The Delegation Doctrine

Imagine taking over an organization whose AI systems are already working. Reports arrive on time. Customers receive answers. Routine decisions move through the business without waiting for a manager. Your predecessor approved the systems, the supplier maintains them, and the employees know how to use them. Nothing in the morning report suggests a crisis.

Then someone asks why a particular decision was made. The employee can retrieve the result but cannot change it. The supplier can explain the product but cannot authorize an exception to your policy. Your predecessor has left. You now own a service whose operations are familiar to everyone and whose authority is clear to no one.

This is a hypothetical situation, but it identifies a duty that survives any particular model or manager. An institution must retain the ability to answer for the power it exercises. Successful execution can make that duty easy to overlook. Work continues, so the organization assumes that the arrangement remains sound. The missing authority becomes visible when a person needs something other than the usual result.

The Delegation Doctrine sets out what must survive as more work moves to machines. It allows substantial autonomy within an accountable arrangement. It also identifies decisions for which human judgment remains necessary, however capable the technology becomes. Its purpose is to make those commitments durable enough to govern choices that a product demonstration cannot settle.

Chapter Twenty-Five supplied the practical test. This chapter gives the eight principles their full statement and explains their relationship. The test asks whether a particular delegation can be defended. The Doctrine explains what that defense must preserve.

The Meaning of Ownership

Delegation assigns work while retaining ownership of the goal, the boundary, the standard of success, and the consequences. That work can include execution, analysis, recommendation, and limited discretion. Abdication occurs when the arrangement leaves no identifiable party meaningfully answerable, or leaves someone answerable in name without the authority, capacity, information, or intention to answer.

Ownership here means authority over a decision. It does not mean owning the software company or knowing how to build its model. A public office still answers for decisions made with privately supplied software. A company still answers for the automated service it chooses to provide. Continuing a system inherited from an earlier manager is also a choice that needs an accountable owner.

Accountability means there is a working route to someone who must explain a decision and respond when it needs correction. Responsibility includes the moral duties people owe one another. Blame is a judgment about who was at fault. These are different questions. An organization can investigate and repair a mistake before deciding whether anyone deserves punishment. A duty to put things right can remain even when no personal misconduct is found.

This distinction also protects useful delegation. If accountability meant personally performing every task or guaranteeing that nothing could go wrong, it would be impossible to organize complex work. The obligation is to retain enough authority and competence to govern the arrangement, provide a meaningful response when challenged, and accept review of the choices that made the outcome possible.

The Eight Principles

Principle 1. Accountability does not transfer. Work, information, and discretion can move. Accountability can be shared or distributed among people and institutions, but it cannot disappear. A system is not an accountable party merely because it produced the output. The people or institutions that authorized, deployed, governed, or relied on the delegate retain answerability according to their role. “The system decided” describes a mechanism; it does not close the accountability chain.

An organization can distribute work across many hands while keeping those roles intelligible. The deploying institution answers for the decision to use a system in a particular setting. A supplier answers for choices within its control and the commitments it made. A professional answers for judgments within the professional role. Assigning an operational role to someone else can change the division of duties. It cannot make the earlier authorization unowned.

In the opening example, the new director need not pretend to have made the predecessor’s decisions. The director does need to establish what the institution is continuing to authorize. The office survives the departure of its occupant. Records and a credible handover allow personal roles to change without leaving the people affected by the service stranded between them.

Principle 2. Capability is not permission. A delegate’s competence is a reason to give it more work. It is never a reason to keep less judgment. As the machine performs more of the execution, the remaining human decisions - goals, limits, definitions of success, deployment, and stopping rules - become more consequential, not less.

A strong performance result answers a question about ability under particular conditions. Permission also requires a legitimate purpose, an authorized scope, and a judgment about the consequences for others. A system that predicts which customers will accept a price has demonstrated a capability. Whether the company should use that prediction, with which information and under which constraints, remains a decision for the company to defend.

The principle gives technical progress a proper place in that decision. Better performance can justify expanding an authorized task when the surrounding conditions support the expansion. It cannot silently supply the authorization. Otherwise, each new feature would revise the institution’s policy merely by becoming available.

