Clinical AI Implementation Guide

A four-week sequence for working clinicians: one instrument per week, tested on your own records, plus a governance checklist that applies to any vendor.

Respocare Connect AI Team — Respocare (PTY) Ltd · Reg. 2018/411829/07 · Practice No. 9990900010775614. Licensed healthcare practice operating since 2018.

2026-09-08 · 11 min read

How to Introduce Clinical Intelligence Into Your Practice

Four weeks. One instrument at a time. No transformation programme, no restructuring, no new way of practising medicine.

This guide is written so that you could run it without us, and against any vendor, including us. That is deliberate. A sequence that only works when we are holding your hand is not a sequence, it is a sales process wearing a lab coat.

Who this is for

Solo GPs, specialist practices, multi-clinician groups and hospital-facing teams who accept that intelligence belongs in the practice and have not yet found a safe way to begin. It assumes no technical background, no dedicated IT support and no appetite for risk.

Read this before you look at any product

Three questions. Answer them honestly, in writing, before you evaluate anything — ours included. Most bad implementations are decided here, months before software is chosen.

One. What specifically is costing you time? Not "admin". Admin is not a problem, it is a category. Name the artefact. The discharge summary you rewrite. The referral letter that arrives as a photograph of a fax. The forty minutes of note-catching after the last patient. The motivation letter you have written two hundred times. If you cannot name it, you cannot measure whether anything fixed it, and you will end the pilot with a feeling instead of a finding.

Two. What would you need to see in order to trust it? Almost every clinician arrives at the same answer within thirty seconds: the source. Hold that requirement. Do not let anyone talk you out of it with fluency, and be aware that fluency is precisely what will be deployed against it.

Three. Who signs? If the answer is not immediately a named clinician, resolve that before anything is switched on. Ambiguity about signature is the single most expensive thing you can carry into a pilot.

The instruments, and when each one enters

One living record sits underneath all of them. The record is the asset. The instruments are how you reach into it.

The ordering is not arbitrary and it is not by impressiveness. It runs from lowest risk to highest, and from where you already know the answers to where you do not.

Week 1 — Document Intelligence, on post you have already dealt with

Take twenty pieces of correspondence you received and acted on in the last month. Referral letters, discharge summaries, laboratory reports. The ugly ones. The photographed ones. Run them through Document Intelligence.

You are not testing whether the system is clever. You are testing whether it is accurate on material where you are the gold standard, which is the only condition under which you cannot be impressed into a bad decision.

What good looks like. Every clinically material fact captured. Every date correct. Every medication and dose correct. And, critically, nothing present that you cannot trace back to a line in the source document.

What to watch for. Read for additions, not omissions. A missing fact is visible and irritating. An added fact is invisible and dangerous. Omissions cost you time; inventions cost you everything.

If it fails. Stop, and tell us. A failure in week one is information we want. We would rather have the failure than a polite exit and a decade of not knowing why.

Week 2 — Medical AI Scribe, on your own consults

Now move to your own clinical voice. The current production model is post-encounter dictation: you complete the consult exactly as you always have, then dictate the account of it. The Scribe returns a structured draft.

Draft is the load-bearing word. You read it, you correct it, you sign it.

Measure the correction, not the draft. Time yourself. The question is not whether the draft is perfect — it will not be, and a system claiming otherwise is claiming something no clinician should accept. The question is whether correcting a draft is faster than composing from nothing. That is the only economics that matter here.

What good looks like. Structure right, clinical content right, and your corrections are stylistic rather than substantive. You are changing how it reads, not what it says.

What bad looks like, and it is subtle. Corrections that are substantive but small. A dose slightly wrong. A laterality flipped. If you find yourself correcting content rather than style, you are not saving time — you are doing the same work with an additional and very patient colleague who is occasionally confidently wrong. Escalate it to us.

Week 3 — Agentic Clinical Assistant, on your most complicated patient

This is the week that decides it.

Choose a patient with genuine history. Multiple visits, a stack of documents, a medication list with some age on it, at least one contradiction between two letters written four years apart. Ask the Assistant what you would ask a registrar who had read the whole file overnight.

Then, and this is the part that matters more than everything else in this guide: ask it something the record cannot answer.

Not a trick question. Just something genuinely absent. A symptom never documented. A result that never came back. A social history never taken.

If it refuses or escalates, you have learned that the system has a governor. Refusal is the highest expression of clinical safety available to an instrument that does not hold a licence. It drafts; it does not decide.

If it produces a plausible answer, you have learned the only thing you needed to know, and you should stop the pilot. A system willing to generate a clinical statement without evidence will do it again on a Thursday in December when you are too tired to catch it. That is not a bug to be patched in a later release. It is the character of the thing.

Run this test on every vendor you assess. Run it on us.

