Governing AI agents: permissions before prompts
Enterprises are ready to let AI do real work. The question is no longer whether agents are capable, but whether they are controllable.
The first generation of enterprise AI answered questions. The next one takes action: sending follow-ups, updating records, approving routine requests. That shift changes the risk profile entirely.
Start with the permission model
An agent should never see or do more than the person it acts for. In practice that means agents inherit role-based permissions from the systems they work in, rather than holding broad service credentials.
Define approval thresholds
Not every action needs a human, but some always should. Discount approvals above a limit, payments, external communications to key accounts: these are written into the agent’s workflow as mandatory approval steps.
Log everything, in one place
Agent actions belong in the same audit trail as human ones. When a customer asks why their credit limit changed, the answer should be one search away, whoever made the change.
- Scope agents to the minimum permissions they need
- Require approval for high-impact actions
- Record every action with its reasoning and inputs
- Evaluate agents continuously against real cases
Governance is not what slows AI adoption down. It is what lets organisations adopt it with confidence.
Want to talk this through for your business?
Tell us where your business is today and where it needs to be. If we are the right partner, we will set up a working session with the people who would build it.