There are only two reasons something can't be done by AI.

The question comes up in almost every AI Framing Sprint, in some form. What can't be done by AI?

There is a short answer. It nearly always comes down to one of two things: judgement, or oversight.

They get confused constantly, and the confusion is exactly where teams get stuck.

Judgement is not always the obviously hard case. It is rarely the dramatic ethical dilemma people picture first. More often it is quieter than that: a borderline decision, a call that depends on context nobody wrote down, the kind of thing a good employee makes ten times a day without noticing they are making it.

Oversight is a different question entirely. It is not asking whether the work can be done well without a human. It is asking who still has to answer for it if something goes wrong, and no amount of capability changes that answer.

Naming which one applies to a given piece of work, judgement, oversight, both, or neither, is usually the moment the guardrails conversation in a Sprint stops being abstract and starts being usable.

None of this means AI cannot touch judgement or oversight work at all. That is usually the wrong question to be asking in the first place.

The right question is which autonomy level fits: Automate, Augment, or Agentic. Augment keeps a human making the judgement call, with AI doing the preparation around it. Automate runs a defined process without asking AI to exercise judgement at all. Agentic goes further, and is rarely appropriate where judgement or oversight sit close to the work.

Every task inside a business carries some risk and some opportunity, usually in roughly equal measure. Getting a small part of it wrong sometimes still leaves less value on the table than leaving it alone entirely.