What Should AI Decide — and What Must Remain Yours?

The short answer

An AI system can own routine execution once a human has already made four things explicit: the goal, the acceptable tradeoffs, where responsibility sits, and the conditions under which the system must stop and escalate. Inside a boundary drawn that carefully, the system is not exercising judgment. It is applying judgment that was already made.

Human judgment remains necessary wherever the work requires deciding what matters, interpreting genuine ambiguity, choosing among consequential tradeoffs, accepting responsibility for the outcome, or determining when the work is actually done. That set does not shrink as models improve. It is not a capability gap waiting to close — it is where the meaning of the work is decided, and delegating it does not transfer it. It only makes ownership harder to locate.

The useful name for the posture is delegation without abdication: real autonomy inside a narrow, stated boundary, with responsibility still visibly held by a person.

Two answers that sound responsible and are not

"A human must approve every AI action." This sounds like control and usually produces the opposite. When a person signs off on every ordinary case without adding anything, one of two things is happening: they are approving at a volume that makes genuine review impossible, or they have become an undesigned routing step the process now depends on — ceremonial clickers and hidden operational middleware, to put names on the two failure shapes. Neither is oversight. Both are worse than an honest boundary, because both look like oversight from the outside.

"Once AI can execute a task, human judgment is no longer required." Execution capability says nothing about whether the goal was right, whether the tradeoffs were acceptable, or whether the output is good enough to stop. Those are separate questions, and they were never the thing being automated. A system can perform a task flawlessly and still be doing work that was not worth doing.

The two mistakes are usually made by the same organization, in different places, for the same reason: nobody drew the boundary, so oversight defaulted to everything or to nothing.

The boundary

The split is not between hard tasks and easy ones. It is between deciding and doing.

What human judgment holds:

• Purpose. What this is for, and why it is worth doing at all.
• Success criteria. What "good" means here, specifically enough that something could fail it.
• Values and tradeoffs. Which costs are acceptable, and which are not, when they conflict.
• Responsibility. Who answers for the outcome — a question no system can absorb on anyone's behalf.
• Ambiguity and exceptions. Cases the boundary did not anticipate, where a policy has to be made rather than applied.
• Stopping conditions. When the work is actually finished, and when to abandon it.

What a system can own, once that is settled:

• Repeatable bounded steps whose criteria are already explicit.
• Retrieval, organization, and transformation — collecting, sorting, reformatting, deduplicating.
• Monitoring — sustained attention over time, which is the thing humans are worst at and systems are best at.
• Ordinary routing of cases that clearly match a known pattern.
• Visible escalation when a case does not match one.

Work that requires creative leaps, taste judgments, or emotional intelligence sits on the human side of this line regardless of how capable the system is. So does any case where the system would have to invent policy in order to proceed. A system that needs to be told why before it can act has been handed the wrong kind of work; what a system needs is constraints. The sequence that keeps this honest is simple: the system surfaces, the human interprets, the system acts on what was decided. Reversing it is how organizations end up with outcomes nobody chose.

A test you can run on one workflow

Pick a single workflow you are considering handing over, or have already handed over, and answer these honestly. This is a thinking aid, not a certification.

1. Has a human defined what "good" means here? Not "accurate" or "high quality" — a definition specific enough that a particular output could be judged to have failed it.
2. Are the limits and acceptable tradeoffs clear? When speed and care conflict in this workflow, is there a stated answer, or does each case get resolved by whoever happens to be looking?
3. Can the system recognize an ordinary case without inventing policy? If handling a case requires deciding something the boundary never addressed, that case is not routine, whatever it looks like.
4. Is there a visible path to escalate ambiguity or consequential failure? Visible matters. An escalation route nobody watches is a route that does not exist.
5. Would a human touch here add judgment — or merely delay? This is the question that separates real oversight from ceremony, and it is worth asking about each approval step you already have.

A "no" is not a verdict against automation. It usually identifies which piece of human judgment has not been made explicit yet — which is a solvable problem, and a cheaper one to solve before delegation than after.

Three ordinary examples

Illustrative only. These are generic shapes, not descriptions of any particular system.

• Refunds. A human decides the refund policy — what qualifies, what the limits are, what the business is willing to absorb. The system processes ordinary refunds inside that policy and escalates the cases that fall outside it. The human is not approving each refund; they are responsible for the policy, and for noticing when the exception rate says the policy is wrong.
• Hiring. A human determines what the rubric values and why — which is a judgment about what the role actually requires, and one with real consequences for real people. The system organizes or screens against that rubric and flags the candidates it is uncertain about. What it must not do is quietly narrow the field on criteria nobody stated.
• Workflow escalation. A human defines the conditions under which a running process must stop and surface itself. The system handles routine cases and reports deviations. The value here comes from sustained attention over time, not from the system deciding what counts as a deviation.

The pattern is the same in all three: the human decision is made once, in advance, and explicitly, and it is the thing that makes the delegation safe. The system's autonomy is real, and it is bounded by that decision rather than by a per-case approval ritual.

What this does not claim

• Automation does not remove human responsibility. Responsibility stays with the person who set the boundary and who answers for the outcome. It cannot be delegated to a process, and a process cannot accept it.
• Not all AI work is suitable for delegation. Some work is not routine and will not become routine by being described more carefully.
• This is not legal, compliance, safety, or risk advice. Regulated decisions carry requirements this page does not address.
• A clear boundary does not guarantee correct outcomes. It makes the decisions inspectable and the failures locatable. That is a different property from being right, and it is the one you can actually build for.

Where this sits

This page explains a general decision boundary. It is not specific to any product, and the distinction holds whether or not you use one.

The related question — whether a workflow you already run is safe to lean on — is a different object, and the one organizations usually need first: a boundary that was never written down is hard to evaluate, and most reliance is inherited rather than designed. The general form of the problem is governable AI action under human authority — AI acting under your name with legibility, bounded delegation, reviewable memory, and inspectable action, so that what a system did can be examined rather than assumed. For the larger object this sits inside, see the Verse.