Use AI on the last stretch of quality, not the first draft. Move a solid A-minus package to A-plus by hunting remaining gaps.
An A-minus package already has a clear ask, real evidence, and a structure the room can follow. What it still has are soft joints: a metric without a definition, a competitive claim with no primary source, a risk that is named but not owned, a timeline that assumes a dependency nobody confirmed, or prose that sounds finished while the backup tabs disagree.
When first-draft AI hurts
Asking a model to generate the first version of a board memo, legal outline, or investor narrative often produces fluent structure with invented specifics. You then spend the review cycle cleaning fiction instead of testing judgment. For consequential work, draft with your sources open. Bring AI in when the package is already coherent and you need a second set of eyes that will not protect your ego.
The refinement pass is also the wrong place to ask for a brand-new strategy. If the recommendation itself is unsettled, step back to war-gaming or decision review. A-plus polish on the wrong decision only makes the wrong decision travel farther.
What the refinement pass should ask
- Which claims would a skeptical director, counsel, or customer reject first, and why?
- Which numbers lack a definition, date range, or source tab?
- Where does the narrative disagree with the appendix?
- What alternative did we dismiss too quickly?
- What would have to be true for this recommendation to be wrong?
- Which sentences imply certainty the evidence does not support?
Verify every factual claim
Run that list through Pingpong against the actual package. Sequential multi-model review helps because later models can challenge earlier soft approvals instead of repeating them. Treat every suggested fact as unverified until you match it to a source. Keep wording changes that clarify ownership or uncertainty. Drop stylistic rewrites that add length without closing a gap. A-plus here means fewer places for the room to catch you, not louder language.
Time-box the pass. Sixty to ninety minutes of review plus source checking usually beats an overnight rewrite that reintroduces old errors. When the remaining gaps are judgment calls only you can make, stop the models and make the call.
Also see bulletproof before you present, catch AI mistakes before you commit, AI decision review, dual-speed AI for leaders, and multi-model review. Walkthrough: first review. Mechanics: how it works.