An AI assistant that only drafts text creates one class of risk. An agent that can read SharePoint, query customer systems, create tickets or change cloud resources creates another. It acts with delegated authority, so its governance must cover identity, data, actions and evidence.
Microsoft's 2026 Cloud Adoption Framework recommends a central, enforceable baseline for AI agents rather than separate rules for every experiment. South African organisations can use that principle without waiting for a perfect enterprise AI programme.
Create an inventory before a policy
Start by recording every agent's owner, purpose, users, data sources, actions, model provider and operating environment. Include low-code agents and vendor features—not only custom code.
An agent without a named business owner should not enter Production. The owner approves its purpose and risk; the technical team implements controls and retains evidence.
Give every agent a distinct identity
Do not share a broad service account across agents. Use managed identities or dedicated application identities, grant only the permissions required and separate development, test and Production access.
Microsoft recommends distinct access controls for each environment. This limits an accidental deployment and makes activity attributable. Apply conditional access and privileged identity management where the supporting service allows it.
Control the data path
Document what the agent reads, what it writes and where prompts, responses, embeddings and logs are stored. Data residency cannot be inferred from the location of the source database alone. Agent services and model providers can have their own processing boundaries.
For POPIA-sensitive workloads, involve the information officer and legal team in the risk assessment. Confirm the operator, purpose, retention, access, transfer and deletion arrangements. A South African Azure region may support part of the architecture, but it does not automatically make the full agent workflow compliant.
Constrain actions, not just answers
An agent that recommends a change and an agent that executes it require different controls. Begin with read-only access. Add human approval for high-impact actions such as changing permissions, publishing content, sending external messages or modifying Production resources.
Set technical limits: approved tools, maximum transaction value, permitted resource groups, rate limits and an emergency disable path. Test prompt injection and untrusted-document handling before enabling write access.
Make activity observable
Microsoft's guidance calls for central observability using tools such as Azure Monitor, Log Analytics and Application Insights. Log the agent identity, user, tool call, result, approval and policy decision without placing secrets or unnecessary personal information into logs.
Alert on unusual volume, denied actions, new data sources, unexpected geographic access and repeated attempts to exceed permissions. Retain enough evidence to investigate an incident and explain a business decision.
Govern cost and lifecycle
Tag resources by owner, environment and cost centre. Set budgets and alerts before broad use. Record the model and prompt versions used for material workflows. Review agents periodically and remove identities, data connections and resources when the use case ends.
OAS Microsoft Azure services can help organisations design the identity, network, monitoring and governance foundation around an agent workload. The technology is only one part: ownership, review and evidence complete the control system.
A minimum approval checklist
Before Production, confirm that the agent has a named owner, a documented purpose, least-privilege identity, approved data sources, separate environments, monitored actions, cost limits, human approval for high-impact steps and a tested shutdown process.
This baseline keeps experimentation possible while making authority visible. That is the central governance question: not whether an agent is intelligent, but what it is allowed to do and how the organisation will know.
Sources
- Microsoft Cloud Adoption Framework: Govern and secure AI agents, updated 26 June 2026 and accessed 20 August 2026.
- Microsoft Cloud Adoption Framework: Governance for AI on Azure infrastructure, updated 10 April 2026 and accessed 20 August 2026.
Plan an identity, monitoring and governance foundation for AI workloads with OAS Microsoft Azure services.