Using AI in delivery without losing control

AI is useful in delivery work for unremarkable reasons. It reads long documents quickly, drafts a first version of something structured, and notices patterns across records nobody has time to compare. None of that requires it to make decisions.

Where it earns its place

• Reading: pulling deliverables, phases and assumptions out of a document into a proposed structure.
• Drafting: a first plan, a first estimate, a first client update — reviewed, then corrected.
• Noticing: the task that has been open too long, the overlap next month, the margin that moved.
• Explaining: why a number changed, with the underlying records to check it against.

Where a human must stay in the loop

Anything that changes a commitment, a plan, a price or something a client sees. Not because the suggestion will be wrong, but because accountability cannot be delegated to a suggestion. The right pattern is propose, review, approve, apply — with the option to correct the proposal rather than accept or reject it wholesale.

Insist on three things

1. Grounding: the suggestion should cite the records it came from.
2. Reversibility: applied changes should be visible and undoable.
3. A trail: who approved what, and when.

Judge it by time returned, not novelty

The question is not whether a feature uses AI. It is whether the work now takes less time and produces fewer surprises. If a suggestion has to be checked as carefully as doing the task by hand, it has not helped.

Set expectations with the team

Say plainly what the assistant does and does not do. Teams trust AI in delivery when its role is narrow and its output is checkable — and stop trusting it the first time something changed without anyone agreeing to it.