Practical AI for healthcare workflows.
DHS uses AI for four things inside a clinical workflow: structuring intake, drafting summaries, surfacing triage signals, and proposing routing. Every output is reviewed by a clinician, and every call is logged with input, model, output, and reviewer.
Case · #4821
via WhatsApp · 10:42
Drafted by AI · clinician approves before the record updates
A structured, clinician-ready case from any inbound patient message.
The model extracts symptoms, duration, language, and flagged keywords from inbound patient messages. It produces a short summary in the clinician's working language, a proposed priority, and a suggested routing target. The clinician then opens the case and decides the next step.
Raw input
Structured output
AIEntity extraction · confidence scoring · triage signal · routing
From inbound message to clinician brief in about fifteen seconds.
Between the moment a patient message arrives and the moment a doctor opens the case, the platform runs five processing steps in roughly fifteen seconds. Every step is logged and reviewable, and the doctor opens a finished brief rather than a raw chat thread.
DHS AI · Clinical intake processorRaw intake
“Mild fever and cough for two days. Could I see a doctor today?”
Extracted
Triage signal
Confidence 91%
Generated summary
AI generated · pending clinician review
Every AI-assisted action is logged with input, output, and reviewer.
Language detection, entity extraction, triage scoring, summary drafting, and routing.
Every inbound message runs through the same five stages and writes its results back into the case record. Clinicians can review the reasoning behind each output, override the proposed routing, and the override is logged for the model to learn from over time.
01
Input captured
WhatsApp · USSD · clinic desk
02
NLPLanguage parsed
Symptoms, duration, channel extracted
03
AIRisk scored
Priority 2 / 5 · Confidence 91%
04
Case structured
Fields populated, summary drafted
05
Reviewed & acted
Clinician confirms, case closed
Four boundaries that protect the clinician, the patient, and the record.
DHS sets the AI boundaries on day one of every project. The four below apply regardless of how the model develops or how the workflow evolves, and they shape how every output is reviewed and approved.
-
Diagnosis stays with the clinician
AI organises the context a doctor needs for a consultation. The diagnosis is made by the clinician based on that context and direct examination of the patient.
-
Treatment stays with the care team
Prescriptions, drug choices, and dosing all sit with the care team. AI may surface relevant guidelines on request.
-
Outputs go through review
Every AI output passes through a named clinician before it changes a record, reaches a patient, or sends a message.
-
Clinical judgement is final
Where a triage signal or summary conflicts with what the clinician sees, the clinician's call stands and the override is logged.
How providers stay in control of AI behaviour in their workflows.
Three mechanisms give the provider full control. A complete audit trail records every AI call. A named clinician reviews each output. Each workflow has its own switch for AI to be on, off, or partial. The full architecture sits in our approach.
Audit trail on every call
Every AI call is logged with input, model version, output, reviewer, decision, and timestamp. Ministry inspections, donor reviews, and internal audit teams can read the same trail.
[10:42:33] AI_ASSIST · intake_v2 · Case #4821 · reviewer: Dr. Kamara · APPROVED
Clinician review in every workflow
Every AI draft is reviewed by a clinician before it touches a record. The reviewer role is configurable per workflow and per risk level, from senior clinician down to senior nurse.
Configurable per workflow
AI can be enabled for intake and triage in one clinic and turned off for pharmacy or dispensing in another. The platform supports any mix of workflows with AI on or off.
DHSAFRICAWant to put AI inside an existing healthcare workflow?
Tell us which workflow is slow today, where the staff repeat themselves, and what should stay manual. We come back with a concrete plan for what to automate, what to leave alone, and where the audit lives.
About a minute · No commitment