From Reactive
Event Correlation
to Autonomous
State Management.

R58B+

SADC telecoms CAPEX

Annual addressable spend

5,847

Duplicate tickets

From a single BGP flap

99.999%

Target SLA uptime

Bankhosa ACN baseline

<1ms

Detection to action

Layer 1–2 Verifier + Reasoner

0

Consulting bill

Refinery SLM replaces your ontology team

The AI Wall.
Why It Exists.
Why It Persists.

South African ICT infrastructure investment has accelerated dramatically over the past decade. Subsea cable capacity has expanded. Regional fibre density has increased. Enterprise WAN complexity has grown. And in response, a new category of tooling emerged — AIOps.

The promise was compelling: feed your network logs into an AI platform, and it would autonomously detect, diagnose, and resolve faults before they became outages. Vendors invested. IT teams integrated. And then — in production — the systems failed to deliver.

Not because the AI was wrong. Because the data it was trained on was rotten. Because the architecture it was layered onto was fundamentally unsuited to intelligence. The AI was being asked to reason from noise.

VISUALISATION · THE AI WALL PROBLEM

MARKET CONTEXT · SOUTH AFRICA

The Overlay Trap

Traditional AIOps tools are architectural overlays. They sit entirely outside the network fabric, consuming logs and events that have already been generated — which means they are, by definition, always operating on historical data.

By the time an overlay system has ingested a log, correlated it with past events, classified the anomaly, and generated an alert — the failure has already propagated. The SLA is already breached. The engineer is already firefighting.

The Garbage In, Garbage Out Silo

In virtually every large South African infrastructure provider, Network Operations, Security Operations, and IT Support functions operate in complete data isolation. Each team uses different logging formats, different severity classifications, different ticketing systems.

When an AI system attempts to correlate across these silos, it creates what Bankhosa calls "Alert Fatigue" — a cascade of thousands of similar-looking tickets from a single underlying fault, burying the root cause beneath thousands of duplicates.

Forced Automation vs. User Insight

Legacy AIOps deployments arrive with generic business rules developed for a global average customer — not for the specific topology, regulatory environment, and operational realities of South African infrastructure.

These systems ignore the institutional knowledge of the engineers who have managed the network for years. They override the "boots on the ground" insights that no generic ML model could ever replicate — and in doing so, they fail in exactly the edge cases that matter most.

Where Palantir needs 6–18 months of forward-deployed consultants to build your data ontology, and Databricks needs a dedicated engineering team to stand up its medallion pipeline — Bankhosa's architecture builds itself. No consulting bill. No professional services ramp.

A dual-layer intelligence framework — Inlay embedded in the control plane, Overlay closing the loop for humans — unified across four autonomous layers from raw data to governed action.

No Consultants. No Ontology Team. Just Clean Data.

Palantir Foundry requires a team of forward-deployed engineers — often 6 to 18 months of professional services — to build your data ontology before a single insight is produced. Databricks demands dedicated medallion pipeline engineers to refine raw data through bronze, silver, and gold layers.

Bankhosa replaces both with a specialised Small Language Model trained to ingest scattered, corrupted silo data — legacy tickets, raw telemetry, unstructured logs, routing tables — and output a unified, structured, continuously refreshed data layer. No consulting bill. No ramp time. The Refinery builds your ontology automatically and keeps it current.

Hallucination-Proof Intelligence at the Network Edge.

A second, lightweight SLM sits at the network edge — inside the Inlay — cross-checking what the LLM Reasoner believes is happening against live ground-truth telemetry before any action is taken. This is the hallucination guard that cloud-based platforms structurally cannot provide: they don't have eyes inside the control plane.

The LLM Reasoner then operates on verified, clean context — making routing decisions, traffic shaping calls, and anomaly classifications with causal reasoning, not correlation. This dual-SLM architecture is also what powers our security offering: the Verifier SLM reading packet metadata for fault prediction uses the same anomaly-detection primitives for lateral movement and DDoS pattern recognition.

Full Autonomy for Routine Faults. Governed Escalation for High-Stakes Decisions.

Every ACN decision executes autonomously — rerouting, traffic shaping, configuration optimisation — preserving the sub-millisecond response that SLA protection demands. The Governor logs every action immutably, with a human-readable audit trail that satisfies regulatory and stakeholder requirements without slowing the system down.

For a defined class of high-blast-radius decisions — large-scale reroutes, financial settlement routing changes, cross-segment topology shifts — the Governor applies configurable approval policies. These can themselves be automated: approve if confidence exceeds 95% and blast radius is below threshold. This is governance without bureaucracy. The opposite of Palantir's human-heavy operational review model.

CAPABILITY COMPARISON · RADAR

The radar chart above compares Bankhosa ACN against conventional AIOps across six critical dimensions. Every axis represents a capability that infrastructure leaders require — and where conventional tools structurally cannot compete.

One Strategic Partner.
Architectural Truth.

Bankhosa is identifying one infrastructure organisation for a 2026 Pilot Deployment. Not a trial. Not a proof of concept with synthetic data. A live deployment on a sub-section of a real production network.

We are looking for an organisation that is tired of AI hype, has experienced the limitations of conventional AIOps, and is ready for ground-truth architectural change.