The Slide Deck Is Not the Data
If you run the numbers through AI to make them better, you may get a cleaner argument and a less accurate meeting
This post is part of my Medium blog.
I hate PowerPoint.
Not because presentations are always bad. Sometimes a few slides are the right way to show a decision, especially when the decision depends on a visual comparison or a small number of things people need to see together.
But most workplace presentations are not decisions. They are group reading exercises where twenty people sit in a room and watch one person click through slides that could have been sent in an email, if anyone had been willing to write the email.
Nobody writes the short paper.
If somebody does write it, nobody reads it. So we make a PowerPoint, put the important thing in the speaker notes, bury the assumptions in an appendix, and schedule a meeting where the entire room can discover the first slide together.
This is how a lot of work gets done.
In Chapter 1 of Redundant, Rob Coleman arrives at a meeting with a slide deck. He is the Director of FinOps, which means he is responsible for making the company's cost reality legible to people who would often prefer a more encouraging version of it.
There is always something to talk about. Something costs too much. A system is running below the utilization someone promised. A cloud provider has a new instance family that could save money if the workload actually behaves the way the model says it does. Someone wants to move platforms and needs numbers that make the migration look inevitable.
FinOps always has a slide.
Rob has done the work. He has built the analysis, checked the numbers, and prepared the deck. He sends it to the CTO's executive assistant before the meeting. This is normal. The assistant is trying to help. The deck could be cleaner. The executive summary could be more concise. The story could be easier for a room full of vice presidents to follow.
So the assistant runs it through a generative AI tool.
The tool improves it.
That is what the tool says, anyway.
By the time Rob gets to the meeting, his work has been summarized by a system that did not ask him what the numbers meant. The slide on the wall says cloud run rate is down thirty percent year over year. The actual number is up thirty percent.
The deck looks better. The number is wrong.
That's the AI experience a lot of people are having now. Not spectacular failure. The dangerous failure is the one that looks like an improvement. The summary is cleaner. The language is more executive. The bullets are shorter. The chart still has the right colors. The number has quietly changed.
Nobody in the room is looking for the error because the point of the transformation was to make the material easier to trust.
This is what these tools do well: they introduce subtle errors into things that already looked finished.
I use generative AI constantly. I know what it is good at. It can help research a topic. It can suggest an outline. It can give you a different way to approach a paragraph. It can take a long document and produce a first-pass summary that helps you find the section you need.
It can also produce a summary that is fluent, plausible, and wrong in exactly the place where nobody thought to check.
The assistant using AI was not the problem. The problem was treating transformation as verification. The deck was sent into a system, came back with better formatting and a more confident tone, and everyone behaved as though the second version had been reviewed by somebody who understood the first version.
It hadn't.
The model had rearranged the material. That is not the same as understanding it.
The failure gets worse in FinOps because FinOps data is rarely neutral in the way people want it to be. Somebody asks you for numbers, but they may not be asking because they want to know what the numbers say. They may already have an idea and want the numbers to make the idea look responsible.
"Can you give me the cost comparison for moving this workload?"
That sounds like an analytical question. It may be. Or the migration may already have been approved in somebody's head, and now they need a slide that turns the decision into a financial argument.
The data arrives. The argument changes.
A three-year cost becomes a first-year savings number. A migration risk becomes a footnote. A range becomes a midpoint because the midpoint fits the slide. The analysis says the savings depend on utilization. The executive summary says the move will save money.
Nothing in the spreadsheet has to be false for the presentation to become misleading.
The selection does the work.
This is one of the reasons FinOps can be a miserable job. You can provide accurate data and still watch somebody move the meaning around until it supports the conclusion they wanted before they asked you the question.
The assistant who runs your deck through an AI tool may not be trying to alter the evidence. They may be trying to help. The executive who reframes your numbers may not think they are lying. They may believe they are clarifying the decision for the room.
That is how the problem survives reasonable people.
Everyone is improving the material. Nobody is checking whether the improved material still means what the original analysis meant.
A short paper would at least make the argument visible. Someone could read the assumptions. Someone could see the caveat that does not fit in the bullet. Someone could ask whether the conclusion follows from the evidence before twenty people spend an hour performing agreement with it.
PowerPoint makes that harder because the format rewards compression. One sentence replaces the paragraph that qualified it. One number replaces the range. One arrow replaces the explanation of what had to be true for the arrow to point that way.
Then the slide gets presented as though the simplicity is evidence of clarity.
It isn't. Sometimes it is evidence that the inconvenient parts have been removed.
Rob's problem in the meeting is not only that an AI summary has changed his number. It is that the room is prepared to trust the summary because it looks like the kind of thing executives are supposed to understand quickly. The polished version has institutional authority before anyone checks whether it is true.
That is a dangerous property for any document.
The same thing happens when a FinOps leader is asked for data to support a decision that has already been made. The requested output is not a report. It is a prop. The person with the title wants the person with the numbers to stand behind the decision, and the slide gives them somewhere to stand.
The numbers can be accurate. The argument can still be dishonest.
This is not an argument for refusing to summarize anything. Executives need summaries. People have limited time. A good summary is one of the most useful things a team can produce because it lets someone see the important part without reading every artifact behind it.
But a summary is a compression of an argument, not an independent source of truth. The person who reads it needs to know what was compressed. The person who generated it needs to be accountable for what disappeared. And if a model transformed it, someone who understands the original needs to check the result.
Otherwise, you have not made the data better. You have made the story harder to question.
In Redundant, Rob sees the bad number on the wall before he begins the meeting. He has the original deck. He knows what the summary changed. He can correct it, but he cannot undo the fact that the room saw the wrong version first.
That is the part people underestimate. Once a polished error is on the wall, the accurate version has to fight its way back into the room. It is no longer just presenting data. It is correcting the atmosphere created by the first version.
The slide deck is not the data.
The executive summary is not the analysis.
And "make it better" is not a verification process.