In most enterprises, an algorithm is already making decisions that once required a person.
It is flagging a loan application. It is routing a customer complaint. It is screening a stack of résumés. Each is a decision. And for each, most companies cannot cleanly answer one question: if the decision is wrong, who takes it on?
This is not a legal question, and it is not a philosophical one about the nature of responsibility. It is an operational question. When the outcome lands, who—by name—owns it?
Executives have started to confront this question. The more useful problem sits one step beyond it. The absence of an AI owner is not the real failure. There is a deeper structural gap that naming an owner does not close, and that gap is the subject of this article.
Executive Summary
Enterprises are being told, correctly, to fix AI Governance by assigning Accountability: name a senior owner for every consequential AI decision, build an audit trail, and make escalation and override work in practice rather than on paper. This is necessary and insufficient. Naming an owner declares where responsibility should land; it does not define what that owner decides, where their judgment ends, and where an AI Agent's begins. Human Oversight and Human-in-the-Loop share the limit: a person can sit inside the loop and still rubber-stamp an output nobody judged. What is missing is a layer beneath governance—the deliberate design of judgment itself. This article introduces that layer as Decision Design and its central concept, the Decision Boundary, and shows through a detailed grant-review example where the line between AI and human judgment should fall. The argument builds on the accountability gap documented by CX Today and the human-involvement requirements in Japan's AI Business Guidelines Version 1.2.
The Governance Gap
The clearest recent account of this problem comes from CX Today. In June 2026, Thomas Walker published "AI Accountability: The Governance Gap Leaders Miss." Read structurally rather than as a summary, the piece makes three moves.
First, ownership of AI Governance is vague. Walker cites a 2025 IAPP survey in which only 28 percent of organizations have formally defined oversight roles for AI Governance. In the remaining roughly seven in ten enterprises, no one formally owns AI compliance or model accountability.
Second, responsibility diffuses. A data science team builds the system, IT deploys it, and a business unit uses it. Ownership thins at every handoff. Walker notes that no single function tends to hold more than a quarter of AI Governance responsibility—IT around 25 percent, risk management 18 percent, a dedicated AI Governance team just 10 percent. The people who built the model do not watch its downstream effects; the people who use it cannot interrogate its internals; the leaders who approved the investment do not monitor its outputs.
Third, oversight after deployment is thin. A model is optimized for objectives that made sense at training time. The business shifts and the regulations shift, but the model stays put, and no one has formally taken on the job of asking whether those original objectives still hold.
From these, the article draws three weaknesses: no single decision owner, no tested escalation path, and legacy risk frameworks applied to systems that drift, learn, and produce outputs no human wrote. Walker's prescription centers on Human Oversight and Accountability. For high-consequence AI, name a senior leader who owns the outcome. Keep an audit foundation that can reconstruct any decision after the fact. Make override and escalation function operationally, not theoretically. His closing test is sharp: stop asking "do we have an AI policy?" and start asking "can we name the person accountable for every important AI decision?"
Why Accountability Alone Is Not Enough
The diagnosis is right. Owners are needed. So is the audit foundation, and so are the escalation paths. Yet something does not resolve.
Does naming an owner actually establish Accountability?
Write one name into a box. Add a title—"AI Decision Owner"—to the org chart. Does that settle who takes on a wrong decision?
Accountability arises after a judgment. If it is unclear whose judgment it was, assigning responsibility afterward does not make it stick. Accountability is an outcome; judgment is a process. Design only the outcome, leave the process undefined, and the name you assigned has nothing to attach to. Naming an owner declares where responsibility should sit. It does not tell you how far that owner's own judgment extends, what they delegate to an AI Agent, and where a human takes the decision back. Without that line, the owner keeps the title "the person accountable" and remains suspended above the actual work.
This unease is worth holding, because the same gap appears far beyond a single company.
