Sovereign claims-document AI for medical schemes
Sentinel is a sovereign document-intelligence system for medical schemes. It reads the unstructured paperwork that floods a scheme every day (claims, pre-authorisations, member correspondence, PMB disputes) and turns it into triaged, drafted, audit-logged work, without a single byte of member health data leaving South African jurisdiction.
A sample pre-authorisation request, processed the way Sentinel would: read, PII tagged, drafted, and flagged. Nothing leaves South Africa. This is an illustration of the workflow, not live member data.
Thandiwe Mokoena, member 884 221 905, requests pre-authorisation for an MRI lumbar spine following chronic lower-back pain over 6 weeks. Referring provider: Dr S. Naidoo (PR 0294817). Plan: Comprehensive. Tariff code 3014.
Drafted response
Pre-authorisation conditionally approved for MRI lumbar spine (tariff 3014). Conservative-treatment history of 6 weeks meets the imaging motivation threshold. Flag: confirm no MRI in prior 12 months before final sign-off.
1 item routed for human review
Member health data is the most heavily protected category under POPIA: special personal information under Section 26, with the CMS watching on top. The global document-AI tools all run inference through US-controlled endpoints. For a scheme, that is a non-starter. Sentinel is built for on-prem GPU or Cassava AI Factory deployment so the model comes to the data, not the other way around.
Documents are classified and PII-tagged at the point of entry. Nothing is sent off-site.
A sovereign RAG layer grounds the model in scheme rules, PMB definitions, and policy, with retrieval logged.
The model drafts a response or triage decision. A human reviews and approves before anything is actioned.
Every inference, document touched, and data path is recorded for a CMS or POPIA evidence request.
Regulatory frame
Sentinel is being built with design partners. If this is your problem, help shape it.
