How do I fact-check an AI-built presentation before I present?
Short answer
Have the AI build a sources-and-confidence table with the deck: every claim, where the number came from, and a confidence percentage. Flag anything under 95 percent, cut any claim with no measured baseline, and archive older files that quote conflicting numbers. Read the table before you rehearse a single slide.
Tuesday night I had an AI build a 33-slide conference talk. The most useful thing it produced sat at the back of the brief: a table listing every claim on the deck, where the number came from, and how sure it was that the number was right.
Twelve rows. Two columns I expected (claim, source) and one I now refuse to present without: confidence, as a percentage.
What belongs on a slide?
A number without a source does not go on stage. A number without a baseline does not go on stage either, even when it would make the talk land harder. The AI keeps the table. You read the table before you rehearse. That is the division of labor.
How the table earned its place
Anything under 95 percent got flagged. One row came in at 90 percent because the article behind it sits behind a paywall. So the slide says the number with the source visible and the caveat spoken out loud.
Anything with no measured baseline got cut entirely. I had no before-and-after data on time saved, so the deck makes no time-saved claim. It would have been the easiest applause line in the room. It is not there.
The table also caught a pricing problem I would have walked on stage with. Three older documents on my drive quoted three different monthly prices for the same offer. The deck carries the current one, and the brief told me to mark the other files ARCHIVE so a contractor never quotes a dead number to a live prospect. That single line probably saved me an awkward email in October.
Show the real thing
Twelve of the 33 slides carry a real screenshot of something that was built and is running. No mockups, and no diagrams of what could exist. When you can show the thing, the adjectives can go.
Why this matters now
AI can produce a polished deck in an evening, and it can produce a plausible statistic in a second. A room of 300 people has 300 phones. The sources-and-confidence table makes your talk auditable by you before it is auditable by them.
The same table works on a proposal, a board update, or a sales one-pager. Anywhere a number carries weight, the reader deserves to know where it came from, and you deserve to know first.
Do this today
- Open the next presentation or proposal you are about to send.
- List every number on it in one column.
- Next to each, write the source.
- Next to that, write how confident you are, as a percentage.
- Cut or soften any row with a blank source.
- Find any two documents in your files that disagree on a number, and archive the older one this afternoon.
Meredith's rule
A number with no source and no baseline does not go on the slide.
Questions
How do I make sure statistics in an AI-generated deck are accurate?
Ask the AI to keep a table of every claim with its source and a confidence percentage. Flag anything under 95 percent and say the caveat out loud on stage. Cut any number that has no source or no measured baseline behind it.
Should I use time-saved statistics in a presentation about AI?
Only if you measured a before and after. I cut a time-saved claim from my own conference deck because I had no baseline data, even though it would have been the easiest applause line in the room. An unmeasured number invites a challenge.
What should I do when my files list different prices for the same offer?
Pick the current price for anything going out, then mark every older file ARCHIVE so nobody quotes a retired number to a prospect. My own deck review turned up three different monthly prices for one offer across three documents.
Related service
AI Architecture Review
Want this worked out for your business?
Bring the real problem to a working session and leave with a plan.
Book a session