Governments are stuck at the same point. Concerned about malfunction and privacy harm from AI Agents that act on their own, regulators are asking developers to build in mechanisms that require meaningful human involvement. Japan's Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry set out this thinking in the AI Business Guidelines Version 1.2, published on March 31, 2026. Version 1.2 formally defines autonomous AI Agents—systems that judge and act on their own toward a goal—and requires human involvement in consequential decisions. It asks organizations to specify which data must never be entered, to keep logs of inputs, outputs, and tool calls, and to design points of error detection and human approval into each stage of a workflow.
The policy detail is not what matters here. The enterprise floor and the national guideline are pointing at exactly the same place. Both say: do not hand everything to the machine. Both say: keep human judgment in the process. Neither yet says, in concrete terms, what "keeping human judgment" actually consists of. That is the unanswered question.
Why Human-in-the-Loop Is Insufficient
Name an owner. Add Human-in-the-Loop. Put an approval button in front of a person.
But who, in that arrangement, has actually judged?
Consider what the person who pressed approve actually approved. If they ratified the AI's conclusion without examining its substance, were they a decision-maker or a checkpoint the workflow passed through?
You can have an owner, a human inside the loop, and an approval step. Every form is present. And still, when something goes wrong, there may be no one who can say, without flinching, "that was my judgment." Assembling the forms of oversight is not the same as designing the judgment inside them.
Here the shape of the problem changes. What is missing is not an owner. What is missing is the design of judgment itself.
The Missing Layer
Governance names the owner. Human Oversight puts a person above the process. Human-in-the-Loop places a person inside it. Each is a real control, and each shares one blind spot: none of them specifies the judgment the human is supposed to hold.
A person inside the loop with no defined decision to make will approve by default. An owner named without a defined scope of judgment holds a title over an empty space. An approver with no line between what they approve and what they refuse is a button-presser, not an accountable party. All three point at the same void: no boundary has been drawn around the judgment.
Regulators are circling the same void from the outside. When Japan's AI Business Guidelines Version 1.2 require meaningful human involvement for autonomous AI Agents, they are demanding that this boundary exist—without yet describing how to draw it. An AI Agent acting on its own chains several decisions together while no human is watching. If each link has no boundary around it, the moment a human should have stepped in becomes visible only after the fact. Accountability that arrives only in hindsight arrives too late.
The layer that is missing sits beneath governance. It is the deliberate design of judgment—who judges, how far, and where the line falls. That layer has a name.
Decision Design
Define it in one sentence, and hold the definition for the rest of this article. Decision Design treats the act of judgment itself as an object of design.
Everything above asked who bears responsibility. Decision Design asks the question one step earlier: whose judgment is this in the first place? Where does an AI Agent's territory end and a human's begin? It fixes that dividing line in advance.
Naming an owner happens after a judgment is made. Decision Design works on the structure that exists before the judgment. It overlaps with naming an owner and sits one layer beneath it. It also resolves the earlier unease: an owner's responsibility floats because no boundary was drawn around the judgment. With no line, the owner has nothing to claim as their own.
Decision Design is not about improving decisions alone; it is about designing the authority structure within which decisions become institutionally legitimate.
What Decision Design Designs
Decision Design specifies six elements of judgment and assembles them into a structure. Together they form a Judgment Architecture—the arrangement of who holds Authority, who exercises it, and how it moves.
It designs authority allocation. Where does the Authority to make a given judgment actually reside—with which person, or which system?
It designs judgment ownership, distinct from Authority. Who actually makes the call: a human, an AI Agent, or a combination? Authority and the person exercising it can diverge, and Decision Design refuses to leave that divergence unmanaged.
It designs delegation. How much of a judgment a human once held is passed to an AI Agent, and under what stated conditions.
It designs escalation, the reverse of delegation. When an AI Agent reaches a situation it cannot or should not judge, to whom does the judgment rise, and on what trigger? The condition and the destination are set in advance.
It designs override. Which human can overturn a judgment an AI Agent made, and under what conditions. If anyone can override at any time there is no boundary; if no one can, the boundary is too rigid to be real.
