Why do AI agents revive an org-design term?
Because agents add a second kind of decision maker. While AI only recommended, decision rights stayed implicit in job titles; an agent that releases an order, replies to a customer or pays an invoice forces the question into the open. German business press now lists decision rights among the leadership calls agent adoption demands, down to examples like an agent releasing supplier invoices only below a set amount. An agent’s configuration is a decision-rights document, whether or not anyone wrote one: its tool permissions and thresholds encode who decided it may act.
How do you tier an agent’s decision rights?
By consequence and reversibility, in three bands. Advisory: summaries, drafts and recommendations that change nothing externally, free within the agent’s scope. Routine: reversible operational actions inside tight limits, run autonomously but logged. Consequential: anything touching money, customers, production systems or regulated records, which routes to a person, the human in the loop, through an explicit tool approval gate. Enforce the bands at runtime, in permissions and gates, never only in prompt wording, because an instruction is a request and a permission is a fact. Widening a band over time, as an agent proves itself, is the managed version of agent autonomy.
What breaks when decision rights stay implicit?
Accountability, first and quietest. When an agent misfires, the question “who approved this authority?” has no answer, and the cleanup lands on whoever notices. Shared ownership fails the same way it does with people: a control everyone owns, no one owns. The working pattern in agent governance is one accountable owner per agent, a written scope the owner can defend, and escalation paths decided before the first exception, not during it. When one agent hands work to another, the authority question compounds, which is why agent delegation chains need the rights to travel with the task. A company that cannot say who decided an agent’s limits has not governed the agent; it has only deployed it.