South African businesses must secure unstructured data before deploying autonomous AI agents

South African businesses are moving quickly to autonomous AI agents, yet the biggest danger may not be the models themselves but the mass of unstructured information they can reach. In practice, that means e-mail archives, shared folders, old SharePoint sites, chat logs and document stores that have been left in place for years and are often poorly governed.

According to Accelera Digital Group customer engineer Anju Mammen, that material is where sensitive records, personal inf...

ormation and obsolete permissions tend to accumulate. Once an AI agent is introduced, those weaknesses can be amplified at machine speed, turning a long-neglected file path or folder into a potential POPIA breach.

That concern is sharpened by the way agents now operate. Google Cloud’s AI Agent Trends 2026 report says these systems are increasingly handling multi-step workflows across different platforms, while also widening the enterprise attack surface through their links to models, data and tools. Mammen argues that, under POPIA, automated collection, analysis or duplication of personal information still counts as processing, meaning organisations remain responsible if an agent accesses data without a clear lawful basis.

The other major problem is access control. Many firms still rely on inherited directory structures in which broad permissions were set up long ago and never fully reviewed. If an agent is given similar workspace-level access, it can move through those repositories without friction, exposing material that was never intended for automated use. Google Cloud’s report underlines the security implications of that kind of expansion, particularly where privacy exposure grows alongside the number of systems an agent can reach.

The answer, Mammen says, is to stabilise the data environment before deploying automation. That begins with mapping the unstructured estate, classifying personal information and identifying high-risk pockets. It should be followed by a full review of inherited permissions, removal of redundant or obsolete material, and the creation of tightly defined agent sandboxes that only expose the data needed for a specific task.

Continuous monitoring is equally important. TechRadar has separately warned that many organisations still lack clear visibility over what agents are active, what they can access and who is accountable for their actions. It has argued for granular guardrails, real-time oversight and auditability, while also noting that low-code workflow layers can help create clearer boundaries between data and AI. Another TechRadar analysis said agents often fail in enterprise settings because they lack context, reinforcing the case for structured workflows and better governance rather than rapid, unchecked deployment.

For South African CISOs and CIOs, the message is straightforward: clean up the data estate, fix the permissions, limit what agents can see and only then expand their use. In a POPIA environment, safe AI deployment is not simply a matter of choosing the right tool; it depends on whether the organisation has built the discipline to control what that tool can touch.

Source: Noah Wire Services