A week of work is full of tasks that never make it onto a plan: chasing a piece of information, hunting for the right document, noticing that a decision never got a reply. Before delegating to AI agents, look at these recurring responsibilities and describe each one as a whole, not as a string of separate requests.
When every step becomes its own request, someone still has to remember the sequence and connect the results. Delegation only happens when the work is described as one piece.
Find what depends on your memory
Start with a simple question: what keeps happening only because I remember to do it? Pick a responsibility that has come back more than once. Write down what triggered it, what information you went looking for and how you knew it was resolved. Include the detours: the ambiguous message, the incomplete answer, the commitment that changed.
Tracking the documents different teams owe you before a meeting is a good example. The work includes checking what arrived, spotting what is missing and telling a delay apart from a change in scope. An automatic list of files covers only part of that.
Also ask whether the task should exist at all. A follow-up can disappear once the deadline or the owner is clear. Fixing the process may work better than an agent propping up the same confusion every week.
Set a clear start and finish
For internal requests, the responsibility might start with the requests from a set period and end with a reviewed view of what still needs a reply. Along the way it can include checking history, flagging open items and drafting responses.
The instructions need to say which sources to check, how to recognize an open request and what counts as a sufficient answer. A thank-you can close a conversation; a "got it" can leave the work open.
Spell out the limits. Drafting a reply, sending it and making a commitment are three different actions. The result should show what was done and what is waiting for a decision.
Choose by the result you want
Maybe the goal is to walk into the meeting with complete information. Maybe it is to stop checking the same things again and again. Each goal calls for different criteria and helps you judge whether the delegation is worth keeping.
A useful test includes routine requests, incomplete information and one exception you already know. Watch whether the agent leaves out something important, invents a priority or marks as done something it only drafted. Good writing matters, but it does not make up for a forgotten request.
If the task has to run on a schedule, check how the schedule is actually set up. A description in plain language does not, by itself, make anything run again.
The leader's job
Delegating well means making the work recognizable. The leader defines the outcome, makes the criteria explicit and watches the exceptions. In AI Thinking by Dani Hayoshi, this is the starting question: who does each step, and how is it checked?
For a concrete case, see how to check what AI email triage left out.
Dani Hayoshi is the author of AI Thinking. Her work connects artificial intelligence, education and entrepreneurship. About Dani.