There's a particular kind of demo that has become very familiar in global health circles. Someone holds up a smartphone, points it at a patient or a scan, and within seconds the screen lights up with a diagnosis that used to require a specialist and a referral two towns away. The room gasps. The slide says "AI is transforming healthcare." And then everyone goes home, and in most clinics, nothing changes.

I don't say that to be cynical. The technology is often genuinely remarkable. The gap I want to talk about is the one between that dazzling demo and a tool that actually, reliably helps a nurse in a busy primary health centre on a Wednesday afternoon when the power is flickering and there are forty people waiting. That gap is enormous. And closing it has very little to do with the AI itself, and almost everything to do with implementation.

So let's talk honestly about AI in primary health care in low- and middle-income countries. What it can really do, where it keeps tripping over the same obstacles, and why the people who understand implementation science are quietly the ones who will decide whether any of this reaches a single patient.

First, the genuinely exciting part

It would be unfair to lead with the problems, because the promise here is real and it's worth being specific about. Primary health care in many LMIC settings runs on an impossible arithmetic: too few doctors, too few specialists, too few diagnostic machines, spread across too many people and too much distance. AI is interesting precisely because it can sometimes put a sliver of specialist-level capability into the hands of a frontline worker who would otherwise have nothing.

Take tuberculosis, which in the WHO African region is the second leading cause of death from a single infectious agent. One of the stubborn problems with TB is diagnosis: chest X-ray machines are expensive, radiologists are scarce, and patients frequently drop out of the process before they ever get a result. In a study presented in 2025, researchers tested an AI-guided lung ultrasound tool that runs on a smartphone-connected probe, used on just over 500 patients in Benin. The AI reached 93% sensitivity and 81% specificity, which actually clears the World Health Organization's benchmark thresholds for a TB triage test.

Read that again, because it's striking. A portable device, interpreted with help from an algorithm, performing at a level a minimally trained health worker could not achieve alone, in a setting where the alternative was often no test at all. That's not hype. That's a real capability that could matter for real patients.

And TB is just one example. Similar AI-assisted tools are being explored for reading X-rays, screening for diabetic eye disease, supporting community health workers through symptom triage on their phones, and flagging high-risk pregnancies. The common thread is the same: take a task that normally needs a scarce expert, and make it doable closer to where people actually live.

The promise of AI in primary care isn't replacing the expert. It's lending a little expertise to the worker who's already there.

Now, the part the demos skip

Here is where it gets real. An algorithm that performs beautifully in a validation study is not the same thing as a tool that works in a health system. Between those two states lies a long list of unglamorous, decidedly non-technical problems. And these are exactly the problems implementation science exists to take seriously.

Let me walk through the ones that come up again and again.

The device needs to actually function where it's deployed. A smartphone tool assumes a charged smartphone, which assumes reliable power, which is not a safe assumption in many of the places that need the tool most. It often assumes connectivity too, and a data bundle that someone has to pay for. A brilliant algorithm is useless if the phone is dead or the network is down on the day a patient walks in.

Someone has to trust it, and trust is earned, not installed. Picture a nurse who has run that clinic for fifteen years. An app now tells her a patient is high-risk, contradicting her own read of the situation. Does she believe the app? Should she? What happens to her authority, and her workload, and her liability if she follows it and it's wrong, or ignores it and it was right? These are human questions, and they decide adoption far more than accuracy figures do.

It has to fit the workflow, not fight it. Frontline workers are already stretched thin. If a tool adds five minutes and three extra steps to every patient, it will quietly be abandoned no matter how clever it is, because there is no slack in the day to absorb it. Tools that survive are the ones that slot into how the work already flows.

The data underneath it may not represent the people in front of it. Many AI models are trained on data from high-income settings, on populations that look nothing like the patients a rural clinic serves. A model that's accurate in one context can quietly underperform in another, and frontline staff have no way of knowing when that's happening.

And someone has to pay for it, indefinitely. A pilot funded by a grant is one thing. A tool that a ministry of health can sustain across thousands of facilities, with maintenance, updates, training, and subscription costs, year after year, is something else entirely. A great many promising health technologies die quietly at exactly this point.

The pattern worth noticing

Almost none of these obstacles are about the AI. They're about power, trust, workflow, context, and money. In other words, they're implementation problems wearing a technology costume.

This is implementation science's whole job

If you read our earlier piece on why good vaccines still miss children, this will feel familiar, because it's the same story in a new outfit. We had a proven intervention (the vaccine) that still failed to reach people because of everything around it. AI in primary care is the identical shape. The intervention works in principle. Whether it reaches anyone depends on the system it's dropped into.

Implementation science is the discipline that studies precisely that: how to take something that works and make it work in the real world, reliably, at scale, in messy and varied contexts. For AI in health, that turns out to be the entire game. The algorithm is the easy part now. The hard part, the part that decides success or failure, is the implementation.

What does an implementation lens actually add? It forces the right questions to the front of the queue, and forces them early, before the money is spent. Questions like these:

Notice that a developer obsessed only with model accuracy will not ask a single one of these. And that is exactly why so many technically excellent tools never make it past the pilot. The science of the algorithm was sound. The science of the implementation was never done.

The algorithm is the easy part now. Getting it used, trusted, and sustained is the hard part, and that is implementation science.

What "doing it well" looks like

The encouraging news is that we're starting to see what a more grounded approach looks like, and it's refreshingly humble. A 2025 examination of smartphone-based AI tools deployed with community health workers across Uganda, Rwanda, and Nigeria didn't conclude with "the model needs to be more accurate." It concluded with a set of distinctly human recommendations: invest in the community health workers themselves through proper training, fair compensation, and professional recognition; understand and align with the local digital ecosystem rather than fighting it; design the tool to fit the local context; and partner with people who deeply understand the health system you're entering.

Look at that list. Training. Fair pay. Local context. Good partnerships. There is barely any technology in it. It reads almost exactly like the old, unglamorous wisdom of strengthening primary health care, because that's what it is. The AI is new. The implementation challenges are the same ones global health has always wrestled with. The World Health Organization made a similar point in its guidance on AI for health, insisting that ethics, equity, and real-world governance be designed in from the start, not bolted on after the fact.

This is the quiet truth behind the loud headlines. The future of AI in primary health care will not be decided by who builds the cleverest model. It will be decided by who does the patient, unglamorous, deeply contextual work of fitting that model into a living health system, with real workers, real constraints, and real patients. That work has a name, and it isn't computer science.

Where this leaves you

If you work in or around primary health care, here's the part I most want you to take away. The arrival of AI does not make implementation expertise less important. It makes it dramatically more important. Every new tool is one more proven-in-principle intervention that will live or die on the strength of its implementation. The people who can think clearly about trust, workflow, context, and sustainability are about to be the most valuable people in the room, even if the room is full of engineers.

You don't need to learn to build an algorithm. You need to learn to ask whether it will survive contact with a real clinic, and to know what has to be true for it to help rather than hinder. That's a learnable skill. It's the same skill that decides whether a vaccine reaches a child or a new protocol changes anything at all. AI just raises the stakes and the urgency.

The demos will keep dazzling. The headlines will keep promising transformation. And the actual transformation, if it comes, will be built quietly by people who understood that the hardest problem was never the technology. It was always the implementation.