In the Age of AI, What Work Must Not Be Delegated?

Geoffrey Chen

After AI Enters the Workflow · Season One: “From Answering Questions to Participating in Work” · Article 12

An organisation discovers that AI can prepare reports, compare policies, rank options, draft correspondence, write code and coordinate a sequence of actions. Each time a task is performed successfully, the next question seems obvious: why not delegate one more step?

Eventually the system is not merely preparing a recommendation. It is defining the objective, deciding which risks are acceptable, choosing whose interests matter, approving an exception and communicating a commitment in the organisation’s name. A person remains somewhere in the process, but mostly to confirm what the system has already framed.

Where should delegation stop?

The answer cannot be a permanent list of skills that machines will never acquire. History makes such lists fragile. Nor is the answer that AI must never be used in morally or legally important work. It can improve search, reveal inconsistency, model consequences and make relevant information easier to inspect.

The limit concerns a different question: which parts of work derive their legitimacy from a person or institution being able to understand the situation, exercise authority and accept the consequences? Those functions can be informed by AI, but they cannot be transferred to a system that has no standing to make the commitment.

“Must not be delegated” does not mean “AI must not be used”

Delegation is not all or nothing. A doctor may use AI to organise a record without delegating informed consent. A judge may use software to search precedents without delegating the judicial decision. A manager may ask AI to compare strategies while retaining authority over the purpose and accepted risk.

The relevant distinction is among assistance, recommendation and final disposition. AI can perform substantial analytical work while a qualified person retains the step that turns analysis into an institutional act.

Human retention must also be real. If a reviewer lacks source access, time, competence or authority to disagree, the workflow has delegated the decision in substance while preserving a ceremonial signature.

Article 14 of the EU AI Act makes this distinction operational for high-risk AI systems by requiring oversight measures that enable people to understand limitations, avoid over-reliance, interpret outputs and disregard, override or interrupt the system. EU AI Act, Article 14

The question is therefore not whether a human appears in the diagram. It is whether human judgement can still alter the decision at the point where legitimacy and consequence arise.

First: setting the objective

AI can optimise only after an objective has been represented. Choosing that objective is not a neutral technical preface.

A hospital can ask a system to minimise waiting time, reduce mortality, treat the most urgent cases first or distribute access more equally. A business can optimise revenue, retention, customer welfare, employee workload or long-term resilience. Different objectives reorganise benefits and burdens.

AI can expose trade-offs and test scenarios. It cannot decide which institutional purpose ought to govern unless people have already supplied a normative framework. If the system appears to choose, it is usually extending patterns from data, instructions and design decisions whose authority still comes from human institutions.

Objective-setting must therefore remain with a role entitled to define the mission and accountable to those affected. The chosen objective should be documented, contestable and reviewed when circumstances change.

Second: deciding who may bear which risk

Risk analysis can estimate likelihood and impact. It does not by itself determine whether the risk is acceptable or who should carry it.

An automated process may save the organisation money while imposing rare but severe costs on a small group. A model may have high average accuracy while performing poorly for people with unusual circumstances. A security control may reduce fraud while denying access to legitimate users.

The decision to accept these trade-offs distributes vulnerability. It requires authority, ethical and legal judgement, consultation and, in many cases, consent. The people who benefit from the automation should not quietly transfer its residual risk to those least able to challenge it.

NIST’s AI Risk Management Framework treats risk management as a socio-technical and organisational responsibility involving governance, context, measurement and management. NIST, “Artificial Intelligence Risk Management Framework 1.0” The framework can inform decisions; it does not make them legitimate on behalf of the institution.

Third: making final dispositions of rights

AI may help classify applications, identify missing documents, calculate eligibility under clear rules and surface relevant considerations. When a decision affects liberty, essential services, employment, welfare or other significant rights, final disposition requires an accountable legal or institutional actor and an effective route to challenge.

The reason is not that human judgement is infallible. Humans make serious errors. The reason is that a rights-affecting decision must belong to a system of reasons, authority and remedy. The affected person must be able to know the basis, submit relevant evidence, contest error and obtain correction.

An AI output can be part of the evidence or analysis. It cannot be the terminal point at which the institution says, in effect, “the model produced this result, so no further reason is available”.

OECD AI Principles include human-centred values, fairness and accountability, placing AI use within obligations to respect rights and ensure responsibility. OECD.AI, “Human-centred values and fairness” and “Accountability”

Fourth: giving consent, forgiveness and exceptions

Some decisions cannot be reduced to consistent rule application because their meaning lies in who grants them.

Consent to medical treatment, disclosure of private information or use of a person’s likeness must come from an authorised person with adequate understanding and freedom. AI may explain options and check comprehension, but it cannot provide the person’s consent.

Forgiveness and reconciliation are relational acts. An AI can draft an apology or suggest language, but it cannot determine that another person’s wrong has been forgiven.

