Organizations describe their AI programs as automation while handing parts of organizational judgment to software. When that transfer lacks defined authority, boundaries, records, and accountability, a gap opens between what the technology can do and what the organization can answer for. Decision Design provides the architecture needed to close that gap.
A Refund No One Could Account For
On a Saturday morning, a refund appears in the admin console of a subscription business. An AI agent read the customer's message, checked the contract and usage history, and issued the refund. The amount fell within company policy. The transaction was processed correctly.
On Monday, a manager asks why the refund was issued. The team cannot answer. The system holds the outcome but not the information the agent consulted or the alternatives it considered. The manager then asks whose judgment the refund represents. The organization has no answer.
The failed explanation exposes a problem that speed and accuracy metrics miss. As a company introduces AI into operations, judgment once exercised by people begins to move to AI systems. If the company increases autonomy without designing that transfer, it can execute a valid transaction yet remain unable to account for it. A recent survey of Singapore and Australian organizations, read alongside Japanese government guidance, shows why this gap deserves attention.
Task Automation Is Becoming Judgment Delegation
A few years ago, AI played a narrow role at work. It summarized text, produced drafts, and suggested how to categorize an incoming request. A person still pressed the final button, and the AI stayed a tool that produced answers.
AI agents now operate external systems. They issue refunds and discounts, change contracts, and suspend accounts. Unless a person checks the last step, the agent's output becomes a real transaction. The AI is shifting from a tool that produces answers to an actor that produces outcomes. Calling it an actor does not make the agent a legal or moral bearer of responsibility. That is exactly why the question matters: someone inside the organization has to own the outcome.
Efficiency describes only part of this shift. The organization processes work faster because part of the judgment people once exercised has passed to AI. The Saturday refund is one example of that broader transfer.
Task automation and judgment delegation are different acts. Automation runs a defined task more quickly and cheaply. Delegation transfers the interpretive work of reading a situation and choosing a course of action. Many organizations believe they are doing the first while they are actually doing the second.
Singapore's AI Autonomy Paradox
Evidence for this shift comes from a survey by the technology firm Insight, reported by IT Brief Australia on 6 July 2026. The respondents were 220 business decision-makers at Singapore organizations with more than 100 employees, alongside 318 in Australia. Insight published a related report on the same theme, Assistance to Autonomy, and reporting by ARN cites the same sample sizes and links the results to that report. The IT Brief article itself does not name the report, so the figures below are attributed to IT Brief's reporting.
According to the Insight figures reported by IT Brief, 37% of Singapore organizations had deployed AI across multiple departments, and 14% had fully embedded AI into their operations. Fifty-five percent of Singapore respondents reported moderate to strong returns on AI initiatives, compared with 41% in Australia. In this sample, Singapore organizations reported both broader deployment and stronger returns.
The same survey records a readiness gap. Only 20% of Singapore leaders said their organizations were very prepared for autonomous AI, while roughly 40% described them as somewhat prepared. Integration with legacy systems was the most frequently cited barrier at 33%, followed by cost and infrastructure constraints at 25%. Half of Singapore's decision-makers said their trust in AI fell in high-risk situations.
Insight calls this condition the autonomy paradox. Organizations expand AI autonomy before their governance, trust, and control capabilities are ready to support it. Oversight then has to catch up with authority already granted.
Why Delayed Governance Is an Incomplete Diagnosis
The survey invites a familiar diagnosis: AI adoption is advancing faster than governance. That diagnosis is accurate but incomplete.
When leaders say governance lags, they often mean that rules and approval structures have not caught up with adoption. The usual response is to set policy, tighten controls, and add oversight. Those measures would not resolve the Saturday refund. The company did not lack a refund rule; the amount complied with one. It lacked an explicit answer to a different question: may this kind of judgment be delegated to AI, and if so, under what conditions? No one had drawn that line.
The real issue is the movement of judgment. Judgment that people held implicitly passes to the AI without being designed. Which choices does the organization let the AI make, and which does it keep for people? Who approves, and who owns the result? Raising autonomy without specifying that boundary is the practical content of the autonomy paradox. Publishing a governance policy does not, by itself, settle any individual boundary of judgment.
Japan's AI Guidelines Point at the Same Gap
The same concern appears in Japanese government documents. On 31 March 2026, the Ministry of Internal Affairs and Communications and the Ministry of Economy, Trade and Industry published the AI Guidelines for Business, Version 1.2 (AI事業者ガイドライン 第1.2版). This edition addresses the autonomous operation of AI agents and use cases in which an agent works with external systems to produce outcomes.
The guidelines are not binding regulation. They set out voluntary practices for businesses. Within that frame, Appendix 1 (page 19, footnote 13) notes that as agent autonomy increases, oversight by humans alone may struggle to keep pace with fast interaction between AI systems, and it points to new safeguards worth considering, such as mutual monitoring among AI systems. Appendix 5 (page 175) addresses cases where AI output is used to evaluate individuals or groups: it calls for reasonable human judgment that accounts for automation bias, and for accountability to the evaluated party when they request an explanation.
