The clarifying question is gone
Here's the safety net you didn't know you had: every vague ticket you ever wrote got quietly repaired by a developer who read it, frowned, and asked what you actually meant. "Validates properly." "Handles edge cases." "Feels responsive." Humans metabolize that mush into questions. The question arrives at standup, you sharpen the requirement, the build proceeds. Your specs were never as good as your delivery record — the team's frowns were doing the last mile.
An AI builder doesn't frown. It reads "make the export robust," picks one of the forty things robust could mean, and implements that one — fluently, completely, and overnight. By morning you have working code, passing its own tests, solving a problem adjacent to yours. Now you're in a rework loop, and rework against a vague spec produces a different wrong thing, because the spec still doesn't say which thing is right.
What the evidence actually shows
The best field data on AI-assisted knowledge work comes from a preregistered experiment on 758 consultants at a major strategy firm, published in Organization Science in March 2026. On realistic tasks inside the AI's capability frontier, people with AI access were about 25% faster, produced work rated about 40% higher in quality, and completed about 12% more tasks — with the lowest performers gaining the most. On a task the researchers deliberately picked from outside the frontier, AI users were 19 percentage points less likely to reach the correct answer than people working alone.
Read that as a delivery lead, not a technologist. The same tool, handed to the same people, made the work dramatically better on one side of an invisible line and actively worse on the other. Nobody hands you a map of the line. Which means the two decisions that matter most now happen before any build starts:
- Is this task on the delegate side of the frontier? (Lesson beat six.)
- Is the ticket specified tightly enough that "done" is checkable? (Everything between here and there.)
The PM job, restated
You already run a version of this discipline for vendors. No PM sends a statement of work reading "build us something good" — you write deliverables, acceptance criteria, and a scope fence, because the vendor will build exactly what's written and invoice you for the rest. AI builders are the fastest, most literal vendor you will ever manage. This lesson takes the SOW instinct you already have and compresses it into a ticket structure a script can lint — so the vague ones get bounced by code at write-time, not discovered by you at review-time, three builds too late.