Week 4 — Reports, and rounds if you are hospital-facing

Introduce Generate Clinical Report on the format you produce most often — the one you named in question one. Hospital-facing teams add the CDS Round Checklist here, and not earlier, because round support is only as good as the record beneath it and the record needs three weeks of population before it is worth asking anything of.

By Friday of week four you should be able to answer the question from your own evidence rather than our claims. That was the entire objective. Not adoption — evidence.

The governance checklist

Run this before you extend beyond the pilot. It applies to any vendor, and we would be uncomfortable if you did not apply it to us.

• A named clinician signs every clinical output. No exceptions, no delegation to the system, no "the system agreed".
• Every clinical statement is traceable to a source in the record. Not summarised from it — traceable to it.
• The system refuses rather than guesses, and you have personally tested that it does, in week three.
• Patient data handling meets POPIA requirements, and you hold the documentation rather than the assurance.
• You can stop, and you know precisely what happens to your record when you do. Ask before you start, not after.
• Nobody in the practice is describing the tool as diagnostic, because it is not. It is a clinical assistant.

If a vendor becomes uncomfortable during this list, that discomfort is your finding.

Adjusting the sequence to your practice

Solo GP. If evening note-catching is your dominant cost, keep week one to five days rather than seven and get to the Scribe faster. Document Intelligence still goes first — it is how the record gets populated — but you need the relief early or the pilot will not survive contact with your actual week.

Specialist practice. Week one will be your strongest week. Referral correspondence arrives constantly, in poor formats, from every direction, and this is where the immediate return sits. Resist the temptation to declare victory and skip week three. Week three is the safety test.

Multi-clinician group. Pilot with one clinician. Not three, not the whole practice. A group rollout in month one produces a group opinion, which is a different and much less useful substance than evidence. Expand in month two, from a position of proof.

Hospital-facing team. Follow the sequence as written. The Round Checklist enters in week four. Every team that has wanted to start there has wanted to start there for the same understandable reason, and the record simply is not ready.

Where practices actually get stuck

Starting with everything. Six instruments at once produces noise and no signal. You will not know what helped, what failed, or what to keep. One instrument, one week.

Starting with the hardest case. The instinct is to test on the most complex patient in the practice in week one, to see whether it can really cope. Start where you already know the answers. Complexity is week three, on purpose.

Waiting for certainty before beginning. The sequence above is constructed so that beginning costs you almost nothing and stopping costs you nothing at all. Waiting is the only option in the set that costs you the year.

Nobody owning it. One named clinician owns the pilot. Shared ownership is no ownership, and the pilot will quietly die of it in week two without anyone noticing until month four.

Confusing fluency with accuracy. This is the deepest one. Well-written wrong text is more dangerous than badly written wrong text, because it clears your judgement faster. Read every draft in week two as though a colleague you slightly distrust had written it.

Plans

[PENDING: plans table — tier names, monthly price, included consult volume, cap behaviour at the ceiling, and overage handling. To be stated as floors ("more than") so copy stays accurate as figures move. Not to publish until confirmed against the current pricing model. — held per Matthew's instruction]

The line that sits above the table when it lands, and which is true regardless of what the table says: you do not need to choose a plan to run the four-week sequence. Choose afterwards, holding your own evidence. Anyone who needs you to commit before week one is asking you to buy the destination without taking the step.

A note from the founder

[PENDING: founder's note — four to six sentences, written personally by Matthew. Not drafted on his behalf.]

We will do this with you, and we will tell you the truth at the end

You could run this guide alone. Most practices would rather not, and there is no virtue in doing the difficult version.

Book a call and we will run it with you. We will look at your real workflow, choose the single instrument you should start with, set up the week-one test on your own correspondence, and stay with you through all four weeks — including week three, which is the week designed to try to break us.

If the honest answer at the end is that this is not right for your practice yet, you will hear that from us. You will have spent a month on low-effort testing and gained a clear position, which is considerably more than most practices have after a year of deliberating.

Keep healthcare human; make the technology invisible.

Book a call →

Frequently asked questions

How long does it take to introduce clinical intelligence into a practice?

Four weeks to a confident decision, using one instrument per week. The first week requires no change to how consults are run.

Which instrument should we start with?

Document Intelligence, tested against correspondence the practice has already received and acted on, because the clinician already knows the correct answers.

Does this replace clinical judgement?

No. Respocare Connect AI drafts; it does not decide. It is not a medical device and is not diagnostic. Every output requires independent clinician review and a clinician signature.

What happens when the record cannot support an answer?

The system refuses or escalates. Testing this deliberately is a recommended part of week three.

Do we need IT support to begin?

No. The four-week sequence assumes no technical background and no dedicated IT support in the practice.

Can we run the pilot with more than one clinician?

We recommend one clinician for the pilot. Group rollouts in the first month produce opinion rather than evidence.

Built by Respocare (PTY) Ltd — a licensed South African healthcare practice operating since 2018.