What would you trust an AI agent to recommend—but not execute?

Real company processes cross departments, systems, owners, approval boundaries, and different definitions of “done.”

We encountered this while building automation for our own company, which led us to a more useful question: where should an agent’s authority stop?

Think about one recurring process inside your company:

• What should an agent be allowed to observe?

• What could it safely recommend?

• Which action must still require human approval?

• What evidence would convince you to grant it more autonomy?

It could be changing campaign budgets, handling renewals, escalating support issues, resolving finance exceptions, or responding to production incidents.

What is the first company process you would place in shadow mode—letting the agent observe and recommend, but not execute?