What not to delegate to AI: draw your personal line

AI can draft, summarize, classify, compare, and transform information quickly. That makes it tempting to ask one question about every task:

Can AI do this?

The more useful question is:

Which part of this task should remain mine?

A model may be capable of producing an output without being the right party to make the decision. It cannot carry professional accountability, repair a damaged relationship, understand every unstated constraint, or notice that you are gradually losing a skill you still need.

Good delegation therefore starts with a boundary, not a feature list.

Assistance is not ownership

Most knowledge work contains several different activities:

1. collecting information;
2. organizing or transforming it;
3. generating options;
4. choosing an action;
5. communicating the choice;
6. and owning the result.

AI is often useful in the first three. The last three require more care.

For example, a model can:

• summarize interview notes;
• identify repeated themes;
• suggest several staffing options;
• or draft a message explaining a decision.

That does not mean it should choose whom to hire, decide whom to dismiss, or send a sensitive message without review.

The output and the responsibility do not travel together. The responsible person or organization remains responsible even when a model produced the recommendation.

Four boundaries to examine

Use four tests before delegating a meaningful task.

1. Consequence

Ask what happens if the output is wrong, incomplete, biased, or misunderstood.

Low-consequence tasks are easier to delegate:

• reformatting notes;
• proposing agenda headings;
• generating practice examples;
• or creating a first-pass checklist.

Higher-consequence tasks need stronger human control:

• deciding access to money, work, housing, education, or essential services;
• making a safety-critical instruction;
• interpreting a contract for action;
• or responding to a serious personnel issue.

Consequence is contextual. A typo in a private brainstorm is minor. The same typo in a dosage, bank account, deadline, or customer commitment may not be.

Do not label an entire activity “safe” or “unsafe.” Identify the particular output and what it can change.

2. Accountability

Ask who can explain and defend the decision afterwards.

If the honest answer is “the model decided,” the task has been delegated too far.

For consequential work, name a human owner before using AI. That owner must be able to:

• understand the relevant evidence;
• detect a weak or unsupported recommendation;
• reject the output;
• explain the final choice;
• and correct the result if it causes harm.

A review step is not meaningful when the reviewer lacks time, authority, context, or skill. Clicking “approve” is not oversight.

3. Relationship

Some work is valuable because another person can see that you paid attention.

Consider:

• giving difficult feedback;
• apologizing;
• recognizing excellent work;
• supporting someone through uncertainty;
• negotiating a disagreement;
• or explaining a decision that affects a colleague.

AI can help you prepare. It can organize the facts, identify an unclear sentence, or suggest questions you may have missed.

But a polished generated message can feel evasive when the real need is presence, judgement, and a genuine response. If trust depends on the recipient believing that you listened and chose the words, do not outsource the entire interaction.

4. Skill retention

Delegation changes the person delegating.

If AI always performs a skill, you get fewer chances to practice it. That may be acceptable for a task you deliberately want to stop doing. It is risky for a skill you need in order to:

• supervise the model;
• respond when the tool is unavailable;
• notice an unusual case;
• teach someone else;
• or advance in your role.

The verification paradox is simple: the less you practise a skill, the less able you may become to check generated work in that domain.

Choose which abilities you want to retain. For those, use AI after a first attempt, for feedback, or for comparison—not as the automatic first mover.

A five-question delegation test

Before handing over a task, answer:

1. What is the worst plausible result of a bad output?
2. Who is accountable for the final action?
3. Can that person independently verify the important parts?
4. Would full delegation weaken trust with someone affected?
5. Is this a skill the team still needs to maintain?

Then choose one of four modes.

This produces a more useful policy than “employees may use AI” or “AI must not make decisions.” It specifies where the boundary sits inside the task.

Examples of a sensible split

Customer complaint

AI may:

• group the issues;
• find relevant policy passages;
• or draft three response structures.

A person should:

• determine what actually happened;
• choose any remedy;
• adapt the tone to the relationship;
• and approve the response.

Hiring

AI may:

• turn an approved role description into a structured interview plan;
• remove duplicate questions;
• or format interviewer notes.

A person should:

• decide what evidence matters;
• investigate inconsistent information;
• make and explain the decision;
• and check whether the process treats candidates fairly.

Learning a technical skill

AI may:

• create practice questions;
• review an attempted solution;
• or explain an error in another way.

The learner should:

• attempt retrieval or problem-solving before seeing the answer;
• reproduce the solution unaided;
• and test the skill on a new case.

Important business recommendation

AI may:

• structure the available evidence;
• expose assumptions;
• generate alternatives;
• or take a deliberate opposing view.

The decision owner should:

• confirm the source data;
• investigate material uncertainty;
• compare trade-offs;
• and record why the final choice was made.

Write a personal boundary worksheet

Create a table with one row per recurring use:

Do not complete the table from memory alone. Ask the people who perform, review, and receive the work. They may see consequences or relationship costs that the process owner misses.

Review the worksheet after:

• a significant error;
• a change in the tool or workflow;
• a new data source;
• a change in who verifies the result;
• or evidence that people are losing a required skill.

Use AI to examine the boundary

AI can help stress-test the policy without choosing it:

Act as a process-risk reviewer.

For the task below, separate:
- information gathering,
- transformation,
- option generation,
- decision,
- communication,
- and accountability.

For each stage, identify:
- a plausible failure,
- who could detect it,
- what evidence they would need,
- whether the action is reversible,
- and which human skill could weaken through repeated delegation.

Do not decide the acceptable risk for me.
Ask for missing context before making assumptions.

[describe the task]

Treat the response as a checklist for discussion, not as approval.

The practical line

Keep a task closer to the human when:

• the consequence of error is serious;
• the decision affects another person’s opportunities or rights;
• the reviewer cannot independently verify the output;
• trust depends on authentic attention;
• or the task exercises a skill needed for future judgement.

Use AI more freely when:

• the task is reversible;
• correctness is easy to verify;
• the model is preparing rather than deciding;
• no sensitive context is exposed;
• and a named person still owns the result.

The goal is not to preserve every manual step. It is to keep responsibility, relationships, and essential competence attached to people while using automation where it genuinely helps.

For a fuller six-dimension life-and-work audit with keep / assist / delegate / never examples across relationships and high-stakes judgment, see the delegation audit: what should stay yours.