Beyond Coding, What Work Are You Delegating to AI Agents?

I’m curious about how people use AI agents outside software development—particularly what separates the tasks that work well from those that don’t.

One possible distinction is whether someone else can trace the result back to the original requirements. I’ll call this traceability.

In coding, we can show how an implementation addresses a requirement and what the tests check. Passing tests doesn’t mean every requirement has been satisfied, but we have ways to examine the result together.

Other tasks offer similar reference points. Information extracted from a contract can be checked against the original text. Reconciled data can be checked against source records. A calculation can be checked against its inputs and formulas. These connections make it easier to explain why an AI-generated result is usable.

We can identify the sources supporting its statements. Yet decisions such as “Why lead with this argument?”, “How strongly should we make this claim?”, and “What should we leave out?” may remain in the requester’s head.

When delegating the work, we may not be able to fully explain what a good result must satisfy. But once we see the output, we can tell that something feels off.

Sometimes the output helps us discover what we cared about in the first place. The facts may be correct, but if the emphasis or priorities differ from what we intended, the presentation still isn’t ready to use.

This can happen when an individual delegates to AI, and when an organization delegates to an employee. The organization’s priorities and expectations may not have been fully communicated when the work was assigned.

The employee uses AI based on their understanding and produces something they’re satisfied with. But when they present it, the response is, “This isn’t the direction we wanted.” Only then do previously unstated criteria become apparent.

In such cases, the gap between the employee’s assessment and the organization’s assessment may partly reflect criteria that were never shared at the outset.

Even work we perform repeatedly isn’t necessarily governed by a fixed procedure. It may contain judgments that have become second nature to an individual or an organization. Delegating the task without conveying those judgments may not produce the expected result.

There’s another complication: other people usually see the output only after its creator has revised it to their satisfaction. Evaluating the finished product doesn’t tell us whether the AI’s initial output was useful or whether substantial human effort made it usable.

Perhaps successful AI delegation tends to occur where the connections between requirements, evaluation criteria, and results are easier to trace.

Even within the same task, some parts can be checked against explicit evidence, while others depend on judgments that haven’t been articulated. That distinction might help us understand which parts are easier to delegate.

This is a hypothesis, not an established conclusion. I’m especially interested in counterexamples: tasks that are difficult to trace in this way but have nevertheless been delegated successfully.

If you use AI agents outside coding, I’d like to hear:

I’d welcome both successful examples and cases that didn’t meet expectations, to help test this hypothesis.