might an always-present second opinion unlock?
Cognitive breathing room
Many clinicians in lower-resource settings see dozens of cases a day across nearly every body system. An assistant that scans for overlooked labs or contradicting symptoms could function as a mental guardrail, freeing scarce attention for patient rapport and education.
Faster guideline diffusion
Local protocols change; international best practices evolve. A live copilot that references current guidance—adapted to regional epidemiology—might shorten the lag between new evidence and real-world use.
A gentler entry point for AI
The Penda pilot emphasised that the clinician stays in charge. Alerts are suggestions, not commands. That framing may lower resistance among professionals who worry that algorithms will one day dictate care.
The cautions that follow close behind
Implementation is half the battle
Penda spent months embedding the tool into existing visit flows. Without that human-centred fit, even a brilliant model can gather digital dust.
Local context is non-negotiable
A copilot trained on U.S. data but deployed in Kenya—or Mexico, or the DR—must learn local disease prevalence, drug formularies, and cultural norms to avoid new forms of bias. This is especially critical for international patients seeking care abroad.
Trust takes longer than a pilot
Early quality gains look promising, yet sustained trust will hinge on transparent audits, ethical review, and clear lines of accountability.
Where this intersects with our own work
heva is not building diagnostic AI today; we are focused on the administrative spine—scheduling, intake, payments—that supports cross-border care in Latin America. Still, three threads from the Penda study map directly to challenges we see daily:
Workflow first, tech second
Our payment agent only gained traction once it mirrored how front-desk staff already triaged inbound messages. The same principle applies to clinical copilots: fit beats flash.
Emerging markets as bellwethers
Kenya's clinics became a proving ground precisely because the stakes are high and practitioner time is limited. We hear similar constraints from surgeons in Mexico City. Low "addressable market" stereotypes miss the point; constraint breeds innovation. This is why heva's mission focuses on emerging markets.
Shared burden, shared reward
If a copilot can trim even a fraction of diagnostic uncertainty, it could pair nicely with the hours we save on admin tasks—together widening the margin for face-to-face care.
What happens next—for us
-
Keep refining the "non-clinical" pieces of the puzzle—intake, payments, documentation—while watching how diagnostic copilots mature.
-
Explore small, localized, administrative aids that complement, not compete with, clinician judgement.
-
Stay in close conversation with the providers who will feel every success or failure first.
If you are experimenting with AI at the point of care—whether in Nairobi, Guadalajara, or New York—we would love to compare notes. What has surprised you so far? Where do you see the biggest gaps?
References
(All observations here are exploratory; none constitute medical or performance claims.)
Related Articles
