In June 2026, a photograph from the G7 summit in Évian captured an unusual seating chart. Beside the leaders of Canada, France, Germany, Italy, Japan, the United Kingdom, and the United States sat the heads of the world's leading AI companies: Sam Altman of OpenAI, Dario Amodei of Anthropic, Demis Hassabis of Google DeepMind, Arthur Mensch of Mistral, and Alexandr Wang of Meta.
Tom Wheeler of the Brookings Institution described the scene as the "sovereigns of geography" sitting down at the same table as the "sovereigns of intelligence." A handful of firms now control the cognitive capability that governments, companies, and citizens all depend on.
This article is not about the proposal exchanged at that table. It is about the question no one put on it. When AI decisions were made in that room, who ultimately decides? Hold onto that question. It returns at the end.
Why AI Standards Are Not Enough
At Évian, the AI companies behaved not as petitioners but as parties to governance. Altman warned the leaders not to "hand off your responsibility to AI labs like ours." Amodei urged states to resist "the temptation to fragment" on AI regulation. Hassabis proposed a U.S.-led standards body working closely with the international democratic community.
Their aim is clear. Fragmented national rules destroy the economic value of a uniform market. So they proposed common AI Standards, with oversight responsibility shared between companies and governments.
Wheeler's response is not to reject the offer but to accept it with two amendments. First, open the standards process beyond the companies to include government and civil society. Second, make the result enforceable. Voluntary compliance produces a suggestion, not a standard. When the G7 created the Financial Action Task Force in 1989, and when it built Basel III after the 2008 crisis, the standards worked because non-compliance meant exclusion from the market. Against technical standards like MCP and A2A that govern how systems behave, Wheeler argues for behavioral standards that bind how companies act.
The argument is sound. Brookings identified the right governance problem. This article begins one layer beneath it.
Standards are a mechanism for applying an agreed content uniformly across many parties. Electrical codes, building codes, and the internet's TCP/IP all work the same way: decide once what must be upheld, then distribute it to everyone.
Standards answer "what must be upheld." They do not answer the question one step earlier: who holds the authority to decide what must be upheld. Wheeler's call to include government and civil society is an attempt to fill this gap. Who sits at the standards table is not a question about the content of a standard. It is a question about who holds the authority to set it.
Standards are how authority is implemented, not authority itself. What the AI companies offered at Évian was a blueprint for plumbing. Who holds the tap remains to be decided.
Governance Is Not the Same as Authority
The same limit applies to the broader word, Governance.
Governance is the framework by which an organization or society governs AI: policy, audit, risk management, accountability. All of it is necessary. But the word is built to arrange the machinery of governing, not to name who finally decides any single case within it.
Consider an AI agent authorized to issue customer refunds on its own. Governance can require an audit trail, set risk thresholds, and appoint an owner. But the line itself — above what amount a human must take over — is not in the vocabulary of Governance to begin with. Arranging the machinery and drawing the line are different acts.
The distinction is precise:
Governance establishes rules. Decision Design structures authority. Governance defines what should happen. Decision Design defines who legitimately decides. Governance prepares the vessel. Decision Design determines where authority resides within it.
Why Human-in-the-Loop Is Not Enough
This missing object of design has already surfaced in policy language. The Japanese government's AI Business Guidelines Version 1.2, jointly issued by the Ministry of Internal Affairs and Communications (MIC) and the Ministry of Economy, Trade and Industry (METI), require mechanisms that ensure meaningful human judgment over autonomous AI agents to mitigate risks such as malfunction and privacy violations.
This is usually read as Human-in-the-Loop: insert a person somewhere in the processing chain. But that reading mistakes the mechanism for the point. The requirement is not to place a human inside a loop. It is to define, as an institution, who retains final Judgment Authority.
When a rule says "keep a human in the loop," what is actually being asked is an allocation of authority: which decisions a human takes on, and which may be delegated to AI. "Where in the loop does the human sit" is only the technical restatement of "where does authority sit." Human-in-the-Loop is a symptom. Authority Structure is the object of design.
Governance is not enough. DX is not enough. Automation is not enough. AI Ethics is not enough. Each matters, and none of them takes the placement of authority itself as the thing being designed.
What Decision Design Actually Designs
Decision Design is a design discipline that takes the act of judgment itself as its object. This article develops it as the cross-cutting concept the debate at Évian was circling without naming.
Decision Design is defined by three questions, answered in the same order every time: what it designs, what it is not, and what problem it solves.
What Decision Design Designs
Decision Design designs the placement of authority. It is not a concept for designing the outcome of a decision. It is not a concept for designing the decision process. Its object is authority itself: which decision is finally made by whom, how far it is delegated to AI, and where a human takes over.
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 making AI more accurate. However capable the model, whether its output may be adopted as a final decision is a separate question. Accuracy improves the material for a decision. It does not decide who holds the authority.
