AI decision support is already transforming care in high-income countries. But the trains only go as far as the tracks beneath them. A Harvard Medical Student working with Better Evidence asks what it will take to bring AI’s promise to clinicians in low- and middle-income countries.

For four months, a 51-year-old woman in Tanzania had a swelling abdomen, pain and weight loss. One hospital evaluated her for an ovarian tumor, but tests were inconclusive. At the next facility, a gynecologist repeated the workup, attributing her symptoms to ascites of unknown cause. Searching UpToDate, a clinical decision support tool, the gynecologist came across a diagnosis that matched her symptoms: peritoneal tuberculosis. Exploratory surgery revealed normal ovaries and biopsies confirmed TB. She started treatment and was well six months later.

This patient reached the health system twice before receiving the right diagnosis. Of the 8.6 million deaths each year in low- and middle-income countries (LMICs) that good care could have prevented, an estimated 5 million occur among people who, like her, reached the health system without receiving quality care. Clinicians in these settings face overwhelming patient volumes and scarce specialist support, with little time to keep up with rapidly evolving evidence.

In US hospitals, UpToDate use has been associated with better outcomes, but cost puts it out of reach in LMICs. Better Evidence at the Global Health Delivery Project, where I am spending a research year, has offered clinicians in LMICs free UpToDate subscriptions since 2009. Its research showed that clinicians have used the tool heavily, but access alone was not enough: clinicians who struggled to fit the tool into their workflow were half as likely to use it.

Generative AI is changing how clinicians find clinical information, and in high-income countries, the trains are already running. In the US, monthly users of one AI platform grew from a few hundred in early 2024 to over 700,000 by mid-2026. In a Better Evidence survey, 93% of respondents said they were willing to learn how AI could benefit their practice. At Penda Health clinics in Nairobi, clinicians using an AI safety net built into their electronic records made 16% fewer diagnostic errors and 13% fewer treatment errors and a later randomized trial confirmed better treatment plans.

But a train is only as good as the track beneath it. I keep returning to one question: what tracks need to be laid for AI to deliver on its promise in LMICs?

The first track is infrastructure. The nature of AI tools means that each clinician question depends on power, a device and internet connection, which are often missing at the point of care. In sub-Saharan Africa, 15% of health facilities have no electricity at all, and only half of hospitals have a reliable supply. This presents a huge obstacle for AI tool deployment but also an opportunity for developers to innovate. Crane AI Labs in Uganda has built EaseHealth to run offline on a phone, though a pilot showed that most workers’ phones lacked the memory to run their software. Funders, in partnership with governments, need to invest more in the power, connectivity and devices needed for AI-empowered care.

The second is local context: tools must reflect local guidelines and medicines lists, and fit into the workflow on the ground. Many of the tools widely used were developed in high-resource settings where guidelines are digitized and frequently updated, and electronic medical records are robust enough for AI tool integration. But this is often not the case in LMICs. The WHO developed the SMART Guidelines to help countries translate their guidelines into a form digital systems can use, and five African countries have already built adapted versions into systems they use. The previous findings from Better Evidence’s UpToDate work and Penda Health’s AI safety net also emphasize the value of understanding clinician workflows and folding these tools into them. Implementers, with continued guidance from the WHO and health ministries, need to drive the incorporation of local guidelines and medicines lists, and be intentional about integration, avoiding parallel systems that only complicate care.

The third track is regulation. A review of 12 African countries found none had AI-specific health regulation, and in 2020 only half of African countries had data protection laws. AI tools deployed within a scanty regulatory landscape leave questions of accountability and data ownership unanswered. However, frameworks are emerging. WHO’s 2024 guidance calls for independent audits of AI tools deployed at scale, and the African Union adopted a continental AI strategy the same year. National regulators now need the mandate and capacity to validate tools, protect patient data and decide who is accountable when AI-informed care goes wrong.

The reality is that even well-laid tracks will not replace the basics. Although UpToDate helped the Tanzanian gynecologist find the right diagnosis, the patient recovered because she received surgery, a lab confirmed TB and the right drugs were available. AI will close knowledge gaps but not supply gaps. But if we lay the right tracks, AI could carry far more patients like her to the right care the first time.