From copilots to agents: redesigning the work, not decorating it.

Most of what is sold as an ‘AI agent’ is a copilot with a different name. The genuine shift is from suggesting to acting — and that requires redesigning the work itself.

A copilot sits beside a workflow. It can draft, summarise, retrieve. A human remains in the loop, executing the steps. The productivity gain is real but bounded — by the speed of the human, and by the fact that the workflow itself has not changed.

An agent acts inside the workflow. It executes the steps, with the approvals, exceptions and shutdown controls that make the execution safe. The human moves from doing the work to designing, supervising and accepting the work.

The mistake we see most often is treating the agent as a copilot upgrade. Drop a tool-using model into an unchanged workflow and you get a liability in production — an agent with no documented approvals, no transaction limits, no audit trail. The seven production gates exist to prevent exactly that.

The redesign is the only part that pays back. If the workflow can be redesigned so that the agent does the routine, the exception paths are documented, and the human keeps the judgement and the relationships — that is where the value lives.

We start every engagement with the redesign. The Blueprint defines which workflows are worth reinventing and the operating model for human-agent teams; the Agent build turns one of those workflows into a production-grade agent; Managed Operations keeps it honest.