It designs accountability continuity. Once a judgment is made, who ultimately takes on its result—and how that Accountability is preserved as the judgment moves across delegation, escalation, and override. Accountability is the inverse of Authority. Design how far someone may decide, and you can design who takes on the result; leave Authority undesigned, and Accountability has nothing to bind to. Decision Design ties the two back together, one judgment at a time.
What Decision Design Is Not
Decision Design is easy to confuse with concepts it is not.
It is not Governance. Governance designs who the owner is. Decision Design designs what that owner judges. Governance sets the address responsibility is sent to; Decision Design defines the judgment that arrives there. The gap looks small and is decisive, because an owner's name means nothing until the boundary around the judgment is drawn.
It is not DX. Digital transformation moves work onto digital rails. Decision Design handles what remains after the move: who decides. Digitizing a process does not draw the boundary around judgment automatically.
It is not Automation. Automation shifts processing from people to machines. Decision Design determines, at the destination, how much of the judgment belongs to the machine and where a human takes it back. Automation does not remove judgment; it relocates it.
It is not AI Ethics. AI Ethics asks whether a judgment is fair and free of discrimination. Decision Design asks who makes that judgment and who bears it. Ethics concerns the merit of the decision's content; Decision Design concerns the placement of the people who hold it.
None of this is a rejection of those disciplines. Decision Design does not belong to any one of Governance, DX, Automation, or AI Ethics. It runs beneath all of them, one architectural layer down, fixing a single thing they each leave open: where to place the holder of the judgment. It complements them precisely because every one of them contains a spot where Decision Design is required.
What Problem Decision Design Solves
Decision Design is not an abstract ideal. It answers the concrete dead-ends already seen.
It solves the Governance Gap. The gap CX Today documents is not merely an absence of owners; it is an absence of drawn boundaries. Naming owners without boundaries leaves the same void one level down.
It solves authority ambiguity. Governance names an owner but does not define the range of judgments that owner should make. Naming the party still leaves the outline of their judgment blank. Decision Design draws that outline.
It solves responsibility diffusion. At each handoff from builder to deployer to user, ownership thins because no boundary marks where one party's judgment ends and the next begins. Boundaries stop the thinning.
It solves the Human-in-the-Loop limitations. A person in the loop whose judgment is undefined turns approval into ratification. Deciding which specific judgment that person takes on is what makes their presence in the loop mean something. Being in the loop and taking on a judgment are different things.
It solves the problem of judgment ownership. An approver becomes an accountable party only once they hold a line between what they approve and what they refuse. Without that line they press a button. With it, they own a decision.
All of these point at one void: no boundary has been drawn around the judgment.
Decision Boundaries
The line that draws that boundary has a name. A Decision Boundary is the line that divides who decides how much and where a human takes the judgment back. Hold this definition, too, for the rest of the article.
State it in the negative to avoid a common misreading. Decision Boundaries are not operational thresholds; they are institutional demarcations of legitimate authority. A Decision Boundary is not a physical partition between AI and human. It defines, as an institutional fact, the moment Authority to judge passes from one side to the other. Under which conditions, from whom to whom, does the standing to make a call move? Fixing that in advance is what it means to draw the boundary.
Why is a boundary necessary at all? When an AI Agent and a human share the judgments inside one workflow, and no line separates them, judgment drifts vaguely between the two. Let the result land while it is still drifting, and the result belongs to no one. The boundary stops the drift.
A Decision Boundary is built from three lines.
The first is the authority line. Up to here an AI Agent may decide; from here a human decides. An AI Agent approves up to a set amount, for instance, and anything above it passes to a person.
The second is the override line. It sets the conditions under which a human can overturn an AI Agent's judgment, and names who may do it. Override available to anyone at any time is no boundary; override available to no one is a boundary too rigid to breathe. Stating the conditions is what keeps the boundary alive.
