AI Agents

Agents are being sold as autonomous staff. In practice, the ones that survive contact with a real business are narrowly scoped, heavily instrumented, and stop to ask a human before anything irreversible. The design question is not how much autonomy you can give, it is where you deliberately withhold it.

• Scoped agents against a defined operational task
• Tool and system access with least privilege
• Approval gates on anything that commits money, goes to a client, or cannot be undone
• Full logging of what the agent did and why
• Fallback and escalation paths
• Measurement against the manual baseline

Six to twelve weeks, run in stages, with the agent shadowing the manual process before it takes any action.

• The agent running against real work
• The approval and audit trail
• The measured comparison against how it was done before
• Defined conditions for widening or pulling back its scope

If the task varies every time, has no clear success test, or carries consequences your business cannot absorb when it goes wrong, an agent is the wrong shape. Automate the parts around it instead.