As AI agents begin making operational decisions in finance, hiring, healthcare, and infrastructure, the global governance debate has focused on liability — who is responsible when AI causes harm. But liability frameworks operate after the fact. They assign consequences once damage has already occurred. The deeper issue lies earlier: how decisions are structured before AI systems are deployed. This article introduces the concept of Decision Design — the deliberate architecture of organizational decision-making in human-AI systems. It explains how three boundaries shape accountable AI governance: the Decision Boundary (organizational governance), the Human Judgment Decision Boundary, and the Governance Decision Boundary. Without explicitly designing these boundaries, organizations risk creating systems where AI influence expands silently while human accountability becomes purely formal.

AI Agent Liability Is the Wrong Debate — The Real Problem Is Decision Architecture

Ryoji Morii / Insynergy Inc.

AI Agents Are Already Making Consequential Decisions

AI agents are no longer a future scenario. They are responding to customer inquiries, booking travel, processing loan applications, and screening job candidates — today, in production systems.

For most of their history, AI systems were tools. A tool is passive: it amplifies human action, but the human remains the actor. An AI agent is something different. It initiates, selects, and executes. When an agent books a flight, declines a loan, or flags a candidate as unqualified, it is participating in decisions that carry real consequences for real people.

This shift raises a question that neither law nor technology has fully answered: when an AI agent causes harm, who is responsible? Japan's regulators are grappling with this question in real time — and what they are finding is worth attention beyond Japan's borders.

What Japan's Regulators Are Actually Saying

Japan offers an early and instructive data point. The country's two lead ministries on technology policy — the Ministry of Internal Affairs and Communications (MIC) and the Ministry of Economy, Trade and Industry (METI) — are currently revising their AI Business Guidelines to address autonomous agents explicitly, for the first time. The revision is expected to be finalized within the month (The Nikkei, March 9, 2026).

The policy direction emerging from this process is telling. At an expert advisory panel convened in December 2025, Hiroshi Nakagawa, Team Director at the Center for Advanced Intelligence Project at RIKEN, stated directly: "The question of liability for damages caused by malfunctions is an extremely important issue. I expect that incidents involving erroneous bookings or purchases will occur frequently going forward." His concern is not theoretical. It reflects what practitioners in Japan are already anticipating as AI agents move into operational use.

The regulatory response has focused on requiring human involvement at critical decision points — the reasoning being that if a human must approve an agent's action, there is always an identifiable party who can be held accountable. It is a coherent instinct, and it reflects how liability has traditionally been structured.

But Japan's own draft guidance on civil liability for AI use offers a candid acknowledgment of the limits of this approach: responsibility "depends heavily on the specific technology and use case," making concrete attribution "currently difficult." Wataru Shimizu, a lawyer at Anderson Mōri & Tomotsune who advises on AI-related risk, put the practical implication plainly: companies face potential liability unless they have taken "realistically possible countermeasures" — and that means building operational structures that include mutual AI checks and defined authority constraints before harm occurs (The Nikkei, March 9, 2026).

What is visible in Japan's regulatory process is a pattern that will almost certainly appear in every jurisdiction as AI agents become more prevalent: governments reach for liability as the primary governance instrument, while simultaneously acknowledging that liability alone cannot answer the structural questions that autonomous systems raise.

Liability frameworks are built to assign consequences after harm occurs. They say nothing about how decisions should be structured before harm is possible.

The Real Problem Is Not Liability — It Is Decision Architecture

Consider a loan officer at a financial institution. Five years ago, she read the application, asked questions, and made a judgment. Today, an AI system scores the applicant and outputs a label: Approve, Review, or Decline. She looks at the output and stamps the form.

Formally, a human made the decision. Practically, something else happened.

Overriding the AI recommendation carries a cost. She must document why she disagreed, justify her deviation, and accept personal exposure for the outcome if things go wrong. Agreeing with the AI requires nothing. Over time, the path of least resistance becomes the default path. Human judgment does not disappear — it is gradually displaced.

This is the structural problem that liability frameworks cannot address. The substitution of human judgment happens before any harm occurs. By the time a liability question arises, the decision architecture has already failed.

This is what we mean by Decision Boundary (organizational governance): the explicit, designed line that separates what AI decides from what humans decide. In most organizations today, this boundary does not exist as a designed artifact. It exists as an emergent pattern — shaped by convenience, system defaults, and institutional inertia rather than deliberate choice.

When the Decision Boundary is undesigned, it drifts. AI influence expands. Human accountability hollows out. And no one notices until something goes wrong.

When AI Advice Becomes Institutional Decision

The loan officer scenario is not exceptional. It is the standard pattern across industries. In hiring, AI screening tools rank candidates before any human reads a resume. In healthcare, diagnostic support systems surface probabilities that shape clinical framing before the physician forms an independent assessment. In infrastructure monitoring, anomaly detection systems flag conditions that trigger escalation protocols.

In each case, the formal structure preserves human involvement. A person signs, approves, or acknowledges. But the substantive judgment has been made upstream — by a system whose reasoning is not fully visible, whose errors are not fully predictable, and whose recommendations carry an institutional weight that is difficult to resist.

This is the Human Judgment Decision Boundary: the threshold at which a human's involvement transitions from genuine deliberation to formal ratification. When this boundary is not explicitly defined, it collapses under the weight of AI recommendation authority. The human remains in the process, but is no longer making the decision in any meaningful sense.