Principle 3. Every delegation has a boundary. A delegation should specify what the delegate may do on its own, what it must escalate, and what it may never do. A delegation without a boundary is an abdication in advance because it gives away whatever the delegate later turns out to be capable of.

A useful boundary includes a destination for uncertainty. When a case falls outside the permission, the system should reach someone who can decide what happens next. That person needs the authority to resolve the exception rather than simply return it to the same process. An escalation queue without an effective destination is an unfinished boundary.

Boundaries also make competent autonomy possible. Within a well-defined task, constant approval can add delay without improving judgment. The owner should be able to explain why routine execution may proceed and why a different action requires review. The permission becomes more useful when its limits are specific enough to enforce and revise.

Principle 4. You did not delegate what you cannot stop. Accountability requires practical oversight: enough evidence, explanation, context, time, authority, and competence for the owner to judge what the delegate is doing and intervene when necessary. This does not require access to a model’s hidden internal reasoning. It requires enough usable information for genuine independent judgment and meaningful disagreement. If a person cannot intervene in practice, the system has not been supervised; it has been released.

Stopping must be understood in relation to the consequence. An owner may be able to pause a workflow, revoke access to a tool, prevent the next transaction, or route an unresolved case to a person. Some completed actions cannot be undone. Where intervention after execution would come too late, the control must operate before that action or at the earlier authorization that permits it, subject to the human-decision requirements of Principle 7.

Visibility serves that practical judgment. A reviewer may need the evidence used, the applicable rule, the action proposed, and the system’s relevant limitations. An impressive account of internal technical detail can still leave those questions unanswered. The standard is whether usable information reaches a competent owner in time to support a real decision.

Principle 5. Speed and architecture move accountability upstream. They do not erase it. Some automated outcomes happen faster than a human can review them individually, while others emerge from data, defaults, thresholds, code, workflows, and institutional design rather than a single final choice. Accountability therefore moves to the earlier human decisions that authorized the system to operate at that speed and through that architecture, within that scope and under those safeguards. Those design and deployment choices must be deliberate, recorded, and owned.

This principle connects the stop requirement to systems operating beyond the speed of individual review. A person may authorize a bounded automated response after examining its conditions and safeguards. The absence of a human click at each execution does not erase that earlier choice. The owner must be able to show what was authorized and how the permission can be withdrawn or narrowed when its conditions fail.

Upstream authorization is not a waiver of Reserved judgment. Chapter Thirteen’s distinction remains: bounded defensive autonomy against incoming munitions or other objects, and carefully bounded autonomous action against materiel targets, may be governable under Principle 3, upstream authorization, and a real intervention path. The intentional decision to apply lethal force against a person remains Reserved. Speed cannot be used to remove a decision from that category.

Principle 6. Consequences must reach the owner. Accountability requires exposure to meaningful review and consequences appropriate to the role. Those consequences may be legal, institutional, professional, reputational, electoral, financial, or employment-related. When benefits flow to one party while harms are externalized to people with no voice and the decision owner remains insulated from review, accountability is mostly ceremonial.

Consequences should reach the authority that could have changed the arrangement. An employee who follows a mandatory process may have little power over its design. A senior manager who removed review capacity may have considerable power despite never seeing the individual case. Investigation should establish those differences. Automatic blame for the last person in the workflow would reward the people who placed themselves farthest from the outcome.

Meaningful consequences can include correcting the decision, repairing harm, changing permissions, independent scrutiny, or sanctions where warranted. Their form should reflect the role and the findings. A system that punishes every reported mistake can teach people to conceal warning signs. Accountability requires enough candor to discover a failure as well as enough authority to respond to it.

Principle 7. Some decisions must carry a human signature. Human involvement is not human accountability. The accountable human must have authority over the decision itself. For irreversible, high-stakes decisions, and especially for Reserved decisions whose legitimacy depends on an accountable human decider, the signer must be an authorized, competent decision owner or designated professional with actual authority over the specific case, enough evidence and context to judge it, practical power to disagree or override, and exposure to consequences. A model trainer, data annotator, outside contractor, generic reviewer, or ceremonial signer who lacks authority over the decision does not satisfy this requirement. Reserved decisions require independent human judgment, not mere ratification. The reason is legitimacy and answerability, not an assumption that humans are always more accurate.

