Martin Wolf is right that humans must answer for what AI decides; no organization can assign that answer until it has designed who held the authority to decide.
In the Financial Times on 17 June 2026, Martin Wolf argued that the world must prepare for an AI revolution it does not yet control. He warned that AI strains accountability, the rule of law, democracy, and the idea of what a person is. He cited Anthropic's own statement that the company now hands a growing share of its AI development to AI systems, and that those systems may one day design and build their successors without human hands on the work. He quoted the investor Vinod Khosla, who expects AI to perform 80 percent of economically valuable work and who cares less about whether mass displacement arrives than about whether governments hold a coherent policy framework ready for the moment it does. Wolf closed on a rule for uncertain times: prepare rather than relax.
Most readers will file the piece under risk. One sentence in it deserves a longer pause, because it carries the problem that lands on the desk of every executive deploying AI agents.
What Martin Wolf Is Warning About
Wolf's deeper warning sits in his account of who answers for a decision a machine makes. He names three parties. The programmer who built the AI. The executives of the company that sold it. The decision-makers of the institution that deployed it. He rejects the position he attributes to Argentina's president, Javier Milei, that AI can run institutions while no person answers for the result. Wolf wants the owners, managers, and staff of a deploying organization exposed to civil and criminal liability when their AI causes harm. He distrusts the self-restraint of AI's makers and calls for a licensing regime modeled on the clinical trials that gate pharmaceuticals, applied across borders so that competition cannot route around it.
Wolf's warning is institutional before it is technical. He is telling organizations that they will own the consequences of decisions their machines make, and that no appeal to the machine's autonomy will excuse them. A board cannot point at the model. A board answers for the model.
Why "Humans Remain Accountable" Sounds Correct but Stays Incomplete
Wolf's sentence reassures because it restates a principle boards already accept: a person answers for the institution's actions. Read the sentence twice and you see what it leaves out. Wolf names three accountable parties across three separate organizations. He does not say which person inside the deploying organization held the decision when an AI agent rewrote a contract, crossed a credit threshold, or settled a customer's claim.
Two questions hide inside the single word "accountable," and they run in opposite directions. Accountability looks backward: after an outcome, who must explain it and carry its cost? Authority looks forward: before the action, who held the right to decide, and how far did that right extend? Accountability needs authority to attach to. You cannot hold a division head accountable for a decision the organization never authorized her to make, or never recognized she was making. Pin liability on a person who held no real authority and you have not assigned accountability. You have selected a name to absorb a cost.
Wolf supplies the accountability principle. He leaves the authority map blank. The organization deploying AI agents draws that map, or operates without one.
Why Regulation Alone Cannot Answer the Question
Regulators have started to press on this, and their reach has a limit worth naming. Japan offers a current case. On 31 March 2026, the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry published Version 1.2 of the AI Business Guidelines. The revision names AI agents and Physical AI as distinct subjects for the first time, and it tells providers to build mechanisms that keep human judgment inside the operation of autonomous systems. It asks for appropriate allocation of authority, meaningful human intervention, and routine review of operational logs, with malfunction, privacy violation, and unintended consequences as the named risks.
Regulators have widened the question past technical performance. They no longer ask only whether a model is accurate. They ask where a person stands in relation to what the model does. A government can require that a human-judgment mechanism exist. A government cannot decide, for your bank or your insurer or your logistics operator, which person should hold which decision, where authority should sit along a specific approval path, or how authority should move from an agent to an analyst to a division head. The guideline sets the requirement. The organization writes the map. Regulators set the floor and leave the architecture to the firm that stands on it.
Japan's guidelines carry no penalty. They are soft law, and a firm can ignore the text without breaking it. The signal still matters, because the liability reasoning Wolf calls for will arrive through other doors. A court weighing a harm will ask who held the authority to prevent it. A counterparty negotiating an indemnity will ask who inside the firm was empowered to commit. An insurer pricing AI exposure will ask the same. Each of them asks the question the guideline now asks, and a firm that cannot answer it discovers the gap at the worst moment, in front of the party least inclined to forgive it.
