At the G7 summit in Évian, France, in June 2026, a small moment marked a shift in the AI debate. This essay starts there. Its purpose is not to explain the news.
Inside the phrase "AI governance" sits a domain that has no name yet. Read to the end and you will see it.
One question runs through everything that follows.
Who decides?
That question exists in the same form from a summit conference room to the desk where you sit. We will return to it at the close.
The material is an essay published in Modern Diplomacy on June 28, 2026: "From the IAEA to the G7: The Contested Meaning of Global AI Governance," by Tuhu Nugraha.
Two Versions of Global AI Governance
In May 2026, hours before a scheduled meeting between the U.S. and Chinese heads of state, Chris Lehane, OpenAI's Vice President of Global Affairs, floated a proposal: a U.S.-led "international agency for AI" that would include China as a member. The model was the International Atomic Energy Agency (IAEA), the body that has managed the civilian use of nuclear technology across rival states. Put everyone at the table, adversaries included, and govern a strategic technology collectively. Call this the IAEA model.
One month later, the G7 summit produced the opposite tone. According to reporting by Axios, Anthropic's Dario Amodei and Google DeepMind's Demis Hassabis leaned toward a more selective framework among democracies: manage access to high-capability models, and the security risks they carry, inside a circle of trusted allies. Call this the G7 model. OpenAI's Sam Altman took a more neutral line, calling for an international forum to agree on shared testing standards and risk evaluation.
The same phrase, "global AI governance," meant "secure legitimacy by including China" in May and "contain strategic risk inside a trusted coalition" in June. The words held; the content reversed.
What is contested here is not how to manage AI. It is who decides. The clash between the IAEA model and the G7 model reduces to that single point.
The Question Hidden Behind Global AI Governance
Most readers treat "AI governance" as a problem of technical control: who manages a technology that might run out of hand, and how. That is not what the Modern Diplomacy essay is about.
The real contest is not who manages AI. It is who sets the rules.
Nugraha calls this the power to define the word "global." What counts as a "frontier model" — the most advanced foundation model. What benchmarks test its capabilities. Who is granted access. Only a small number of companies and states can decide these things.
The essay names a double asymmetry. The first is technical: the set of actors able to define a frontier model is small. The second is narrative: those same actors frame the very language used to talk about governance.
India, Brazil, Kenya, South Korea, and Egypt are sometimes invited into the G7 dialogue. But the essay draws a line between attending a forum and shaping its design. The power to take a seat is not the power to set the shape of the table or its agenda.
So Global AI Governance is a question about the allocation of authority before it is a question of technical control. Who holds the authority to decide, and who receives the decided outcome downstream. That is where the contest sits.
Why This Contest Over Authority Matters Now
Three structural shifts pushed this contest into the foreground now.
First, AI became a matter of national security. Shortly after the G7 discussions, the U.S. government imposed export controls, and Anthropic had to cut off foreign access to its own models, Fable 5 and Mythos 5. According to Reuters, the company restricted access more broadly to comply with the rules. Even among democratic allies, technical solidarity has a limit. The moment AI became strategic infrastructure, states began calculating their own discretion.
Second, AI companies acquired state-scale influence. According to Stanford HAI's AI Index Report 2025, roughly 90% of notable AI models released in 2024 came from industry, up sharply from about 60% in 2023. Frontier research and development shifted its center of gravity away from universities and public institutions toward private firms that control capital, compute, talent, and data. The result was visible at the G7 table, where AI company CEOs sat and spoke not as observers but as negotiating parties.
Third, "global" became a political word. As shown, its meaning inverted between May and June. When the political situation changes, the scope of "global" changes with it. A word that appears neutral and universal is in fact defined from someone's position. That is the point the essay keeps pressing.
The three shifts look separate but converge on one place. The deciding party is moving toward the few, and toward the corporate. The question does not change. Who decides?
The Same Question, at Smaller Scale
So far this has been about states and companies. But does "who decides?" belong only to the stage of international politics?
Scale it down.
The same question sits in the corporate boardroom. Which investments does the executive team decide, and where does authority pass to the field? Scale down again, into a department. Does this approval come from the manager or the individual handling it? One level lower, it sits on your desk. Did you make this call, or did a system, or someone in between?
State, company, department, individual. The scale differs. The shape of the question does not change by a single character. Who decides? How far is delegated, and where does responsibility resume?
There is one more shared feature. In most cases, no one has drawn the line. It exists by custom, by drift, by "that's how it's done." In international politics, treaties and export controls fight to draw the line. In organizations, most of the time, no fight even occurs.
Who decides? Hold on to the question. It returns, harder.
Government Guidance Is Already Moving
This is not an abstract thought experiment. It has already entered the language of regulation.
Governments are asking developers to build "mechanisms that require human judgment" for autonomous AI agents, with malfunction and privacy risks in mind. That thinking appears in the AI Business Guidelines (Version 1.2), published on March 31, 2026, by Japan's Ministry of Internal Affairs and Communications and its Ministry of Economy, Trade and Industry.