The signature represents an exercise of authority over the case. The institution must supply the conditions that make it valid. A reviewer with no time to examine the evidence, no power to change the answer, or no safe way to disagree cannot provide what this principle requires. Adding a signature field while removing those conditions leaves the underlying delegation unresolved.

Tier 3 and Reserved remain distinct even when both require a human decision. Tier 3 responds to irreversible, high-stakes consequences. Reserved concerns judgments whose legitimacy depends on an accountable human decider. Its established examples include a criminal sentence, the intentional decision to apply lethal force against a person, removal of a child, denial of liberty, and disclosure of a terminal diagnosis. Greater measured accuracy alone does not settle who may exercise that authority.

Principle 8. Accountability decays. It must be renewed. Delegation can quietly become abdication through habit, over-trust, deskilling, and the gradual loss of the capacity to intervene. Accountability is therefore not a one-time authorization. Boundaries must be re-read, outputs sampled, overrides exercised, and human competence deliberately preserved.

Renewal makes the earlier principles continuing obligations. The opening organization might have had a sound arrangement when it launched. Staff departures, new connections, or a growing workload could have changed what the original approval meant in practice. A leader taking over should examine the present system rather than treating an old signature as evidence of current control.

An absence of overrides may reflect good performance. It should prompt an examination of whether disagreement remains possible, not a quota for unnecessary interventions. Controlled exercises and the outage test can help establish whether people still detect errors and recover when the tool is withheld. Safe test conditions matter. Competence should be demonstrated without exposing people to preventable harm merely to prove that a safeguard exists.

Current Evidence: Principles and Their Implementation

Reviewed September 13, 2026.

International guidance already recognizes obligations that persist beyond a model’s output. UNESCO’s Recommendation on the Ethics of Artificial Intelligence, adopted in November 2021, places human rights and dignity at its foundation. Its treatment of human oversight retains ultimate human responsibility and accountability. Its accountability provisions call for traceability, auditability, and mechanisms for oversight and assessment. These are normative commitments. Adoption of the Recommendation does not establish that every institution has implemented them.

The OECD’s Due Diligence Guidance for Responsible AI, published February 19, 2026, explains how businesses that develop or use AI can check for harm and respond to it. Due diligence here means continuing work: build responsible conduct into company rules, identify actual and possible harm, act to stop, prevent, or reduce it, track the results, explain what was done, and help put matters right when appropriate. The examples can be adapted; they are not an exhaustive checklist. The underlying business standards are voluntary. The guidance offers practical steps, but its publication does not prove that businesses follow them or that harm is prevented.

I read these documents as support for treating governance as work that continues throughout use. Neither establishes this book’s precise Tiers or Reserved boundary, and neither validates the Accountability Test as a measured instrument. The Doctrine brings its own judgment to the question of which authority should remain human. That judgment must be defended directly, rather than borrowed from the standing of an international organization.

The evidence most useful for future revision would show whether people can challenge consequential outcomes, whether interventions work under ordinary workload conditions, and whether remedies reach those harmed. A comparison must also examine the available alternative. An AI-assisted process can have defects and still improve on a harmful manual one. The relevant question is whether the proposed arrangement improves outcomes while preserving the authority and safeguards the decision requires.

Accountability Analysis

The principles work together. Naming an owner achieves little unless that owner can see what is happening and intervene. A written boundary needs review as the work changes. Earlier authorization for fast automated action still carries consequences and cannot replace a required human decision. An institution cannot satisfy the Doctrine by choosing the easiest promise and leaving out the conditions that make it work.

Return to the director inheriting a functioning service. An immediate shutdown of every automated task could interrupt useful work and harm people who depend on it. Continuing everything without examination would also be a decision. The director should identify the consequential actions, establish their owners and permissions, and prioritize gaps according to the stakes, reversibility, and applicable prohibitions or Reserved requirements. The Chain, Tiers, and Renewal provide the practical method already developed in Chapter Twenty-Five.

