AI agents in production: what actually matters

Building a demo of an AI agent takes an afternoon. Putting one in production in front of customers or real business data is another discipline.

The difference is controls: what the agent can and cannot do, what happens when the model's confidence is low, who reviews consequential actions, and how every decision is logged so it can be audited later.

Our practical rule: every production agent needs least-privilege permissions, a human approval point for irreversible actions and a manual fallback path. If the agent fails — and it will, sometimes — the business keeps running.

With that framework, agents stop being an experiment and become infrastructure: they handle enquiries, prepare documentation, move data between systems and scale without adding headcount.

Have a candidate process? Tell us about it and we will tell you frankly whether an agent is the right tool.