The third is the escalation and delegation line. Under which conditions an AI Agent raises a judgment to a human, and, in the other direction, how much a human hands to an AI Agent. These two set the direction the boundary moves—escalation lifts judgment toward the human, delegation pushes it toward the AI Agent.
To design Decision Boundaries is to write these three lines, workflow by workflow, as concrete conditions. And the record that these boundaries produce matters as much as the boundaries themselves. Decision Logs do not merely record outputs; they preserve accountability continuity across distributed judgment processes. When a judgment moves across the authority line, gets overridden, or is escalated, the log is what carries Accountability with it—so that the answer to "whose judgment was this?" survives the handoffs.
A Practical Enterprise Example: Grant Review
Leave the abstraction and draw the lines all the way through, in one setting: grant review.
From the moment an application arrives to the moment an award is decided, judgment is not a single act. It splits into stages. At each stage, decide whether an AI Agent or a human stands there.
Stage 1 — Formal check (AI Agent). An AI Agent checks for missing fields, absent attachments, and late submissions. This sits inside the AI Agent's Authority. Formal defects can be returned automatically. No human is involved.
Stage 2 — First-pass eligibility sort (AI Agent). An AI Agent sorts, at first pass, whether an applicant meets the program's eligibility requirements. It is not allowed to render a verdict. It outputs only three labels—clearly eligible, clearly ineligible, gray—with its supporting reasons. Clear cases it sorts; gray cases it escalates to the next stage.
Stage 3 — Subsidy-eligibility determination (this is the Decision Boundary). Gray cases, unprecedented project types, and high-value applications must not be determined by an AI Agent. Here the Authority to judge passes from the AI Agent to a human. The AI Agent hands over the issues and the evidence, but it does not decide whether the application qualifies. This single line—"from here, a human"—is the Decision Boundary. The flow runs from AI Agent, to human escalation, to final judgment, and this is the point at which it turns.
Stage 4 — Award or denial (human). The final call is made by a human alone. Denials in particular are made by a human. The state to preserve is one where a person can say "I turned this down," not "the AI rejected it." This is where Accountability binds: the human who made the determination takes on its result, and the Decision Logs record that it was theirs.
Put these four stages on a single page and the location of the boundary is obvious at a glance. Formal check, AI Agent. First-pass eligibility, AI Agent. But from subsidy eligibility onward, human. An AI Agent is not taking work away, and a human is not drowning in every case. The institution has simply fixed the one point at which the Authority to judge moves.
The same method transfers directly to hiring, procurement, and contract review. The subject changes; the lines you write do not. Authority, judgment ownership, delegation, escalation, override, and accountability continuity—write these six as concrete conditions in the language of the work. That is what it means to put Decision Design into practice.
Conclusion
Naming an owner is the starting point of AI Governance, not the finish line.
Decision Design is not a philosophy for raising the quality of AI's decisions. It is a Judgment Architecture for designing judgment itself in an age when AI Agents and people share the work. Who decides. How much is delegated, and where a human takes the judgment back. That line is not left to drift, unexamined.
Before asking whether you can name the person accountable, ask whether you have drawn the boundary of the judgment that person takes on. The two questions from CX Today and from Japan's AI Business Guidelines Version 1.2 both lead here.
Naming a responsible owner for AI does not make the problem disappear. The problem disappears not when an owner is chosen, but when the Decision Boundary is designed. The challenge is not assigning responsibility after a decision. The challenge is designing legitimate judgment before a decision.
Sources
- Thomas Walker, "AI Accountability: The Governance Gap Leaders Miss," CX Today, June 29, 2026. https://www.cxtoday.com/ai-automation-in-cx/ai-accountability-the-governance-gap-leaders-miss/
- Ministry of Internal Affairs and Communications and Ministry of Economy, Trade and Industry, "AI Business Guidelines (Version 1.2)," March 31, 2026. https://www.meti.go.jp/shingikai/mono_info_service/ai_shakai_jisso/20260331_report.html
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.