The guidelines do not mandate uniform human approval for every action an AI agent takes, and they do not treat a human in the loop as a guarantee of safety. An organization must combine final judgment, monitoring, recording, and stop mechanisms according to the use case and its risks. Although the Singapore survey and the Japanese guidelines address different contexts, both indicate that wider delegation requires a corresponding control architecture.
The Question Organizations Have to Answer: Who Is Deciding?
An organization already using AI can ask three questions. Which judgments has it delegated to AI? Who authorized that delegation and owns the resulting decisions? Can the organization reconstruct and explain each judgment when challenged?
In most workplaces, AI adoption advances in units of tool and task: which tool should automate which process. The question of which judgment to delegate, and how far, dissolves into day-to-day operations instead of becoming an explicit design object. The team in the opening scenario did not fail through negligence. No one had designed the boundary around the delegated judgment.
The missing element is a designed boundary between human and machine judgment. More capability cannot supply that boundary, and more rules do not automatically define it at the level of an individual decision.
What Is Decision Design?
Decision Design is a judgment architecture framework, proposed by Ryoji Morii, founder of Insynergy Inc., that treats the act of judgment itself as an object of design. It specifies, for each recurring decision, who may decide, what evidence supports the judgment, which authority has been delegated, where the delegation ends, what triggers human intervention, who approves or owns the result, what must be recorded, and how the organization reviews and updates the boundary. Its central construct is the Decision Boundary, the line that separates what the organization delegates to AI from what it keeps for people.
Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate. The distinction matters because a faster or more accurate refund decision still fails the organization if no one can say who was entitled to make it. Technical permission to execute a refund is not the same as legitimate authority to do so. Producing a correct output is not the same as being able to account for the outcome it caused. Decision Design works on the second half of each pair: the authority, the boundary, and the record that let an organization stand behind what its AI did.
Decision Design complements AI governance, AI ethics, digital transformation, and automation. It connects organization-wide principles to specific judgment processes; it does not supersede those fields or seek to sharpen an individual's cognition.
How Decision Design Differs from AI Governance and Related Frameworks
Four established frameworks perform necessary work but do not, by themselves, define the boundary around an individual judgment. Decision Design connects their organization-wide aims to that operational unit.
Governance sets the policy, control, and oversight for AI use. It defines who is accountable and which principles govern operation across the organization. A policy does not descend on its own to the unit of a single judgment, such as whether the AI may issue this refund to this customer today.
Digital transformation changes how work and business are done. It removes paper, connects processes, and runs decisions on data. It addresses how work changes, not which judgment inside the changed work a person should own.
Automation runs processing and tasks without manual effort, improving speed and cost. Some automation carries stop conditions for exceptions. Being technically able to halt is distinct from deciding where a process should stop given customer impact and accountability.
AI ethics states what is desirable and which principles to uphold, setting the direction on fairness, transparency, and accountability. Ethics names the principles to honor; it is not the implementation procedure that applies those principles to one refund judgment.
All four are needed, and none of them, on their own, settles who may make a given judgment and under what conditions. Someone has to translate policy, transformation, automation, and principle into the actual unit of judgment. That translation is the work of Decision Design.
Decision Boundaries Define Legitimate Authority
A Decision Boundary specifies who may decide what, under which conditions, and how far that authority extends across a relationship: human and AI, staff and management, or company and customer. A permissions table records access. A Decision Boundary also accounts for the type of judgment, its impact, uncertainty, reversibility, accountability, and effect on personal data. Those factors place a judgment in autonomous execution, human approval, or escalation. The same refund can fall into different categories depending on the amount and the customer's circumstances; the boundary defines the conditions that move it from one category to another.
Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority. An operational threshold asks whether a system can act. An institutional demarcation asks whether it should be allowed to, on whose authority, and with what recourse if the action turns out to be wrong. Setting a refund ceiling in code is an operational threshold. Deciding that refunds below that ceiling may be executed by the AI, that refunds touching a legal claim may not, and that a named business owner answers for either outcome is an institutional demarcation. The second is what Decision Design draws.
How to Design Boundaries for a Customer-Service AI Agent
Return the abstraction to the opening case. Consider an e-commerce or SaaS company where an AI agent reads customer inquiries, checks contract and usage records, and proposes or executes refunds, discounts, contract changes, and account suspensions. Begin by sorting the judgments in this workflow into three operating categories.
The first category is autonomous execution, where the AI may proceed from judgment to action. A small refund that meets internal policy may fit here. The organization still records the basis for the judgment and the result, and keeps the action reversible where possible.
The second category is human approval. The AI prepares a recommendation but does not execute it. For an exceptional discount or a change to contract terms, the AI presents the available options and supporting evidence. A staff member reviews and approves the action, and the Decision Log identifies that approver.
The third category is human escalation. When a case involves a legal claim, serious complaint, or sensitive information, the AI stops deciding and transfers it to a designated person or team. The boundary specifies the recipient and response deadline. The customer receives a clear notice that the case is on hold pending human review.
