Classified by what the decision does to a person.
Not by the technology that produced it or the business unit that owns it. Most regulated AI in a South African bank or insurer falls into a small number of profiles.
Adverse outcome requiring a reconstructable reason
Credit decline, claims repudiation, account closure.
NCA section 62, the Policyholder Protection Rules, Treating Customers Fairly outcomes.
Threshold and prioritisation
Fraud alert ranking, AML triage, complaint routing.
FICA section 29 and FIC Directive 5 of 2019, under which knowledge of suspicious activity is deemed from the moment an automated system generates an alert.
Suitability of advice
Product recommendation, next best action, guidance surfaced in a digital channel.
FAIS General Code of Conduct and its record of advice requirements.
Information with a duty of care
Servicing responses, policy explanations, disclosure generation.
Differential access and pricing
Underwriting, risk-based pricing, screening.
Each profile carries a different evidentiary burden, and therefore a different requirement of the model serving it. The same model can be entirely appropriate for one profile and inadequate for another, with nothing about the model having changed.
Three layers, built for a South African estate.
Open source foundation
The governance mechanics are open and self-hostable: model inventory, approval routing, lineage, audit trail. This layer is commoditised and we treat it as such. It is free, inspectable, and runs inside your environment. Your security review is of code you can read.
The regulatory content library
Decision profiles, obligation mapping, and model fit criteria for South African financial services. Maintained and versioned as the regulatory position develops. This is what a global platform cannot supply and an institution should not have to build.
Continuous monitoring
Your estate changes faster than the law does. Models are swapped, prompts edited, retrieval sources updated, workloads deployed. Taara flags when something behind a regulated decision changes, tested against the standing obligation, so a technical change does not silently alter a validated decision.
Model selection is a three-way trade.
Not a two-way trade between cost and capability. It is a three-way trade that includes where the inference happens.
Amazon Bedrock became available in the Cape Town region in November 2025. The frontier Claude 4.5 models are reachable from Cape Town through global cross-region inference, which routes the inference itself to supported commercial regions worldwide. Logs, knowledge bases and stored configurations remain in af-south-1. The processing does not.
AWS offers geographic inference profiles, which confine routing within a defined geography, for customers with data residency requirements. Those profiles exist for the US, the EU, Japan and Australia.
There is no African equivalent.
In South Africa, model capability and data location are the same decision. The models with the strongest reasoning, long-context and instruction-adherence characteristics are reachable from Cape Town only through global routing. The models that stay in country are the smaller ones.
For some workloads that is not a trade at all, because a smaller model is genuinely adequate. Extraction, classification and short summarisation have converged. For others, a document-heavy affordability assessment or a multi-constraint disclosure, the capability difference is material, and choosing capability means choosing offshore processing.
Taara makes that trade explicit per decision profile, so it is made deliberately rather than as a side effect of a model ID.
No global routing or evaluation platform accounts for this, because it is not a constraint their customers face.