AI should draft and challenge. It should not decide.

Every planning tool has shipped an AI feature in the last two years, and most of them are pointed at the wrong half of the job. The pitch is usually some version of "it prioritises your backlog" or "it estimates for you" — the parts that look most impressive in a demo and are least defensible in a meeting.

The useful split is not about how good the model is. It is about what happens when the output is wrong.

The test: reversible and verifiable, or not

Ask two questions of any task you are considering handing to a model.

Can a human verify the output in seconds? Not "check it carefully over an afternoon" — glance at it and know. Sorting a pasted request into facts and assumptions is verifiable in seconds by whoever read the request. A capacity estimate is not verifiable at all; it is only ever confirmed or refuted by events months later.

What does a wrong answer cost? A bad draft costs the seconds it takes to delete a line. A bad commitment costs a quarter, and the cost lands on people who never saw the model's output.

Where both answers are cheap, use the model freely. Where either is expensive, a person decides — and the model's job shrinks to putting the right things in front of them.

What that leaves AI genuinely good at

Turning prose into structure. Someone pastes a Slack thread. A model produces a candidate title, problem statement and outcome. This is close to ideal: the input is unstructured, the output is verifiable at a glance by the person who pasted it, and a wrong field is one edit.

Finding what a plan has not answered. Models are good at noticing absence, which humans are bad at — we read what is on the page. "This plan names a dependency on the platform team but never says what is needed from them or by when" is a useful sentence, and it is checkable instantly.

Challenging a too-clean claim. When someone marks a demand as having no known risks, a model that has read the shape can ask one pointed question. Sometimes the answer is "yes, genuinely clean". Sometimes it is "…ah, the migration". Either way it costs one sentence.

Drafting the boring parts of a summary. A one-page brief from structured data the human already approved. Reformatting, not deciding.

Notice the pattern: every one of these produces something a person then accepts, edits, or throws away — in a few seconds, with full context.

What it must not decide

What gets committed. A commitment allocates other people's time against other people's priorities. It is the single most political act in delivery planning, and the model has no access to the context that actually determines it.

Who owns something. Ownership is a social fact. A model can suggest a plausible name from who has been mentioned; whether that person accepts is not derivable from text.

How much capacity something takes. A number produced without knowledge of the codebase, the team, or what else they are carrying — and delivered with the same fluency as a number that came from the team. The fluency is the danger. A team's own rough estimate carries a signal that a generated one does not: someone will be held to it.

Whether something is ready. Readiness is a judgement about whether you have enough evidence. Outsourcing it is outsourcing the entire point of having a process.

Why "advisory" is not enough on its own

The common defence is that the AI is only a suggestion and the human can always override. True, and insufficient — because the interface decides how often that actually happens.

A pre-filled field with a confident value gets accepted. A field that says "suggested: 3 FTE-months — confirm with the team" gets thought about. A suggestion you must explicitly accept or reject leaves a record of a decision; one that silently becomes the value leaves a record of the model's output with a human's name on it.

So "AI suggests, human decides" is a claim about the UI, not about intent. If the product cannot show you which values a person actually chose, it is not what is happening.

The part nobody likes discussing

Anything a user types is untrusted input, and that includes the text you feed to a model. A pasted request can contain instructions aimed at your prompt — and a pasted request is, by design, the main input to this kind of product. Treating demand text as data rather than as instructions is not an edge case here; it is the normal case.

The right place for that rule is one chokepoint every call goes through, not a line repeated in each prompt. Repeated in each prompt, it is one forgotten copy-paste away from not being there.

How this is drawn in DeliverySheet

The rule is "AI suggests, the human owns" and it is enforced structurally rather than promised in the marketing.

AI drafts a demand from pasted text, produces the clarification brief separating facts from assumptions from open questions, suggests a next step at the decision point, and challenges a "none known" readiness claim. Every one of those is labelled as AI-generated with a confidence level, and the accept-or-reject is a separate recorded event from the generation — so the log distinguishes what the model proposed from what a person kept.

What it never does: decide a commitment, set capacity, assign ownership, or judge readiness. Capacity comes from the delivering team with a confidence level attached. Commitment requires a person, and committing early requires four recorded fields. Deeper shaping fields — impact, known facts, assumptions, unknowns — are human-owned and deliberately excluded from the AI draft; you can ask for a suggestion per field, but nothing pre-fills them.

And the injection guard is exactly the single chokepoint described above: every AI call in the product routes through one function that appends it, so a new prompt cannot accidentally ship without it.

None of this is a bet against models getting better. It is a claim that the line is drawn by accountability rather than capability — and that line does not move when the model does.