Exceptions are also revealing. An organisation may depart from a rule because an unusual case exposes a conflict between the rule’s wording and its purpose. AI can identify similar precedents and consequences. The authorised decision-maker must accept the fairness, consistency and future implications of the exception.

Delegating these acts would confuse informational assistance with the standing to alter a relationship or obligation.

Fifth: making public commitments

Organisations make commitments through contracts, official statements, professional certifications, published research and public policy. AI can prepare these materials, but the commitment must be adopted by a person or body with authority.

This is why major publishing policies do not treat AI systems as authors. The International Committee of Medical Journal Editors explains that AI tools cannot be authors because they cannot be responsible for accuracy, integrity and originality; human authors must remain accountable and disclose relevant use. ICMJE, “Use of AI by Authors”

The same structure applies outside publishing. A generated contract clause becomes the company’s commitment only when authorised representatives adopt it. A generated government statement becomes public policy only through legitimate institutional action. Fluency can prepare the language; it cannot supply the office, mandate or accountability.

Sixth: deciding when AI itself must stop

A system should not be the sole judge of whether it may continue operating. The organisation must retain the ability to suspend use when evidence is insufficient, harms emerge, the context changes or affected people cannot obtain remedy.

This includes operational stop controls and governance decisions. Technical teams need the ability to interrupt an agent. Risk owners need thresholds for pausing deployment. Senior officials need authority to withdraw a use case even if performance metrics remain commercially attractive.

The Australian Government’s responsible AI policy establishes governance expectations for government use, including accountable officials, transparency and management of AI use. Australian Government, “Policy for the responsible use of AI in government—Version 2.0”

A stop decision may be difficult because organisations have invested in the system and redesigned work around it. That makes the retained authority more important. Dependence should not eliminate the capacity to reconsider delegation.

What work is especially suitable for delegation

Drawing boundaries does not require minimising AI. Many tasks are strong candidates for extensive delegation when they are well specified, low risk, observable and reversible.

Examples include format conversion, duplicate detection, preliminary classification, search across approved sources, comparison against explicit criteria, test execution, anomaly flagging, draft preparation and routine scheduling within narrow limits.

Even here, the surrounding system matters. Data access should be proportionate, outputs should preserve sources, actions should be logged and unusual cases should escalate. Delegation works best when completion can be demonstrated rather than inferred from fluent language.

The objective is not to reserve every meaningful action for humans. It is to use machines most deeply where evidence and control are strongest, thereby preserving human attention for decisions that require authority, relationship and responsibility.

“Humans are ultimately responsible” must become a concrete structure

The phrase is correct but inadequate. Which human? Responsible for what? With which information and power?

A credible design should identify at least:

  • who defines and periodically reviews the purpose;
  • who approves data, models, tools and permissions;
  • who sets validation and deployment criteria;
  • who adopts rights-affecting or public decisions;
  • who monitors outcomes and receives complaints;
  • who can suspend the system;
  • who can explain, reverse and remedy an error.

The retained human functions must be supported with time, evidence, expertise and authority. Otherwise the organisation delegates control while retaining human blame.

The human remainder is not a mysterious skill machines can never learn

It is tempting to locate an eternal human advantage in creativity, empathy or common sense. AI capabilities may continue to change, and humans do not exercise those qualities perfectly.

The more durable boundary is institutional and normative. A person or organisation can occupy an office, receive a mandate, owe a duty, give consent, make a promise and be required to repair harm. These are not merely information-processing capabilities. They are relationships of standing and obligation.

AI may become better at modelling the considerations relevant to these acts. Better advice can improve human judgement. But the source of legitimacy remains the authorised actor who can be questioned and held to the consequence.

Conclusion: legitimacy and consequence cannot be delegated away

AI can perform growing portions of the analytical, linguistic and operational work surrounding a decision. The boundary should not be drawn around a fixed set of tasks or a romantic image of human uniqueness.

It should be drawn around the point at which an output becomes an authorised commitment affecting others.

The principle is:

We may delegate the production of options, evidence and implementation. We must not delegate away the human or institutional act that makes the objective legitimate, accepts the risk, disposes of rights or undertakes responsibility for the consequences.

This does not slow AI adoption by definition. Clear retained responsibilities make deeper delegation possible elsewhere. When purpose, authority, review and remedy are explicit, AI can act confidently within a bounded domain.

The work that must not be delegated is therefore not whatever AI currently performs poorly. It is the work through which people and institutions say: this is what we are trying to do, this is the risk we accept, this is the decision we authorise, and these are the consequences for which we remain answerable.

Primary sources and further reading

Continue reading: Explore the After AI Enters the Workflow series.

Discover more from Geoffrey Chen

Subscribe to get the latest posts sent to your email.