What is new in Version 1.2 is a formal definition of the AI agent. The guidelines describe it as an entity that judges and acts autonomously toward a goal on a human's behalf, connecting with external systems to carry a sequence of business processes across functions. Precisely because autonomy is high, the guidelines call for appropriate assignment of authority, appropriate human intervention in judgment, and periodic review and reporting of operational logs.
The point not to miss is that the guidelines do not merely ask that "a human judges." They ask for a design: how authority is set, at what moment human judgment intervenes, and how far it reaches.
The question moves from "does a human judge or not?" to "how do we design who judges, and how far?"
What Governance Cannot Fully Explain
Line these up and a gap appears.
In international politics, actors fight over who sets the rules. In regulation, governments require "human intervention in judgment." But neither reaches the design of judgment itself.
The existing vocabulary cannot fill that gap. Align the terms and compare what each one designs.
Governance designs rules. DX (digital transformation) designs how work is done. Automation designs how processing runs. AI Ethics designs the values to uphold.
All four matter. None of the four designs judgment itself. Rules sit above judgment; work is the vessel that holds it; automation executes it; ethics points its direction. Judgment itself — who decides, and how far — falls into the gaps between the four.
So what designs that gap?
Decision Design: The Name for the Missing Layer
What was missing is a view that treats the act of judgment itself as an object of design.
That is Decision Design™.
The order matters, though. AI did not create this problem.
The location of judgment was ambiguous long before AI. Approvals no one can later explain. Sign-off lines that exist only by custom. Organizations that ran on "the higher-ups decide, more or less." The boundary of judgment was never drawn. Things still worked, because humans quietly reconciled the accounts.
When AI agents began to act autonomously, that reconciliation stopped working. Machines began inserting judgments into the gaps that human discretion used to fill. In that moment, the location of judgment — invisible until then — became visible all at once.
Decision Design became necessary not because AI created a new problem, but because AI made an old ambiguity visible. It is not a concept invented on a whim. A domain left undesigned for decades has finally been given a name.
The earlier question — "who decides?" — is answered here. The same question that states, companies, and societies were fighting over now lives inside your organization's daily judgments, in the same form. Only the scale differs.
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
What Decision Design designs
Decision Design designs the allocation of decision authority: which judgments an actor may make, how far that authority extends, where it passes to another actor, and how a judgment that exceeds its range is escalated. The object of design is not a deliverable or a system, but the structure of the act of judgment.
What Decision Design is NOT
It is not a mechanism for managing AI. It is not a technique for evaluating model safety. It is not an ethics charter that declares values to protect. It is not a project to replace work with digital processes. It sits next to these, but its object differs. The object of Decision Design is the boundary of judgment: who decides, and how far.
What problem Decision Design solves
It solves the problem of judgment being handed back and forth between humans and machines without anyone noticing. When AI agents begin to act autonomously, a judgment is made in fact even though no one decided it explicitly. Decision Design dissolves this "decided-without-deciding" state by drawing the boundary of judgment deliberately.
What a Decision Boundary Is
The central concept of Decision Design is the Decision Boundary™.
A Decision Boundary is the smallest unit at which judgment authority is allocated: the line that separates who decides, and how far. If Decision Design is the discipline, the Decision Boundary is the unit that makes it operable.
Compare a world with boundaries to one without, and the difference is sharp.
With no boundary, an AI agent auto-executes an ¥800,000 refund, and afterward no one can say who approved it. With a boundary, the same refund catches on a line — "any amount above ¥500,000 requires human approval" — routes to the accounting manager, and leaves a record. Same system, same amount. The only difference is whether a boundary was drawn.
A single boundary is built from five elements.
Decision-maker — who actually makes the judgment: a human, an AI agent, or a combination.
Authority — how far that actor may decide alone, bounded by amount, scope of impact, and reversibility.
Escalation — when a judgment exceeds the authority, to whom it is raised: a superior, a specialist function, or a human as such.
Override — the path by which a human can cancel or overwrite a judgment the AI made. Only when this exists does delegation become safe delegation.
Human Review — the mechanism by which a human checks the AI's judgment before or after the fact: every case, a sample, or only cases past a threshold.
Drawing a boundary means deciding each of these five explicitly rather than leaving them vague. Naming the decision-maker alone is not enough. Only when authority range, escalation target, override path, and review frequency are all set does a single boundary exist.
Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority.
Why Governance Alone Is Not Enough
Decision Design fills the gap that opens between the four existing concepts. Align what each designs and does not design, and the shape of the gap appears.
Governance designs the rules and responsibilities of the whole organization. But it does not descend to who carries each judgment, and how far.
DX redesigns how work is done. But it does not ask who decides in the reworked process.
Automation designs the auto-execution of processing. But it does not draw the line between what is delegated to the machine and what a human resumes.
AI Ethics designs the values to protect. But it does not show where in today's judgment flow those values are embedded.
All four design the surroundings of judgment: the rules, the work, the processing, the values. None designs the placement of judgment itself — who decides, and how far.
