Decision rights are a system-design choice

What the paper develops

The evidence is clear. The authority is not.

A portfolio review can have the right data and still fail to act. An agent workflow can detect an exception and still keep moving. In both cases, the control loop stays open when no one has been assigned the right to advance, redirect, pause, or stop the work.

That is not mainly a meeting problem. It is a system-design choice. An org chart names positions; it rarely tells a cross-functional workflow who can close a particular consequential decision.

Design the decision, not just the committee

For each consequential recurring choice, write a short decision contract: who recommends, who must agree, who decides, and who performs. Ordinary operating choices should name one accountable decider; collective bodies need an explicit closure rule. Make veto rights narrow and require a veto to carry an alternative or an escalation path.

The model borrows a useful distinction from Bain's RAPID work without treating RAPID or RACI as doctrine. The point is to separate the case, the control right, the choice, and the follow-through. Clear rights do not repair weak evidence, poor incentives, or missing capacity. They make it possible to see who must act on those problems.

One decider is the ordinary operating default, not a claim that every decision belongs to one individual. A board, fiduciary committee, safety authority, works council, or separated approval function may legitimately share the right. In those cases, the contract names the quorum, decision rule, chair, challenge path, and missed-deadline treatment. Collective authority remains real; ambiguity does not.

Put the rights where work crosses a boundary

A role map in a slide deck is documentation. A role map in intake, a stage gate, a decision record, or an agent permission boundary is governance. The workflow should not let a consequential commitment advance without the required disposition.

For a portfolio, the recurring choices include admitting demand, releasing capacity, accepting an exception, redirecting funding, pausing work, and stopping it. Different decisions may have different deciders. The PMO does not need to own all of them; it needs to make the route from evidence to authorized action visible.

A portfolio capacity collision shows the point. The portfolio office can recommend a sequence using dependency and capacity evidence. A regulatory owner may hold a bounded veto over any option that misses a binding obligation. The executive accountable for the combined operating consequence closes the allocation, and delivery leaders update staffing, scope, and dates. The model does not make the trade-off painless. It prevents delay from masquerading as neutrality.

The same contract governs an agent fleet

A named human owner may use an agent to prepare a recommendation or perform an action inside a tested limit. A policy rule can enforce a pre-authorized stop, while the named control function holds the veto right. Consequential exceptions and recovery remain assigned to accountable people or governance bodies.

Delegated execution attaches to the action, threshold, and operating state—not to the agent as a whole. Aggregate thresholds also matter because several locally authorized actions can combine into an unauthorized outcome. Override privileges, escalation paths, stop mechanisms, and recovery ownership belong in the system.

The fleet therefore needs both local and aggregate controls. A refund agent may act within a tested amount, yet repeated credits across linked cases can cross a customer or financial boundary that no single action reached. A cumulative threshold should stop the chain, route the exception, and identify who can restart the fleet. The person who owns recovery remains named even when an agent performs the corrective steps.

Use formal governance where consequence earns it

Not every choice deserves a matrix. Reversible, low-cost decisions should stay light and local. Formal mapping earns its cost when a wrong choice is hard to reverse, difficult to contain, or likely to commit other work.

Measure whether the design closes choices instead of counting role charts. Track the time from a decision-complete recommendation to a disposition, work that continued because no decision arrived, vetoes that carried no alternative, and whether the performer changed the record on time. Those measures reveal whether evidence, authority, execution, or recovery broke the chain.

The test is simple: can the team name the decision, the recommender, any legitimate veto, one decider, and the performer? If not, accountability still belongs to no one.