AI in Healthcare in South Africa: Beyond the Chatbot
- Matthew Hellyar
- Aug 9
- 11 min read
A
rtificial intelligence has entered healthcare at an unusual moment.
Medicine has never generated more information, yet the people responsible for interpreting that information have never been under greater pressure. A single patient can move through consultations, pathology, imaging, prescriptions, referrals, hospital admissions and specialist reviews, leaving behind a growing clinical record at every stage. At the same time, doctors are being asked to see more patients, complete more documentation, navigate more software and make decisions within increasingly complex systems of care.
Into that environment has arrived a new generation of artificial intelligence capable of producing remarkably fluent medical language.
The possibilities are difficult to ignore.
A consultation can be transformed into a structured clinical note in seconds. A long medical document can be summarised almost immediately. A physician can ask a question in ordinary language and receive a sophisticated response without navigating a traditional software interface. Reports that once required repetitive manual drafting can increasingly be prepared with the assistance of intelligent systems.
This represents genuine progress.
It also creates a risk that healthcare mistakes the most visible expression of artificial intelligence—the chatbot—for the final form of the technology.
It is not.
The deeper opportunity for clinical AI is not simply to create better conversations between doctors and machines. It is to build better connections between clinicians and the enormous amount of information surrounding their patients.
That difference has shaped the way we think about artificial intelligence at Respocare.
Despite the name Respocare Connect AI, we do not think of what we are building as an AI product in the conventional sense. Artificial intelligence is an enabling technology. It is extraordinarily important, but it is still only one part of the system.
The service exists in what is built around that intelligence: the patient record, the clinical context, the retrieval of relevant information, the workflows, the governance, the traceability of patient-specific claims and, ultimately, the medical professional who remains responsible for the care of the patient.
As artificial intelligence becomes more deeply embedded in healthcare, this distinction will matter more than ever.
The first generation of clinical AI solved the obvious problem (AI in healthcare South Africa)
Documentation was an obvious place for generative AI to begin.
Few clinicians entered medicine because they wanted to spend large portions of their working day typing notes, recreating histories, preparing referral letters or converting a consultation into several different administrative formats. Documentation is essential, but the burden associated with producing it has become one of the most visible frustrations in modern clinical practice.
The medical AI scribe is therefore an important development.
It takes something that already happens—the clinical consultation—and helps turn it into structured documentation. The clinician speaks, the system listens, and a draft is prepared for review.
That can return meaningful time to a medical professional.
But the consultation taking place today is rarely the whole clinical story.
A patient may have presented with the same symptom four months ago. A medication may have been started by another clinician and subsequently discontinued. A laboratory abnormality may have been slowly evolving across several sets of results. A specialist may have raised a concern in a letter that is now buried among dozens of other documents. Treatment may have been tried previously and failed.
A scribe can tell us what happened in the room today.
Clinical intelligence must eventually help us understand how today relates to everything that happened before it.
That is a much more difficult problem.
It is also a far more interesting one.
Medicine is longitudinal
Healthcare is often documented as a series of encounters, but patients do not experience illness that way.
Disease develops through time.
Symptoms emerge, disappear and return. Investigations change. Treatments are introduced, adjusted or abandoned. Diagnoses become more certain or less certain as new evidence appears. A clinical decision that seemed entirely reasonable in January may look very different when viewed alongside the information available by June.
This temporal dimension is fundamental to medicine.
A haemoglobin result on its own is a number.
Five haemoglobin results across six months are a trajectory.
A medication list tells us what a patient may be taking today.
A longitudinal medication history can tell us what was tried before, what changed, what failed and what happened afterwards.
A diagnosis written in the latest consultation may appear definitive.
Placed beside previous investigations, specialist correspondence and treatment response, it may become something that deserves another look.
This is why the clinical record has such an important role in the future of healthcare AI.
The question is no longer simply whether an artificial intelligence model can answer a medical question.
The more meaningful question is whether an intelligent clinical system can find the right patient-specific information, understand where it sits in time, distinguish evidence from assumption and present that information back to the medical professional in a form that is genuinely useful.
That requires far more than a chatbot.
It requires architecture.
The patient record is where intelligence becomes clinically meaningful
Consider two questions.
The first is: What are the common causes of progressive anaemia?
The second is: How has this patient's haemoglobin changed over the last year, what treatments were introduced during that period, and what happened after each intervention?
Both are medical questions.
Only one can be answered predominantly from general medical knowledge.
The second requires access to the patient's history.
