Trust and Guardrails Governing Autonomous AI Agents on Azure

This session looks at how authorization changes once software can interpret intent, choose tools, and act on a user's behalf. The old question, is this user allowed to call this API, becomes more contextual: should this specific agent take this specific action, for this user, under these conditions. That answer depends on identity and permissions, but also on business context, risk, reversibility, and potential impact. A core idea is that agents deserve their own identities and narrowly scoped permissions rather than inheriting whatever the human they serve is allowed to do. An agent that drafts proposals should not automatically delete records or move money just because its user can. We will look at designing human approval into the workflow without killing the value of automation, drawing the line capability by capability between what can run on its own and what needs a person to sign off. Policy, including spending limits, data handling rules, and approval thresholds, should live in a dedicated layer between the agent's proposed action and execution, not buried inside a prompt.