A familiar idea
Delegation is not new. A partner asks a manager to handle a client’s correspondence. A director gives an assistant access to their diary but not their bank account. A power of attorney can be broad or limited to one matter. In each case, someone with authority passes on part of it, for a purpose.
AI delegation applies the same idea to AI agents. When you ask an agent to do work for you, you are delegating. The question is whether that delegation is explicit and bounded, or implicit and unlimited.
The implicit version
In many set-ups today, delegation is all or nothing. You sign an AI tool in with your account, and it can do whatever your account can do, indefinitely. A partner’s AI inherits the partner’s access to every client, every mailbox folder and every approval right.
That is rarely what anyone intended.
What an explicit delegation says
A proper delegation records:
- Who delegated. The person granting authority.
- To which agent. The specific AI client or workflow.
- What scope. Which clients, projects or systems.
- Which actions. For example, read and draft but not send or approve.
- For how long. Until the task is done, a date, or until revoked.
- With what conditions. For example, approval required before external communication.
The delegation can never exceed the delegator’s own authority. A manager cannot delegate a partner’s approval rights they do not have.
Delegation chains
Agents increasingly call other agents or tools. A person delegates to an assistant, which asks a specialist agent to run part of the job. Each step should narrow, never widen, what is allowed. The final action should still be traceable back to the person who started it.
Why delegation matters for audit
With explicit delegation, an audit record can say: the agent created this draft, acting for Sarah, under her delegation for year-end reminders, on ABC Limited. That is far more useful than “Sarah did it” when she did not, or “the AI did it” with no one accountable.
Delegation in regulated work
In professional services, the person delegating remains responsible for the work. Delegating to AI does not transfer professional responsibility. Explicit delegation helps by making clear what the AI was allowed to do, so a person can review it properly.
A checklist
- Can each AI agent be given less authority than the person using it?
- Is delegation limited in scope and time?
- Can a person see and withdraw what they have delegated?
- Does the audit trail record the delegation used?
- Do delegation chains only ever narrow authority?
How Kroy approaches it
Delegation is one of Kroy’s core primitives: who authorised whom? On every request, Kroy checks what the person has delegated to the agent, alongside the person’s own permissions and the organisation’s policies. A delegation can be narrowed to specific entities, Skills and actions, can expire, and can be withdrawn at any time.