Suppose the initial review finds that routine correspondence has a clear boundary, while disputed outcomes cannot reach anyone able to change them. The first finding may support continued delegation. The second requires a decision about the affected actions. The institution might restore a competent review route, narrow the automation, or suspend the unsupported authority. The remedy should address the actual failure, with continuity arrangements where interruption would create its own serious consequences.

The director must then examine why the gap existed. Perhaps the operating instructions were clear but employees lacked access to the relevant record. Perhaps the supplier offered an override that the institution never enabled. Perhaps management set a workload that made review impossible. Each explanation points to a different corrective action and a different accountable role. Treating all three as employee error would preserve the defect.

An institution should also allow its own judgment to be challenged. A capable owner can still defend an unfair rule or an inadequate remedy. Independent review, professional standards, and the voice of affected people limit that danger. Accountability cannot mean simply that someone powerful is willing to take credit or accept criticism. The person answering must face a process that can examine the grounds of the decision and require a meaningful response.

There are costs. Evidence must be retained appropriately, reviewers need time, and some competence must be maintained even when routine execution moves elsewhere. Smaller organizations may need shared expertise or a narrower permission because the necessary review capacity is expensive. That is a reason to make safeguards proportionate and usable. It is also a reason to count their cost before claiming that a deployment saves money.

The Doctrine can be misused as a defense of familiar jobs or existing hierarchies. A professional should not be able to invoke human judgment to protect poor performance from comparison. Nor should an organization classify everything as exceptional until meaningful delegation becomes impossible. Where a bounded task can be performed well and the required accountability survives, the principles support giving the machine the work.

The harder limit appears when the proposed efficiency depends on removing the very capacity that makes the delegation legitimate. A budget may show a saving from eliminating reviewers while continuing to describe their judgment as a safeguard. The institution must then change the budget, the scope, or the claim. An unavailable safeguard cannot remain part of the justification for proceeding.

The Doctrine is not a defense of human judgment against machine competence. It is a defense of the accountability loop that allows competence to be corrected, governed, and legitimately used.

Check an inherited promise

Return to the director who inherits an automated service. Choose one promised safeguard, such as a route to human review. Explain what would make it usable today: an authorized person, the necessary records, enough time, and power to act. If a condition is missing, identify what must change before the organization can rely on that safeguard.

Accountability Verdict

The director’s obligation is to make the inherited service answerable in the present. That means knowing which decisions the organization is authorizing, who can examine and change them, and what happens when someone challenges the result. The old approval becomes part of the history. Current authority requires current conditions capable of supporting it.

The eight principles do not promise error-free institutions. They establish the obligations that remain when institutions use powerful delegates, including the obligation to learn when an arrangement fails. Better evidence may justify broader autonomy, a narrower permission, or a revised judgment. The Doctrine itself should remain open to reasoned challenge. Its claims deserve the same scrutiny that the book asks of a deployment.

What we should preserve is an institution’s capacity to make a defensible choice and remain answerable for it after execution has moved elsewhere. That capacity can support ambitious use of AI. It can also require a refusal when the proposed authority exceeds what the institution can govern. The remaining question is personal: what does it mean for any of us to put our name to a decision when a machine has done much of the work?

Sources and Notes

1. UNESCO, Recommendation on the Ethics of Artificial Intelligence, adopted November 2021. Official Global AI Ethics and Governance Observatory presentation: https://www.unesco.org/ethics-ai/en/recommendation. Accessed September 13, 2026. See the responsibility and accountability, and human oversight and determination provisions. This is normative guidance, not an evaluation of deployment outcomes. Source record: DA-CH26-S1. Archive reference: pending.

2. Organisation for Economic Co-operation and Development, OECD Due Diligence Guidance for Responsible AI, OECD Publishing, Paris, February 19, 2026. https://doi.org/10.1787/41671712-en. Official report and executive summary: https://www.oecd.org/en/publications/oecd-due-diligence-guidance-for-responsible-ai_41671712-en.html. Accessed September 13, 2026. See the executive summary’s due-diligence measures and implementation-example qualifications. Source record: DA-CH26-S2. Archive reference: pending.