The Delegation Map: What Machines May Do, Recommend, or Never Touch

Every organisation already delegates. Boards delegate to management, managers to teams, teams to procedures and systems. Delegated authority is one of the oldest instruments of governance, and most organisations have written schedules that say who may approve what, up to which amount.

What changes in the Age of Agency is the arrival of a new kind of delegate. It is fast, tireless and increasingly capable. It does not tire, forget or leave. And it is not accountable in any way a court or a regulator recognises. When it makes a mistake, the responsibility returns to the people and the organisation that gave it authority.

This essay introduces the delegation map, the central tool of The Age of Agency: a deliberate answer to the question of what machines may do, what they may only recommend, and what they must never touch.

Three rings of authority

The map sorts every decision into one of three rings.

The outer ring: the machine executes. Decisions an agent may make and carry out within explicitly authorised limits. Matching invoices under a value cap, routing service requests, reconciling data, scheduling and preparing first drafts are typical examples. People design the limits and review samples of the results, but they do not approve each action.

The middle ring: the machine recommends. Decisions where the agent analyses and proposes, and a named person decides. Credit decisions above a threshold, pricing proposals, supplier changes and customer remediation offers often sit here. The agent does most of the work; a person carries the authority.

The inner ring: people decide. A narrow core that stays with people because of its consequence, its irreversibility or its need for legitimacy. Decisions about an individual's livelihood or liberty, clinical judgement, hiring and dismissal, and safety-critical operations typically belong here. Machines may inform these decisions, but their outputs are inputs, never verdicts.

The rings are not a ranking of how clever the machine is. A system may be perfectly capable of making an inner-ring decision. The question is whether it should.

Placing a decision on the map

Five tests help place a decision.

1. Consequence. What is the worst plausible outcome of a wrong decision, for the organisation and for the people affected?
2. Reversibility. Can a wrong decision be detected and undone, and at what cost? A payment can sometimes be recalled; a dismissal or a medical intervention often cannot.
3. Legitimacy. Would the people affected accept a machine making this decision? Some decisions require a person not because a person decides better, but because only a person can be held to account in a way others accept.
4. Evidence. Is there reliable evidence that the agent performs well on this decision, across the cases it will meet, including unusual ones?
5. Speed of harm. If the agent goes wrong, how quickly does damage accumulate? Errors that compound at machine speed need tighter limits.

A decision that scores low on consequence, high on reversibility and has strong evidence can move to the outer ring. A decision with severe or irreversible consequences, or one that requires legitimacy, belongs in the inner ring regardless of how good the system is.

Why the answer differs by industry

The same decision can sit in different rings in different organisations. Approving a refund may be an outer-ring decision for a retailer and a middle-ring decision for a bank, where it can signal fraud or breach regulation. A scheduling decision is trivial in an office and safety-critical in a hospital or a power station.

The map must therefore be drawn by each organisation, for its own decisions, in the light of its own regulation, customers and risk appetite. Industry guidance can provide a starting point, but it cannot provide the answer.

The quieter cost of standing still

Most discussion of AI risk focuses on reckless automation: agents given too much authority, too soon. That risk is real. But there is another cost, less visible and often larger: the cost of keeping decisions in the inner ring that could safely move outward.

An organisation that sends every routine decision to a person, out of caution, does not avoid risk. It moves the risk elsewhere: into slower service, higher costs, staff overwhelmed by low-value work, and competitors who drew their maps more deliberately. Standing still is also a decision, and it also has consequences.

The purpose of the map is not to minimise automation or to maximise it. It is to make each choice deliberately, with the reasons written down.

A living map

The map is not drawn once. As systems earn trust through evidence, decisions move outward. When they fail, or when circumstances change, decisions move back in.

That movement needs rules. Moving a decision outward should require evidence: performance data across representative cases, a review of failures, and approval by a named owner. Moving a decision inward should be possible immediately, by anyone accountable, without having to argue the case first. An agent's authority should be easier to withdraw than to grant.

Organisations that manage the map this way gain something valuable: a record of what they delegated, why, on what evidence and with whose approval. That record is exactly what regulators, auditors and boards will increasingly ask to see.

Questions for the board

1. Do we have a list of the decisions we have delegated to AI, including AI inside the software we buy?
2. For each material decision, which ring is it in, and who decided?
3. What evidence would move a decision outward, and who must approve it?
4. Can anyone accountable pull a decision back to people immediately?
5. Which decisions are we keeping with people out of caution rather than reason, and what is that costing us?

What follows

Once decisions are delegated, work changes shape. Some roles narrow, others expand, and value begins to move between roles, firms and industries. That process, which I call agentrification, is the subject of Part III.

This essay introduces Part II of the forthcoming book The Age of Agency. The delegation map is also applied in advisory work through Kiros Intelligence Group as the AI Delegation Map.

Views are the author's own. The frameworks described are applied methodologies of Kiros Intelligence Group, not regulatory standards.