Decision Design steps one level below these four and makes the boundary of judgment its object. It therefore does not compete with Governance. Inside the framework Governance sets, Decision Design draws the concrete placement of judgment.
The Same Structure, from the Planet to the Enterprise
Return to the essay that opened this piece.
In Modern Diplomacy, states, companies, and societies fought over who sets the rules. IAEA model or G7 model. Include China, or fence it inside a democratic coalition. This was a fight over where to draw a planetary Decision Boundary.
A fight of the exact same structure runs every day inside organizations, at a smaller unit. Who approves this expense? How far do we let an AI agent decide on credit? May the AI send this email automatically, or does a human check it first?
The scale shrinks from the planet to a single department. The shape of the question holds. Who decides? How far is delegated, and where does responsibility resume? In international politics, treaties and export controls draw the line. In organizations, Decision Design draws it.
How to Implement Decision Design
To keep this out of the abstract, work through a concrete case: deploying an AI agent that autonomously processes invoices. Proceed in this order.
1. Break the work into stages. Split the agent's work into stages — reading the invoice, matching amounts against purchase data, drafting the journal entry, executing payment. For each of the four stages, decide which are delegated to autonomy and which require a human. Skip this and hand "invoice processing" wholesale to the AI, and no boundary can be drawn.
2. Draw the Decision Boundary (authority). Draw the line along amount and counterparty. For example: payments under ¥100,000 to existing counterparties execute autonomously; payments of ¥100,000 or more, or involving a new counterparty, require human approval. That concrete line — "¥100,000," "existing versus new" — is the form of authority. Set the threshold from the historical error rate and the operational load per case.
3. Define escalation. Cases that exceed the boundary — over the limit, a matching mismatch, a new counterparty — escalate to an accounting staff member. Cases above a further amount, say ¥1,000,000, require managerial approval. Document the escalation targets and conditions before going live.
4. Set Human Review frequency. Decide how autonomously executed cases are checked. Sample sub-threshold auto-processing monthly. Review every case near the threshold on the next business day. Vary the intensity of review by amount band. If a human reviews every case, autonomy had no point.
5. Provide an Override path. Give humans a path to cancel a payment the agent executed within a set window — say, 30 minutes from the scheduled execution. Delegating to autonomy and being unable to reverse are different things. Delegation becomes safe only in a form that can be overwritten.
6. Keep a Decision Log™. Record which judgment was made, by whom (human or agent), and on what basis. This corresponds to the "review and reporting of operational logs" that Version 1.2 requires, and it becomes the material for revising boundaries later. Without a log, you cannot verify whether the boundary was right.
Compiled into a single view, the above becomes an authority matrix. Put stages in the rows and decision-maker, authority, escalation, override, and human review in the columns, and it doubles as an operational document you can distribute as is. In the first month, run the matrix as a working draft; watch the record of errors and escalations, and redraw the thresholds each month. A boundary is not drawn once and finished.
Authority matrix versus RACI. Distinguish this clearly from the existing RACI chart (Responsible, Accountable, Consulted, Informed). RACI assigns roles to a task: who executes, who is accountable, who is consulted, who is informed. RACI is a static table of task allocation; it does not anticipate an autonomous decision-maker like an AI agent, or the exception handling when a boundary is exceeded. Decision Design designs not who owns the task, but the boundary of judgment itself: how far to delegate, where to resume, how to escalate past the limit, how to override. If RACI is a "who owns it" table, the Decision Boundary is a "how far one may decide" line. The two can complement each other, but they are not the same.
Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes.
Conclusion
Decision Design is not an idea for managing AI. It is an idea for designing judgment.
When we began reading the Modern Diplomacy essay, we thought we were discussing Governance — who manages AI, and which international body fits. By the end, something else is in view.
The clash between the IAEA model and the G7 model, the government guidance on "human intervention in judgment," the approval on your desk — at the root of all three was one and the same question.
Who decides? How far is delegated, and where does responsibility resume?
That boundary exists in the same form in a planetary conference room and in your organization's approval flow. And in most cases it is left in place unnoticed. International politics can at least fight over the line. Organizations, most of the time, arrive at a decision with no fight at all.
To redraw that line deliberately. To design judgment itself — not Governance, not DX, not Automation, not AI Ethics.
That is Decision Design.
This essay was never only about Global AI Governance. That was the surface. From the first line, it was about authority — about who decides.
Primary Sources
- Tuhu Nugraha, "From the IAEA to the G7: The Contested Meaning of Global AI Governance," Modern Diplomacy, June 28, 2026.
- Stanford Institute for Human-Centered Artificial Intelligence (Stanford HAI), AI Index Report 2025 — approximately 90% of notable AI models in 2024 originated in industry.
- Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry (Japan), AI Business Guidelines (Version 1.2), published March 31, 2026.
- G7 Summit 2026 (Évian, France); International Atomic Energy Agency (IAEA).
- Reuters (reporting on Anthropic's response to export controls); Axios (reporting on AI executives' remarks at the G7).
- United Nations, Global Digital Compact (adopted September 2024).
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.