The Hidden Limitation of Human-in-the-Loop
"Keep a human in the loop" has become the reflex answer, and it hides a failure mode that an audit will not catch until harm has already occurred. A human in the loop holds a seat on a workflow. Authority is the power to use that seat. The two come apart more often than most operating models admit.
Consider the analyst who receives an AI-drafted contract amendment with a deadline measured in minutes, no practical way to reconstruct the agent's reasoning, and a workload built on the assumption that approval is the default. She occupies the loop. She holds no authority. Her sign-off records a name without recording a judgment. The organization has installed a procedural human, present for the audit trail and absent from the decision.
Hollow oversight follows. A review the reviewer cannot refuse turns into a formality, and the agent's output passes through unexamined while the firm mistakes the record of a review for the fact of one. The damage compounds at the next step, because accountability now lands on the procedural human. The analyst's name sits on the approval, so the analyst absorbs the liability Wolf wants assigned, while the people who set the deadline, sized the workload, and chose to deploy the agent stay out of frame.
A human in the loop restores accountability only when that human holds real authority: the time to examine, the information to judge, and the standing to refuse and have the refusal respected. An intervention the person lacks the power to make is not oversight. The guidelines ask for meaningful human intervention, and the meaning turns on whether the person can act, not on whether the person appears in the diagram. The fix is not a better-trained reviewer or a longer checklist. The fix is an operating model that grants the reviewer the authority the diagram already implies she has.
The deeper limitation is architectural. Human-in-the-Loop treats governance as a moment of intervention. It does not define how authority is allocated before the intervention, how authority moves after the intervention, or how accountability remains continuous across the process. The result is oversight without architecture.
Decision Design
The condition under all of this is an absence of design. Organizations hold governance policies, transformation roadmaps, automation pipelines, and ethics principles. Few hold a deliberate map of where decision authority sits and how it moves once an AI agent enters the work. That map is the object of Decision Design.
What Decision Design Designs
Decision Design designs the allocation of judgment authority within an organization. It specifies who holds a given decision, how far that authority extends, where authority transfers to another person, under what conditions a decision escalates, when a person may override an agent's action, and when a process must stop. Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate.
What Decision Design Is Not
Decision Design is not governance. Governance sets rules and checks compliance against them. It operates above the decision and does not assign the specific authority for a specific judgment.
Decision Design is not digital transformation. Transformation changes how work flows. It leaves untouched the question of who answers when an automated step makes a consequential call.
Decision Design is not automation. Automation removes human steps for speed and cost, and the removed steps were the points where authority used to be visible.
Decision Design is not AI ethics. Ethics defines what an AI should do in principle, and principle does not tell a named manager whether this contract is hers to approve or counsel's to review.
What Problem Decision Design Addresses
Decision Design addresses four conditions that surface once AI agents join the work. Authority ambiguity, where no one in the organization can say which person holds a decision. Accountability fragmentation, where several people each touch the process and none owns the outcome. Invisible delegation, where the organization hands authority to an agent and keeps no record that a transfer took place. And the governance gap left when policy, transformation, automation, and ethics each assume some other discipline mapped the authority that none of them mapped.
Together, these conditions create what may be called a Governance Gap. The operational capacity of organizations expands through AI, while the institutional structures that allocate authority remain largely unchanged. The result is not a lack of governance. It is a misalignment between governance structures and operational reality.
Within this framework, Decision Boundaries define how authority moves. Decision Logs preserve how authority moved. One governs the transition. The other preserves the institutional memory of that transition.
Decision Boundary
The unit of that map is the Decision Boundary. A Decision Boundary marks the point at which judgment authority transfers from one holder to another: from an AI agent to a person, from one person to a more senior one, from an operating team to a control function. Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority.
