The questionHow do you keep customer-identifiable data out of the AI?
Agents operate on network and infrastructure data — topology, configuration, performance counters, asset registers, and are not connected to subscriber-facing systems unless the operator explicitly provisions it. Scoping is decided at integration time with the operator's data protection officer, and enforced at the gateway before any agent or query runs.
What we can evidence
Exclude subscriber identifiers as a configuration decision, not a code change.
The operator specifies which database views, tables and columns each agent pipeline can access, decided at integration time with the data protection officer. Excluding subscriber identifiers requires no platform customisation.
Reject out-of-scope requests before any database is touched.
Per-pipeline access enforcement is applied at the gateway layer. A request falling outside the granted scope is refused before any agent runs or any query is initiated — the enforcement point is upstream of the thing being protected.
The default is network data, not subscriber data
The datasets AiDN is built to reason over are topology, configuration states, performance counters and asset registers. These are the datasets where the operational value sits, and they are largely free of personal data by nature. Subscriber-facing systems are not connected unless the operator explicitly provisions that connection — the absence is the default state, not a setting someone has to remember to apply.
POPIA and GDPR alignment in practice
- Automated analysis is advisory by default — the platform's conclusions are recommendations until a human acts on them.
- Consequential actions sit behind a human-in-the-loop confirmation gate, so no automated decision has legal or operational effect without a person.
- Every conclusion is supported by an explainable evidence trail, which is what the auditability requirements on automated decision-making actually demand.
- Data scoping is documented at integration time, giving the DPO an artefact to review rather than a behaviour to trust.
The aim is an AI platform a data protection officer can evaluate and endorse without accepting unusual compliance carve-outs.