The system has to locate the relevant results. It must understand their chronology. It may need to identify medication changes, previous diagnoses, consultation notes or specialist correspondence. It then needs to bring those pieces together without quietly inventing information that is absent from the record.
This is where artificial intelligence begins to change character.
It stops being only a generator of language and becomes a way of working with clinical information.
Healthcare already has a great deal of data. The greater challenge is often turning that data into usable context at the precise moment a clinician needs it.
The relevant information may already exist somewhere in the record. The doctor may simply not have ten minutes to find it during a consultation.
That problem becomes increasingly important as clinical records grow.
Digital healthcare solved an important problem by making information storable and retrievable. The next generation of healthcare technology has an opportunity to make that information easier to understand.
An electronic record can hold the information.
A clinical intelligence layer can help the medical professional work with it.
This changes what good clinical AI should look like
Once artificial intelligence begins operating against patient-specific information, the standard has to change.
Fluent language is no longer enough.
A system can produce an answer that sounds clinically sophisticated and still be wrong. In ordinary consumer technology that may be frustrating. In healthcare it can be consequential.
Clinical AI therefore needs a relationship with evidence.
When a doctor asks a patient-specific question, the system should first determine what relevant information is available. It should retrieve that information, preserve the distinction between what is documented and what is inferred, and use the evidence as the basis for its response.
In simple terms:
Find the evidence first. Reason second.
That principle is central to the architecture we are building at Respocare Connect AI.
It is also why we believe the conversation about artificial intelligence in healthcare cannot remain focused solely on which underlying model is the most powerful.
The model matters enormously. But so does everything around it.
How is clinical information organised?
How does the system decide what information to retrieve?
Can the clinician trace an important statement back to the underlying record?
What happens when the record contains conflicting information?
What happens when information is missing?
Does the system know the difference between “this is not documented” and “this never happened”?
How is the output reviewed?
What actions are recorded?
Who remains responsible for the clinical decision?
These questions are not secondary to the technology.
In healthcare, they are part of the technology.
A doctor should not have to become a prompt engineer
One of the stranger consequences of the generative AI revolution has been the rise of prompt engineering as a skill expected of ordinary users.
There is value in learning how to communicate clearly with an intelligent system. But healthcare should be careful about shifting the burden of poorly designed technology onto clinicians.
Doctors have spent years learning medicine.
They should not need another qualification in how to phrase questions for a language model.
A well-designed clinical system should understand the task from the clinical workflow surrounding it.
If a clinician needs a referral letter, the technology should know what kind of information is normally relevant to that task.
If a doctor asks for a longitudinal summary, the system should know that chronology matters.
If a medical professional asks when a medication was started, the system should search the authorised patient history rather than require the doctor to manually find the correct document and paste it into a chat window.
If the necessary information is unavailable, the technology should be capable of saying so.
This is where agentic AI becomes particularly important.
The word “agentic” is becoming fashionable, and as often happens with new technology, it risks becoming so broadly used that it loses its meaning. Its real significance is not that an AI system suddenly becomes autonomous in some science-fiction sense. It is that the system can use tools, retrieve information and work through a defined sequence of actions in order to complete a task.
For healthcare, that might mean receiving a clinical request, identifying which patient information is relevant, retrieving the appropriate clinical events, preparing an output and returning it to the healthcare professional for review.
The intelligence does not disappear.
But the workflow becomes part of the intelligence.
That is the shift from chatbot to clinical system.
Why Respocare Connect AI was built around the record
Respocare did not enter healthcare through artificial intelligence.
We were already here.
For years, our work has involved medical professionals, patients, healthcare services and the operational realities that exist beyond the technology industry. That experience inevitably influenced how we approached the development of Respocare Connect AI.
We did not begin by asking where we could place an AI model.
We began by looking at the friction already present in healthcare.
A doctor finishes a consultation and still has documentation to complete.
A patient's history is distributed across months or years of records.
The same information has to be reconstructed for a referral, a clinical summary, a report or a medical motivation.
An important result sits somewhere in the record but has to be manually found.
A previous treatment response is documented, but not necessarily visible when the next decision is made.
Different pieces of the patient's history exist, but the clinician remains responsible for assembling them into a coherent picture.
These were information and workflow problems before they became AI opportunities.
That distinction matters.
Respocare Connect AI has therefore been developed as a connected clinical environment.
The medical AI scribe is part of it, but not the whole of it. A consultation can contribute to a living clinical record. Patient information can be represented as clinical events rather than simply as disconnected files. The agentic clinical assistant can retrieve context from that record. Clinical reports can be generated from existing patient information.
