Who owns the agent? An operating model for AI in operations

Every agentic workflow eventually hits a case it should not decide alone: an angry key account, a claim above the threshold, an invoice from an unknown supplier, a regulation that changed last week. In that moment the question is not what the model does. It is who decides, and who is allowed to change what the agent does next time. Most programmes discover then that they have not answered it.

An operating model for agentic workflows needs six roles: the workflow owner, who is measured on the output; the solution owner, who may change prompts, rules and integrations; the operator, who watches the numbers and handles incidents; the risk owner, who signs off the control set and its changes; the data owner, who decides what the agent may read; and the sponsor, who pays and decides on the next workflow. In a mid-sized company these are six people who already exist, not a new department.

Write down, per workflow, what the agent may do alone, what needs a person, and who that person is. Then write down who may change these rules and how the change is tested. This document is the operating model. Dashboards, committees and policies come after it, and only if the document does not do the job on its own.

Agents in production change constantly: a prompt is tuned, a threshold is moved, a model version is swapped by the provider. Each change can shift behaviour across thousands of cases. The operating model has to say who may make which change, what test set it runs against, and who releases it. Without that, the first silent regression will cost more trust than the launch gained.

Do not design an enterprise AI operating model on paper. Run one workflow with the six roles filled by name, for one quarter. Keep what worked, write down what the roles actually had to decide, and apply that to the second workflow. By the third, you have an operating model that matches your organisation rather than a consultant's diagram.