Requiring "human involvement" — as regulators increasingly do — addresses the symptom without diagnosing the condition. Presence is not judgment. Approval is not accountability. What matters is whether the human at the boundary has the information, authority, and structural incentive to exercise genuine independent judgment.

The Governance Problem

When organizations cannot trace who actually decided something, governance breaks down.

After an AI-assisted decision causes harm, the investigation typically produces the same set of unanswerable questions: Was the AI recommendation flawed? Was human review inadequate? Was the decision criterion itself miscalibrated? Was the AI's authority over the outcome ever formally defined?

In the absence of designed decision structure, these questions cannot be resolved — not because the answers are hidden, but because the structure required to generate the answers was never built.

This is the Governance Decision Boundary: the organizational architecture that defines, at each decision point, who holds authority, who bears accountability, and what the conditions are for human override versus AI delegation. Without this architecture, governance becomes reactive — responding to failures rather than preventing them, assigning blame rather than designing for accountability.

The governance failure is not technological. It is structural. Organizations have invested heavily in AI capability without investing in the decision architecture that makes AI capability governable.

Decision Design: What It Is and What It Is Not

Decision Design is the discipline of designing the structure of decision-making itself — not the tools that support decisions, but the organizational logic that determines how decisions are made, by whom, under what authority, and with what accountability.

What Decision Design Designs

Decision Design addresses four elements:

Decision authority — Which decisions belong to AI, which belong to humans, and which require both? This is not a binary. It is a structured classification based on the nature of each decision: its reversibility, its consequence, the degree to which it requires contextual judgment, and the accountability it carries.

Accountability allocation — Who is responsible for the outcome of a decision, and under what conditions? Accountability cannot be assigned after the fact if it was not designed before. The failure mode of AI-assisted decision-making is not that accountability disappears — it is that it was never placed anywhere specific.

Human-AI role structure — What does the human actually do at each decision point? Is the human a reviewer, an approver, an auditor, or a genuine decision-maker? These are different functions with different accountability implications, and they should be designed explicitly rather than assumed.

Override conditions — Under what circumstances can a human override AI? Under what circumstances must they? Designing these conditions is not optional — it is the primary mechanism by which human judgment retains organizational meaning.

What Decision Design Is Not

Decision Design is not prompt engineering. The question of how to instruct an AI model is a technical question. Decision Design asks who in the organization is responsible for the outputs of that model, and what the structural conditions are for human intervention.

Decision Design is not model optimization. Improving AI accuracy reduces error rates. It does not define who owns the decision when an error occurs, or how the organization should respond when AI and human judgment diverge.

Decision Design is not workflow automation. Automating a process changes who performs a task. It does not determine who holds authority over the outcome of that task. Automation and accountability are different problems.

What Problems It Solves

Decision Design addresses three converging problems: the erosion of clear accountability when AI provides advisory outputs that humans nominally approve; the governance challenge of agentic AI systems that operate across multiple steps without defined human intervention points; and the organizational tendency to treat human presence as equivalent to human judgment.

Toward Practical Governance

Decision Design translates into specific structural implementations:

Judgment Layer — A designated layer within operational workflows where human decision-making is concentrated. Rather than distributing human review across every step, the Judgment Layer identifies where consequential decisions actually occur and ensures that human authority and accountability are anchored there.

Human Review Gate — A defined checkpoint in an AI-assisted process where human approval is not merely formal but structurally meaningful. At a Human Review Gate, the human has access to the full reasoning basis for the AI recommendation, the authority to override without penalty, and explicit accountability for the outcome. The gate is not a bottleneck — it is the point where accountability is formally assigned.

Decision Ledger — A structured record of who decided what, on what basis, at what point in the process. The Decision Ledger is not a compliance log. It is the organizational memory that makes it possible to evaluate decision quality over time, identify where AI influence has exceeded its designed boundary, and refine the Decision Boundary based on evidence.

AI Agent Governance Structure — For agentic systems that operate across sequential steps, an explicit definition of each agent's decision authority, the conditions under which one agent can trigger another, and the points at which human approval is required before the chain continues. Without this structure, agent chains produce accountability gaps — outcomes that no defined party owns.

The Governance Debate Needs to Shift

Japan's experience is instructive precisely because it is not exceptional. A major economy, with sophisticated institutions and genuine policy intent, is finding that liability frameworks and human-involvement requirements — while necessary — do not resolve the underlying structural problem. The same discovery will be made, in sequence, by every regulatory body that takes AI agent governance seriously.

The current regulatory instinct — require human involvement, clarify liability after the fact — is not wrong. It is incomplete.

Liability frameworks establish consequences for failure. They do not produce the conditions for success. An organization that designs its decision architecture well is less likely to produce the failures that liability frameworks are built to address.

What is missing from the current governance conversation is a design vocabulary — a way of talking about decision structure that is precise enough to be actionable, and broad enough to apply across industries and regulatory contexts. Decision Design, and the concept of Decision Boundary in its organizational, human judgment, and governance dimensions, is an attempt to provide that vocabulary.

The question is not only who is liable when an AI agent causes harm. The question is whether the organization had designed, before deployment, a structure in which human judgment remained meaningful, accountable, and traceable. If the answer is no, the liability question is downstream of a deeper failure — a failure of governance architecture that no regulation, by itself, can prevent.

Ryoji Morii is the founder of Insynergy Inc., a Tokyo-based management consulting firm specializing in AI governance and Decision Design. Decision Design™ and Decision Boundary™ are conceptual frameworks developed by Insynergy Inc.