Using AI is often described as automation or assistance. Both terms can hide the more important organisational decision: what has been delegated, under what conditions, and with what continuing responsibilities?
Delegation can be small. A system may draft text while a person retains every consequential decision. It can also extend across a process, allowing the system to choose tools, sequence actions, contact other systems or continue until a goal appears to be complete. Treating these arrangements as versions of the same “AI use” makes governance unnecessarily vague.
Specify the object of delegation
The first question is what is being handed over. It may be:
- producing or transforming information;
- evaluating options;
- recommending a decision;
- making a bounded decision;
- taking an action;
- deciding what to do next;
- coordinating other tools or agents.
A workflow can contain several of these at once. Naming them separately helps reveal where judgement has moved and where a person still needs to intervene.
Delegation includes a mandate
A useful delegation arrangement gives more than a goal. It specifies permissions, constraints and stopping conditions. It establishes what the system may access, which actions require approval, how far it may continue, and what should cause it to pause or escalate.
These choices are often embedded in configuration rather than written as organisational decisions. A permission setting, trigger or retry rule can determine how much practical discretion the system has. Technical configuration is therefore part of the mandate.
Checking is part of the design
Delegating a task does not remove the need to evaluate its result. It changes the evaluation problem. The person checking an output may need different information from the person who originally performed the task. They may also have less contextual knowledge because the system completed intermediate steps out of view.
Checking arrangements should match the task. Some outputs can be reviewed individually. Others require sampling, monitoring patterns over time, comparing outcomes across groups, or examining unusual cases. A generic instruction to “verify the output” transfers responsibility without specifying a feasible method.
Responsibility does not follow the task automatically
An organisation may delegate performance while retaining legal, professional or moral responsibility. That is not inherently contradictory, but it becomes unstable when the responsible person lacks control over the system’s mandate or cannot reconstruct what happened.
Good delegation keeps responsibility connected to governing capacity. The people accountable for an outcome need access to relevant information and influence over the conditions that produce it.
Delegation changes over time
Tasks are often handed to AI gradually. A system starts as an optional aid, becomes the default route, and eventually shapes the workflow around itself. Each individual change may look minor even when the cumulative shift is substantial.
This is why delegation should be reviewed as an evolving arrangement. The central question is not only what the system was originally authorised to do. It is what people and systems now do in practice, and whether the mandate, controls and responsibilities still match that reality.