Most institutions cannot list where AI participates in a regulated decision.
Some of it was built in-house and registered as a model. Much of it arrived inside vendor platforms, procured as a claims system or a servicing tool, and was never classified as a model at all.
Discovery is the first engagement and it stands on its own — four to six weeks, scoped to one regulated decision path, typically lending or claims.
What it produces
A full inventory
Of where AI participates in that decision path, including decisioning embedded in vendor systems.
A decision profile classification
For each step in the path, mapped against the profile library.
The attached obligation
Mapped to the applicable South African instrument for each profile.
A comparative model assessment
Candidate models tested against each profile on your own workload rather than general benchmarks — which satisfy the requirement, which do not, and what the gap costs in capability, rand and residency exposure.
A specification of what would need to change
The output is an artefact a model risk committee, an internal audit function or a supervisor can read. It is useful whether or not anything further follows.
How this is built.
Every institution's estate is different, so the first engagement with each is partly bespoke. Discovery has to reach into systems that were never designed to be inspected, and connectors have to be built for whatever is actually there.
That work is priced per engagement. It is not, however, built per engagement. Each connector and each discovery pattern is built as a reusable component, so the second institution is faster than the first and the tenth is faster again.
The intent is a product, not a consultancy. Early clients pay for bespoke work and receive the benefit of everything built before them.
Taara is built. What it needs now is contact with real estates.
We are working with a small number of South African institutions to validate the decision profile library against their own workloads, customise it to how their governance actually runs, and contribute to the comparative benchmarking base: how an institution's workloads and failure patterns compare against a sector reference class.
That last part is the thing no institution can build alone.
Design partners validate the profile library against their own estate and contribute to the sector reference class. They shape the assurance approach while the local requirements are still forming.
If your institution is scaling AI into regulated decision paths and the assurance question does not yet have an owner, we should talk.
Talk to us