Decision Design is not Automation. Automation asks whether a task can be handed to a machine. Decision Design asks whether a judgment may be handed to a machine, and if so, how far. The first is about speed and effort. The second is about authority and accountability.
Decision Design is not DX. DX moves paper and personal habit into digital systems. It renews the container of a decision. It does not touch where authority over that decision sits.
Decision Design is not AI Ethics. Ethics debates the values a decision should satisfy: fairness, transparency, accountability. It disciplines the quality of a decision. Who exercises that valued judgment, and within what bounds, lies outside it.
Decision Design is not Governance. Governance designs rules. Decision Design designs the Authority Structure. Rules define what may be done. An authority structure defines who may decide.
What Problem Decision Design Solves
Decision Design solves the authority vacuum created when AI became able to execute judgment.
Until now, humans held a monopoly on judgment, so no one needed to design who decides. Role and responsibility roughly coincided. When AI can perform part of a judgment, that coincidence breaks. AI can execute a decision but cannot bear responsibility for it. The agent that acts and the agent that answers for it come apart. The authority vacuum opens in that gap.
Standards and Governance implement and discipline authority. Neither is the first act of placing it. That first move is Decision Design.
What Decision Boundaries Define
The central instrument of Decision Design is the Decision Boundary.
A Decision Boundary is the line separating what AI may complete on its own from what a human must take on. Every judgment splits into a region AI can close and a region a human must own. Designing that boundary means deciding four things:
Who decides. What may be delegated to AI. What must remain under human authority. Where legitimate authority changes hands.
These four are written into the institution as documents, not left to individual discretion or the mood of the moment. The boundary also fixes its own exceptions: what may breach it, and who bears responsibility when it is breached.
Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority.
Decision Logs Are Accountability Infrastructure
A boundary that no one records decays. This is the role of the Decision Log.
A Decision Log is not log management. It is not a technical record of outputs kept for debugging. It is Accountability Infrastructure: the continuous record of who held authority over which decision, where the boundary was crossed, and on what grounds.
Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes.
In a system where humans and AI agents make interleaved decisions, judgment becomes distributed. Distributed Judgment without a log dissolves accountability — no one can later reconstruct who actually decided. The Decision Log is what keeps authority traceable across that distribution.
Implementation: An AI Agent Approval Workflow
Take a concrete case. A company puts expense approval in the hands of an AI agent. The naïve design lets the agent read the receipt, check it against policy, and approve end to end. Decision Design instead asks first where the Decision Boundary falls.
The Decision Boundary here is set in three layers. Small claims that fully match policy: the agent approves to the end. Claims above a set amount, or where policy admits interpretation: the agent assembles the material and hands the decision to a human. Any claim involving a counterparty with a past fraud signal: a human takes it on regardless of amount.
The Authority Structure makes these layers institutional rather than personal. It names who holds final authority in each layer, and who answers when an exception breaches the boundary — not the individual reviewer's discretion, but a documented allocation.
The Decision Log records each pass: which layer the claim fell into, whether AI or a human held authority, what the exception was, and who owned it. When someone later asks "who approved this," the log answers. Accuracy may widen the layer delegated to AI over time. What decides how far it widens is not accuracy but authority design.
The same structure governs a financial credit approval. Format checks and routine scoring sit on the AI side of the boundary; there is little room for judgment. Exception approvals and final denials sit on the human side, where value judgment enters, denial demands an explanation, and the outcome leaves an irreversible mark on an applicant's business or life. The AI may present a score. The Judgment Authority to grant or deny stays with the human reviewer, and the Decision Log preserves how a conflict between score and reviewer was resolved. Skip this design, and the score becomes the decision in practice while human review degrades into ritual. Authority migrates to AI that no one intended to move — and no one can say whether "the AI declined" or "a human declined."
Who Ultimately Decides
Return to the opening question. At the table in Évian, who ultimately decides? No one at that table decided. The sovereigns of geography and the sovereigns of intelligence debated the plumbing called Standards and adjourned without naming who holds the tap — the Judgment Authority that flows through it.
What must be designed sits before Standards. Brookings identified the right governance problem. Decision Design identifies the layer beneath it: who holds authority, how far it is delegated to AI, where a human takes over. Governance is not enough. The age of Enterprise AI and autonomous AI Agents needs a discipline whose object is the placement of authority itself. That discipline is Decision Design.
AI will keep getting smarter. Left undesigned, authority flows to AI by default. Machines can execute judgment. Only humans can decide who holds the authority to judge.
Sources: Tom Wheeler, "G7 should accept AI standards offer, but make it enforceable," Brookings Institution, July 1, 2026; AI Business Guidelines Version 1.2, Ministry of Internal Affairs and Communications (MIC) and Ministry of Economy, Trade and Industry (METI), 2026.
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.