Delegate something other than a copy of yourself

Delegation is not handing unpleasant work to a person nobody remembered to give the context to.

For someone else to carry out a task, he needs:

• the expected result;
• quality criteria;
• a deadline;
• authority;
• resources;
• constraints;
• check points;
• a way to ask a question.

A common error is to describe the action but not the goal. The performer follows the instruction conscientiously and produces a result nobody needs.

Another error is to hand over the goal with no constraints and expect the person to guess a dozen conditions that exist only in the head of whoever set the task.

A good assignment answers the questions:

What has to change?

What must not be broken?

How will we know the work is accepted?

In what cases should you stop and come back to me?

A check halfway through the deadline is more useful than a check on the last day. But the number of check points depends on the performer’s experience and the price of an error. Constant supervision strips delegation of its point and of responsibility.

You should not demand contact at the slightest doubt. Then every question comes back to whoever set the task. Better to define a threshold:

• what the person decides himself;
• what he reports;
• what requires approval.

Assigning one task to several performers in parallel sometimes shortens the time or produces alternatives. Sometimes it simply burns resources and creates competing versions. It is justified when the work really is divisible or when the value of an independent solution exceeds the cost of duplication.

A delegated result will almost inevitably differ from the picture in the head of the person who set the task. That is not a defect if it meets the requirements.

The main sign of delegation is that the other person received not only the work but part of the right to decide how to do it.

AI is a special kind of performer. It is fast, patient, able to offer many options, and it has neither authority nor reputation nor any responsibility of its own for the consequences.

It particularly needs context, quality criteria, a policy on sources, an acceptable level of uncertainty and an explicit ban on filling missing data with guesses. Otherwise the system will conscientiously amplify an incomplete brief.

A model’s answer also reflects the frame set by the person asking. If the task is stated too narrowly, the result can be excellent inside a mistaken boundary. So it helps to ask separately for the missed premises and the strongest alternatives.

The main sign of this kind of delegation is that the person decided in advance where the machine draft ends and human sign-off begins.

A classic mistake by whoever sets the task is to give the route instead of the address. “Call these twenty customers” instead of “find out why this segment stopped renewing.” The performer dutifully calls, reports twenty calls, and both sides then discover that the knowledge they needed lay in five conversations and one export from billing. A goal allows for a method better than the one known to the person setting the task; an instruction leaves no room for that.

Hence a practical test: if the performer found not a single deviation from your plan over the whole job, what was delegated was not the task and the responsibility but your hands.