Decision-support workflows can reference relevant clinical history. The medical professional remains responsible for reviewing the work.
None of those capabilities is revolutionary in isolation.
The significance lies in connecting them.
The patient record becomes the common thread.
Clinical AI must also know when not to answer
There is a quality that receives far less attention in discussions about artificial intelligence than intelligence itself.
Restraint.
A system designed for healthcare should not feel compelled to have an answer to everything.
If the available record does not contain the information required to answer a question, that absence should be surfaced.
If the evidence is conflicting, the conflict should not quietly disappear inside a confident summary.
If a conclusion requires inference, that inference should not masquerade as a documented fact.
This is one of the areas where healthcare can demand something better from AI than ordinary consumer technology does.
In a clinical environment, “I cannot establish this from the available record” can be an excellent answer.
It tells the doctor something important.
It preserves uncertainty rather than concealing it.
And it respects the boundary between an intelligent assistant and a medical professional.
The best clinical AI may therefore not be the system that answers the most questions.
It may be the system that knows most reliably which questions it should not pretend to answer.
Skepticism will make clinical AI better
There is considerable excitement around artificial intelligence in healthcare, but there is also understandable skepticism.
We think healthcare needs both.
Clinicians should ask difficult questions of technology companies.
They should ask how a system was tested, what information it can access, where that information is processed, how patient data is protected, how errors are investigated and how the system behaves when it encounters uncertainty.
They should question extraordinary claims.
They should be suspicious of demonstrations that confuse a polished output with proven clinical reliability.
They should expect companies operating in healthcare to understand that building a powerful technology and building a trustworthy clinical service are not the same thing.
None of this slows responsible innovation.
It improves it.
Healthcare has earned its high standard through the consequences of getting things wrong.
Artificial intelligence does not deserve an exemption from that standard simply because it is new.
At Respocare, skepticism does not make us uncomfortable. It is useful. It forces better questions, better testing and better engineering.
We are not interested in asking clinicians to trust AI because it is impressive.
Trust has to be earned by the system around it.
South Africa should participate in defining what good clinical AI becomes
There is another reason this conversation matters.
The development of healthcare AI should not be understood as something happening exclusively somewhere else and eventually arriving in South Africa as a finished product.
South African healthcare has its own realities.
Our regulatory environment matters.
Our healthcare infrastructure matters.
Our public and private sectors present different challenges.
Our clinicians work within local systems that technology designed for another environment may not fully understand.
Patient privacy must operate within South African law. Clinical deployment must take into account local workflows and standards. The role of regulators, including the Information Regulator and SAHPRA where applicable, will become increasingly important as intelligent systems move closer to consequential clinical functions.
That should be viewed as part of the maturation of the industry.
The objective cannot be to deploy as much artificial intelligence into healthcare as quickly as possible.
The objective is to determine where intelligence is genuinely useful, how it can be deployed responsibly and what safeguards are required when software begins participating more deeply in clinical work.
South Africa should be part of shaping those answers.
The future of healthcare AI may be quieter than we imagine
The popular image of artificial intelligence is highly visible.
A screen.
A chat window.
A question.
An answer.
The future of clinical AI may ultimately be much less theatrical.
The most useful intelligence may sit quietly inside the clinical workflow.
It may retrieve the relevant history before a consultation.
It may organise a complex timeline.
It may recognise that the information required for a report already exists elsewhere in the patient's record.
It may prepare documentation while the clinician moves on to the next task.
It may surface a previous treatment failure that would otherwise require several minutes of searching.
It may show exactly where a patient-specific statement came from.
And sometimes it may simply tell the doctor that the record does not contain enough evidence to answer the question.
The technology becomes less visible precisely because it has become more useful.
That is the future we find compelling.
Not artificial intelligence replacing the practice of medicine, but intelligence embedded around it.
Not another system asking for a doctor's attention, but a system designed to return some of that attention to the patient.
Not a chatbot pretending to be a clinician, but clinical technology that understands its role beside one.
This is the philosophy behind Respocare Connect AI.
Artificial intelligence is important to what we are building. It would be impossible without it.
But AI is not the purpose.
The purpose is to create a better way for healthcare professionals to work with the information already surrounding their patients—to reduce unnecessary administrative burden, preserve clinical context and make increasingly complex records easier to navigate.
Because the future of healthcare will not be determined by how much artificial intelligence we manage to place inside it.
It will be determined by whether we use intelligence to make healthcare meaningfully better.
Respocare Connect AI Agentic Intelligence for the Art of Medicine.
Comments