Amount alone does not set the boundary. The design combines reversibility, potential harm to the customer, sensitive information, model confidence, departure from precedent, legal and contractual exposure, and the presence of a dispute. A small refund that carries a legal claim moves to a person. A high-frequency, fully reversible action leaves room for autonomous execution. Conditions must be stated in terms a system can evaluate and an auditor can check, not in vague language such as "if it feels risky."
The Operating Components of Decision Design
Six mechanisms turn a Decision Boundary into an operating practice. Together, they identify the judgments that require boundaries, assign authority, define intervention, preserve accountability, and provide a basis for revising the design.
Decision Inventory identifies the judgments embedded in a workflow before the organization tries to govern them. It reveals where people or AI systems interpret circumstances, compare options, and choose an action. The result is a documented list of judgments already delegated to AI, along with judgments that the organization may delegate in the future.
Boundary Matrix assigns each identified judgment to autonomous execution, human approval, or escalation. For every judgment, it specifies who may decide, under which conditions, which exceptions change the classification, and which business owner remains accountable. The matrix converts a general policy into an explicit allocation of decision authority.
Escalation Trigger defines the verifiable conditions that stop an AI process and return a case to a person. These conditions may include an amount, an irreversible action, detection of a legal claim, or a mismatch in identity verification. A single keyword should not determine a consequential escalation without supporting context. The trigger specifies when the AI must stop, who receives the handoff, and how the organization handles false positives.
Decision Log preserves the inputs, policies consulted, options considered, output, model or rule version, approver, and result of a judgment. Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes. An ordinary system log tells an engineer what the software did. A Decision Log also shows who was entitled to make the judgment, which evidence supported it, and who owns the consequence. Because storing personal data creates its own exposure, the record defines its purpose, access rights, retention period, and deletion conditions.
Override gives an authorized person the ability to stop, reverse, or correct an AI judgment. It ensures that delegated authority remains recoverable when circumstances change or the AI makes an error. The organization records the intervention, its reason, the person who made it, and the resulting action.
Boundary Review uses errors, exceptions, complaints, reversals, and operating data to test whether a Decision Boundary remains appropriate. The review may widen autonomy when evidence supports further delegation or narrow it when the current design creates unacceptable risk. Each change leaves a documented rationale, allowing the organization to show why the boundary moved.
Each mechanism needs a business owner. That owner tracks the autonomous-execution rate, referral rate for approval, reversal rate, complaint rate, and number of judgments the organization could not explain. These measures support evidence-based changes to the boundary. Model accuracy remains relevant, but it cannot show whether the delegation of authority is working as designed.
Redesigning the Saturday Refund
Decision Design makes the omissions in the Saturday refund specific.
If the amount sat within policy, the refund might have belonged in autonomous execution, though amount alone would not settle it. A Boundary Matrix would have fixed the applicable conditions, the exceptions, and the business owner who answers for the result. A Decision Log would have preserved the contract details and policy consulted and the reason for the action, so the Monday question had an answer. An Escalation Trigger would have returned the case to a person the moment it detected a legal claim or a mismatch in identity verification.
The unease that morning was not that the AI issued a refund. It was that the organization could not say on whose authority, or under what conditions, it had delegated the judgment the AI carried out.
Organizations that let AI produce outcomes need more than rules for managing the technology. They need a method for deciding which judgments to delegate, to whom, under what authority, and with what means of explanation and recourse. Drawing those boundaries before expanding autonomy gives the organization a basis for authorizing AI action and answering for its consequences.
FAQ
Is Decision Design just another name for AI governance?
No. AI governance sets policy, control, and oversight across the organization. Decision Design translates those principles into the unit of a specific judgment by defining delegation, intervention, and accountability. It determines whether an AI may make a given judgment under stated conditions and identifies who remains responsible for the result.
Where should an organization start with Decision Design?
Start with a Decision Inventory. List the judgments that current workflows already delegate to AI, then classify each one as autonomous execution, human approval, or escalation. Use criteria such as financial impact, reversibility, personal-data exposure, legal consequences, and the availability of recourse. Begin with the judgments that could produce the greatest harm or accountability burden.
Do Japan's AI Guidelines for Business require human approval of every AI decision?
No. Japan's AI Guidelines for Business, Version 1.2, are voluntary guidance rather than binding regulation, and they do not require human approval of every AI decision. For the evaluation of individuals or groups, the guidelines call for reasonable human judgment and accountability. They also note that human oversight alone may prove insufficient for highly autonomous AI. The appropriate controls depend on the use case and its risks.
References
- Insight, Assistance to Autonomy: How much control are leaders willing to give up as AI becomes more autonomous? (survey of 220 Singapore and 318 Australian business decision-makers at organizations with more than 100 employees).
- IT Brief Australia, Singapore firms scale AI as governance lags behind, 6 July 2026.
- ARN, Australian orgs reach 'autonomy paradox' with AI, reporting on the same Insight survey.
- Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry, AI Guidelines for Business, Version 1.2 (AI事業者ガイドライン 第1.2版), published 31 March 2026.
- Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry, AI Guidelines for Business, Version 1.2, Appendices (別添) (PDF; cited passages: Appendix 1, page 19, footnote 13; Appendix 5, page 175).
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.