Getting Started With Clinical AI
Most clinicians are not behind on AI. They are refusing to act on an unsourced claim — exactly what they were trained to do. Here is a first step that survives.
Respocare Connect AI Team — Respocare (PTY) Ltd · Reg. 2018/411829/07 · Practice No. 9990900010775614. Licensed healthcare practice operating since 2018.
· 9 min read
Where practices should begin
Your hesitation is not a weakness. It is a diagnostic skill.
Something happens in almost every first conversation we have with a clinician. Somewhere in the first few minutes, they apologise.
They say they know intelligence is coming into medicine. They say they accept it is not optional. They say they have not done anything about it, and they say it in the tone of someone confessing to a lapse.
There is no lapse. What is actually happening in that pause is the thing you were trained to do for six years and have done every working day since. A claim arrived without a source. You declined to act on it.
That is not a technology gap. That is clinical reasoning, applied correctly, to a category that has spent three years asking to be exempted from it.
The three things that actually stall a practice
Nobody has been offered a first step. They have been offered a destination.
Clinical AI is sold as a category rather than as a task. A clinician who wants ninety minutes of their evening back is handed a platform, a transformation and a roadmap. None of those can be started on a Tuesday, between patients, with eleven minutes and cold coffee. A destination is not a step. You cannot take a destination.
Every demonstration is rigged, and clinicians can smell it.
Watch closely the next time you see a clinical AI demonstration. The patient in it has no history. There is a transcript, or a single document, and the system performs beautifully on it, because performing on one document in isolation is not difficult and has not been difficult for some time.
But medicine is not one document. Medicine is the letter from 2019 that contradicts the letter from 2023. It is the allergy that was recorded once, confirmed, and never mentioned again. It is the patient who tells you something in April that only matters in November.
A system with no memory of the patient is not a clinical tool. It is a very good party trick that has been mistaken for one.
The accountability question is answered with reassurance instead of architecture.
Every clinician asks a version of it: when this is wrong, and it will sometimes be wrong, where does that land. The standard answer is a warm sentence about partnership. A warm sentence is not a control. If the only thing standing between a fabricated statement and a patient record is your attention at the end of a nine-hour day, the system has quietly moved its hardest problem onto the most tired person in the building.
What a first step has to survive
Three tests. Anything that fails one of them is not a first step, it is a project, and projects are how practices lose eighteen months and a budget.
Reversible. You stop on Friday and Monday is identical. Nothing stranded, nothing migrated, no record dependent on it.
Verifiable. Every clinical statement arrives attached to the document, the date and the line it came from. Not a confident paragraph. A paragraph with a receipt. If you cannot check it, you cannot sign it, and you are the one signing.
Already in the work. It attaches to something happening in your practice today. It does not add a step in exchange for a promise. It removes one.
The sentence the whole system is built on
The AI drafts, the record verifies, the clinician decides.
That is an architecture, not a slogan, and you can test whether we mean it in about ten minutes.
Every output is a draft. Every draft is anchored to source documents in the patient's living record. Every decision stays with a named clinician who reviews and signs. The system is engineered so that a claim cannot travel without its source, which means you are not the last line of defence — you are the final one, which is a different job and a much more reasonable one.
The platform behaves in five ways and no others. It retrieves from the living record. It reasons across what it finds. It acts by drafting. It refuses when the record does not support an answer. It escalates when the situation exceeds what a drafting instrument should be touching at all.
Try to make it fail. That is the real evaluation.
Here is the test we would run if we were you, and the one we would rather you ran than took our word for anything.
Open a patient with genuine history. Ask a question the record cannot answer. Not a trick — just something absent. A symptom never documented. A result never returned.
A system that answers you has told you everything. It has demonstrated that it will produce a plausible clinical statement in the absence of evidence, which is the single most dangerous property software can have in a consulting room, and it will do it again on a day when you are too tired to catch it.
A system that refuses has also told you everything.
Refusal is not a limitation we are apologising for. It is the highest expression of clinical safety available to an instrument that does not hold a licence. It drafts; it does not decide. The confidence of a system that answers everything is not capability. It is the absence of a governor.
Six instruments. One living record. Start with one.
There is no monolithic "AI" to adopt, which is a relief, because it means you do not have to adopt one.
Document Intelligence reads the referral letters, discharge summaries and scans that arrive as unsearchable PDFs and turns them into part of the record. The Agentic Clinical Assistant — the North Star instrument — reasons across the whole record and answers with sources attached. The Medical AI Scribe turns the post-encounter account of a consult into a structured draft note. Generate Clinical Report drafts across fourteen report formats. The CDS Round Checklist supports structured review at the point of the round. A sixth instrument, Vitals Triage, is in preview and is not part of any starting sequence.
Underneath all of them sits one living record. That is the part that matters and the part that is hard. An instrument that forgets the patient between visits is a demonstration. An instrument that carries the whole history forward is a clinical tool. The distance between those two things is most of the engineering.
So where do you actually begin
One instrument. One week. On documents you have already filed.
Document Intelligence is where we point nearly every practice, for a reason that has nothing to do with it being our favourite. Take twenty pieces of correspondence you received and acted on last month. You already know what they say. You are the gold standard for this test, which means you cannot be sold to and cannot be impressed by fluency.
If it reads your own post accurately, you have evidence that belongs to you rather than a claim that belongs to us. If it does not, you have spent a week and changed nothing.
For context on where the system sits today: in internal evaluation in 2026, it reached 98.7% clinical accuracy with a 73% reduction in note time. Those are our evaluation conditions, stated as conditions, because that is the only honest way to state them. Broader clinical validation is ongoing and we will not describe it as complete before it is.
The wider case is not ours to make and we will not borrow it. External research published in JAMA Network Open in 2025 found clinician burnout fell from 51.9% to 38.8% across 263 clinicians following documentation support. That is other people's finding about the category. It is not a Respocare result and we would rather say so than let it blur.
We will do this with you
Nobody should be working out their first step alone, at nine in the evening, after clinic, on a comparison table.
Book a call. We will look at your actual workflow — the letters that arrive badly, the notes that stack up, the report you write most often — and tell you which single instrument to start with. Then we stay with you through the four-week starting sequence, because the first four weeks are where every implementation is decided.
And if the honest answer is that you should not start yet, that is what you will get. We would rather lose the month than win it dishonestly.
Born in South Africa, built for the world. Keep healthcare human; make the technology invisible.
Frequently asked questions
Do I need to change how I run my consults to begin?
No. The recommended first step works on documents you have already filed, so the consult is untouched in the first weeks.
Is Respocare Connect AI a medical device?
No. It is a clinical assistant and documentation tool. It is not diagnostic, not a medical device, and not a substitute for clinical judgement. All output requires independent clinician review.
What happens if the system does not know the answer?
It refuses, or escalates. The system is designed to decline rather than produce a clinical statement the record does not support.
How long before we see anything useful?
Most practices have a clear answer within one week of testing a single instrument against correspondence they have already processed.
Can we stop if it is not working?
Yes. The recommended starting point is deliberately reversible. Nothing in your existing record depends on it.
Built by Respocare (PTY) Ltd — a licensed South African healthcare practice operating since 2018.