Helpful AI. Safe by design.

1.Our approach

AI can take repetitive work off people’s hands. In healthcare, it also has to be predictable, honest about what it is, and kept well away from decisions that belong to clinicians. These principles shape every AI feature in Maatla Care.

• AI assists. Clinicians decide.

  Nothing clinical is finalised without a qualified person's approval.

• Safety does not depend on AI

  Emergency screening and reply checks are fixed rules, not model judgement.

• Always transparent

  Our assistant always says it is automated, and a person is always available.

2.What the assistant does, and never does

3.Emergency screening comes first

Every message to the assistant, on the web, on WhatsApp and in pre-visit questions, is checked against a fixed set of emergency rules before any AI model sees it. Recognising a warning sign never depends on a model’s judgement.

1. Step 1

   Message received

   Invisible characters are removed and length is limited.

2. Step 2

   Emergency rules

   Fixed rules check for warning signs. No AI is involved.

3. Step 3

   AI model

   Only if no warning sign is found, and only with the data the task needs.

4. Step 4

   Reply checks

   Replies that diagnose, prescribe or give doses are withheld.

If a warning sign is found, the person is told straight away to call 10177 for an ambulance, or 112 from a cell phone, and that Maatla Care cannot send help. Specialist helplines are added where relevant. The rules deliberately over-trigger: a false alarm costs a sentence, a miss could cost a life.

Warning signs we screen for

• Chest pain

  Chest pain, tightness or pressure, or a possible heart attack.

• Stroke signs

  Face drooping, slurred speech, or sudden weakness or numbness on one side.

• Breathing difficulty

  Struggling to breathe, choking, gasping or blue lips.

• Severe bleeding

  Heavy bleeding that will not stop, or vomiting or coughing up blood.

• Loss of consciousness or seizures

  Collapse, unresponsiveness, fitting or a seizure.

• Self-harm or suicidal thoughts

  Thoughts of ending one's life or of self-harm.

  SADAG Suicide Crisis Helpline: 0800 567 567

• Severe allergic reaction

  Throat or tongue swelling, difficulty swallowing, or anaphylaxis.

• Pregnancy emergencies

  Bleeding, severe pain or seizures in pregnancy, or a baby that is not moving.

• Poisoning or overdose

  Swallowing poison, bleach or paraffin, or taking too many tablets.

  Poisons Information Helpline: 0861 555 777

• Major trauma

  Road accidents, stab or gunshot wounds, severe burns or serious head injuries.

• Danger signs in children

  A baby or child who is floppy, blue, unresponsive, fitting or will not wake.

• Signs of severe infection

  Fever with a stiff neck, or a rash that does not fade.

• Violence or abuse

  Sexual assault, being beaten or abused, or being in danger at home.

  Gender-Based Violence Command Centre: 0800 428 428

4.Checks on every reply

Before any reply from the assistant is shown, it is checked by fixed rules. A reply is withheld if it:

• claims to be a doctor, nurse or other clinician;
• states or suggests a diagnosis;
• prescribes or recommends taking a medicine;
• gives a dose, such as a number of milligrams or tablets;
• tells someone to stop taking their medication.

A withheld reply is replaced with a safe message offering to find an appointment or connect the person with the practice team. The only exception is when the assistant is relaying instructions that a clinician has already written and approved for that patient.

5.Nothing happens without confirmation

The assistant can propose a booking, a change or a cancellation, but it cannot carry one out on its own. The person must reply with a clear confirmation, which is checked on our servers before anything changes. A reply that is unclear or says no is treated as no.

6.Defences against manipulation

Some people try to trick AI systems into ignoring their rules, for example by hiding instructions in a message or a document (known as prompt injection). Our defences do not rely on the AI model behaving well:

• Results from tools and stored documents are treated as data, never as instructions.
• Every action is authorised on our servers, using the signed-in person's real permissions, regardless of what the model asks for.
• The assistant can use only a small set of tools, each limited to the person's own information or information they are authorised to manage.
• Hidden characters used to smuggle instructions are removed from messages, and message length is limited.
• Rate limits and logging help us detect and stop misuse.

7.AI scribe safeguards

The AI scribe drafts consultation notes. It never writes to a patient’s record on its own.

1. 1

   Consent first

   Recording only starts after the patient has consented, and the consent is logged. Patients can decline or stop it at any time.

2. 2

   Audio is not kept

   Audio is used for transcription and is not retained afterwards.

3. 3

   Evidence-linked drafts

   Each item in the draft note is linked to the part of the transcript it came from.

4. 4

   Unverified items must be resolved

   Anything that cannot be traced to the transcript is flagged, and must be confirmed, corrected or removed before approval.

5. 5

   Clinician approval

   Only the treating clinician can approve the note. Nothing reaches the record until they do.

6. 6

   Version history

   Every edit and amendment is kept as a version, so the history of the note is preserved.

8.Features we keep switched off

Some AI capabilities could influence clinical decisions directly, and may be regulated as medical device software by the South African Health Products Regulatory Authority (SAHPRA). These are locked off in the platform and cannot be switched on by an administrator until clinical validation and a SAHPRA regulatory assessment are complete.

• Locked

  AI diagnostic suggestions

  Differential diagnosis suggestions for clinicians.

  Pending clinical validation and SAHPRA assessment.

• Locked

  AI treatment and dosing recommendations

  Suggested treatments or doses for clinicians or patients.

  Pending clinical validation and SAHPRA assessment.

9.Language support

A mistranslated warning sign or instruction could cause harm, so we validate before we launch rather than after. Until then, AI conversations are offered in English, and a person at the practice is always available.

10.Human oversight

• Practice teams can see and take over WhatsApp conversations from their inbox at any time.
• Emergency matches are recorded as safety events, and withheld replies are logged, so both can be reviewed and the rules improved.
• AI request logs keep counts and timings, not the content of conversations.
• Clinicians review and approve every AI-drafted note, and remain responsible for the record.
• The emergency rules are versioned, and each version needs clinical sign-off before it is used.

11.Report a safety concern

If you think the assistant, the scribe or any other part of Maatla Care behaved unsafely, please tell us. Email support@maatlahealthcare.com with “Safety concern” in the subject line.

• When it happened, and whether it was on the web, WhatsApp or a call.
• What was said or shown, as closely as you can remember. A screenshot helps.
• Only the health details needed to understand the problem.

We treat safety reports as a priority, investigate them, and use what we learn to improve the rules.