Audiences AIFourteen Pages, No Holes Showing
The deck arrived on a Sunday night with a one-line message: would love your read before Thursday. Fourteen pages. Five recommendations. A founder I have known for years had built it with an AI assistant in an afternoon, and it read better than most decks I had reviewed all year. Clean charts, a summary page a board would applaud. I skimmed the first page and called him instead of reading on. What follows is a composite of two engagements; numbers are rounded and identifying details withheld.
The call was short. I told him I would give the deck a proper review once he sent me four answers:
- What decision are you asking me to help you make?
- Which few numbers does the recommendation stand on, and where did each one come from?
- What information was missing when this was built?
- What evidence would change your mind about the lead recommendation?
It took him a full day to assemble the answers. That day did most of the work: it meant tracing each headline number back to wherever it was born, and two of the trails ended somewhere surprising.
The lead recommendation was to put more money behind the highest-revenue category. The revenue figure turned out to be gross, before returns and cancellations, and returns in that category ran near 35 percent. On net contribution the category placed fourth. The number one growth bet in the deck was the fourth-best place to put the next unit of spend.
The second finding was worse. The category the deck wanted to cut had been out of stock for eleven weeks in its two best-selling sizes. Demand had not fallen; availability had. The model read a supply failure as a demand signal and recommended finishing the job the stockout had started.
Nothing was hallucinated
This is the part I keep retelling, because it cuts against the fear most executives carry into these reviews. They are braced for fabrication: invented citations, figures with no source behind them. Obvious nonsense gets caught. What almost slipped through here was different in kind. Every number in the deck was real. Every claim was consistent with the data the model had been handed. The analysis was faithful to its inputs and silent about everything its inputs left out.
Sit with how the hole disappeared. A returns column that never made it into the export does not appear as a gap; it appears as nothing at all, and to a model, nothing looks like a fact about the world. The system completes the picture from whatever is present, the way autocomplete finishes a sentence: fluently, plausibly, with no flag that a word was ever missing. Gap-filling is the default behavior. Disclosure has to be forced.
Polish then disarms the reviewer. A ragged internal spreadsheet invites questions; a typeset fourteen-pager with a table of contents signals that the questions were already asked and settled. The better an artifact reads, the less scrutiny it draws, and AI has made the best-reading artifact the cheapest one to produce. That inversion is the governance problem underneath this whole story. Effort used to be a rough proxy for diligence. That proxy is gone, and most review habits were built on it. The old instinct that a heavy, well-made document means somebody sweated the details now selects for exactly the documents that deserve the most doubt.
The standing intake
The second engagement in this composite ended the same way: different category, same silent fill, caught later and at higher cost. After that one I stopped treating my four questions as a phone call and turned them into paper. Anyone who wants me to review an AI-assisted recommendation now sends a one-page intake first, covering four things:
- The decision being asked for, in one sentence.
- What was assumed, excluded, or unavailable while the analysis was built.
- The few numbers the recommendation depends on, each with its source named.
- The evidence that would cause the author to withdraw the recommendation.
And one name at the bottom: the person who stands behind the work. The AI can draft every page. A human still owns the recommendation and the quality of the evidence underneath it, and if nobody will put their name on that line, the review is already over, because that absence is the finding.
The intake costs its author about an hour, and it changes the meeting. Half the time the author finds the hole while filling in item two, and the review I get asked for becomes a different, better one. The other half, we spend the hour on the decision itself instead of on archaeology. Either way the fourteen pages stop standing in as the evidence and go back to being what they always were, an argument.
The intake catches decks, and decks are the symptom. Somewhere upstream of every silent fill is a person who learned that arriving with an answer beats arriving with a gap. Punish honest uncertainty in enough meetings and your people will fill the holes quietly and confidently long before any model does; the model just made the filling cheap and gave it a table of contents. So the last thing I ask has nothing to do with the artifact. What happens, in your organization, the next time somebody says, "We don't know yet"?
If one of these decks is sitting on your desk ahead of a real decision, send it over.