A boundary is not a step in an approval workflow and not a number on a risk dashboard. A workflow step routes a task. A risk threshold flags a value. A Decision Boundary answers a question neither of them asks: at this point, who now holds the right to decide, and what may they do with it? Four transitions give a boundary its content.
Delegation transfers authority outward or downward, from a person to an agent or to a subordinate, and a real delegation states its limits rather than assuming them. Escalation transfers authority upward when a case exceeds the holder's mandate, and a real escalation names the trigger and the receiver in advance rather than leaving the move to discretion under pressure. Override is the authority a person retains to reverse an agent's action, and it exists only where that person holds the standing and the means to exercise it. Suspension is the authority to halt a process, assigned before anything goes wrong rather than improvised after harm appears, and given to a holder who can act in time.
Boundaries define where authority moves. A trace of those movements has to survive, or accountability dissolves the moment a decision passes through several hands. Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes. A log that captures what an agent produced tells you the result. A log built for Decision Design records who held authority at each boundary, what that holder was empowered to do, and where the authority went next, so that Wolf's accountable person can be identified after the fact rather than selected after the fact.
A Decision Boundary is therefore not a workflow artifact. It is a governance instrument. Its purpose is to make authority visible, auditable, and institutionally traceable.
How Authority Moves: Three Enterprise Cases
Picture a contract amendment handled by an AI agent inside a mid-sized enterprise. The agent drafts a revision to a supplier's payment terms. A Decision Boundary sits between the agent and a contract manager: the agent holds authority to draft, not to commit, and the boundary states the limit. The manager holds authority to approve standard revisions inside a defined commercial range. A pre-named condition triggers a second boundary between the manager and legal counsel, so that any change to liability, indemnity, or termination language transfers authority to counsel without waiting on the manager's discretion. The agent's accuracy is a separate matter for separate controls. The design question is where authority rests at each step and where it moves, and the firm answers it before the first contract runs, not after a dispute.
Take AI-assisted customer service. An agent resolves routine claims under authority delegated for a bounded set of cases. A boundary returns authority to a human representative when a claim exceeds a defined exposure. A further boundary lifts the decision to a manager when a legal risk surfaces, such as a disputed liability or a regulatory complaint. The quality of the agent's replies is one concern. The governing concern is the movement of authority: which conditions return the decision to a person, and which conditions raise it to someone with the standing to bind the firm.
The same logic governs AI-assisted risk decisions. An agent scores exposures and prepares recommendations, and a boundary reserves the risk-acceptance decision itself for an officer who can answer for it. The agent informs the judgment. The officer holds it. The boundary also fixes a suspension point: if the agent's inputs drift outside the range it was validated on, authority to continue reverts to the officer before the next decision runs, rather than after a loss reveals the drift. In each case the enterprise decides, in advance, who is authorized to decide, and that prior choice is what lets the firm name an accountable person later.
Back to Wolf
Martin Wolf is right. Humans must answer for what AI decides, and a firm should carry civil and criminal exposure when its machines cause harm. Accountability is the correct principle. It is also the second question. A firm can assign accountability only after it has answered the first one: who was authorized to decide?
If an AI agent acts and harm follows, the firm that mapped its Decision Boundaries can point to the person who held authority at the relevant point, show what that person was empowered to do, and trace where the authority moved. The firm that drew no map can point only to whoever's name landed on the approval, which is how the cost comes to rest on a procedural human while the people who shaped the system stand clear. Wolf's principle protects no one until the authority behind it has been designed.
Wolf is right that humans must remain accountable. But accountability cannot be assigned where authority was never designed. The governance challenge created by AI is therefore not simply one of oversight, compliance, or ethics. It is a problem of judgment architecture. Organizations that fail to design authority structures will discover accountability only after failure. Organizations that design authority structures can preserve accountability before failure occurs. That is the difference between governance as reaction and governance as architecture. That design discipline is Decision Design.
Decision Design is a judgment architecture framework proposed by Ryoji Morii, founder of Insynergy Inc., for structuring authority, accountability, and decision boundaries in AI-augmented organizations.