Why AI Presentation Generators Get the Numbers Wrong (and How to Check a Deck)

The slide looks finished, the chart is on brand, and the number in the title was never calculated. Here is where the wrong figures come from, and how to catch them in ten minutes.

AI presentation generators get numbers wrong because a language model writes text that is plausible, and a number is just more text. Unless the tool actually runs a calculation against your data, the figure in the slide title was generated to fit the sentence, not computed. Even when you upload a spreadsheet, the same text-writing process often does the summarizing, rounding and comparing, so every number in an AI-made deck needs checking before it reaches a meeting.

That is not an argument against using AI for decks. The wrong numbers come from a handful of predictable places, and each one has a quick check.

Generated, not computed

A language model produces a sentence one piece at a time, choosing what is most likely to come next given everything before it. Ask for a slide on "Q3 performance for a mid-size software company" with no data attached, and the model will produce a title, three bullets and a chart that look exactly like a Q3 performance slide. The 18% revenue growth in the title is there because a growth figure belongs in that sentence and 18% is a believable one. Nothing was looked up and nothing was added up.

How often this happens is hard to pin down, because it depends heavily on the task. One 2026 write-up reported that independent testing of 20 data-driven slides found only 12% of the statistics had verifiable sources and 41% were fabricated outright. Twenty slides is a small sample and the testing is not published in detail, so treat it as an illustration rather than a rate. The broader research says the same thing with more nuance: models invent far less when the answer is sitting in a document in front of them than when they answer from memory, and we went through why published rates range so widely in AI hallucination statistics for 2026.

For a deck, the practical point is simple. A prompt-only deck has no source, so every number in it is in the "answer from memory" category.

Six ways the numbers go wrong

Attaching your own data removes the first problem and leaves the rest. These are the failure modes worth knowing by name, because each one leaves a specific fingerprint on the slide.

1. The number was written, not looked up

The prompt-only case above, but it also happens inside decks built from real data. A slide about a region the data does not cover, or a metric the file does not contain, gets filled anyway, because the tool's job is to produce a slide and a slide with a number in it looks more finished than one without. The fingerprint: a figure that is suspiciously round, or a benchmark ("industry average churn is 5%") with no source.

2. Your data was summarized and rounded

Many tools treat an uploaded spreadsheet like a document: the model reads the file, or a slice of it when the file is too large to read whole, and writes about it. A total in that kind of summary may be the model's estimate of the column, not the sum of it. Rounding compounds it: say March net sales are $1,284,600; the slide says $1.3M, the takeaway says "up about 9%" on a February of $1.19M, and the real change is 7.9%. Each step was a reasonable approximation and the result is a figure nobody can reproduce. The difference between the model writing the arithmetic and a database doing it is the central finding of how accurate AI is at analyzing data.

3. The percentage is taken over the wrong total

A pie chart shows the top five product lines, and the label says Product A is 34% of revenue. It is 34% of the five lines on the chart. Across all fourteen product lines it is 22%. This is the most common share error in any deck, AI-made or not, but AI tools make it more often because the chart and the percentage are both built from whatever rows were in front of the model at the time, and the top five were what it was given. Win rates have the same trap: won deals divided by all deals is a different number from won divided by closed deals, and the slide rarely says which one it used.

4. The comparison uses mismatched date windows

"Q3 revenue is down 31% on Q2." Q2 is a complete quarter. Q3 is the first eight weeks of the quarter, because that is all the data there was when the deck was built. The decline is mostly the four missing weeks. The same thing happens with month-to-date against a full prior month, a fiscal year against a calendar year, or this year's year-to-date against all of last year. The fingerprint is a dramatic change that nobody in the business noticed happening.

5. The takeaway quotes a figure that is not in the chart

The chart shows monthly revenue by region. The bullet beside it says "Northeast grew 14% while margins held steady." There is no margin anywhere on the slide, and possibly nowhere in the data. Takeaways are where a model is most tempted to add context, because a good takeaway usually connects the chart to something else, and the something else is the part that gets invented.

6. The chart and the text were produced separately

In many tools, the chart is drawn in one step and the words in another. The chart might come from the actual rows while the bullets come from the model's summary of the file, or the title might be written before the chart exists. Nothing forces them to agree, so a slide can show a bar at $412K and a title that says $430K. This is also why regenerating a slide can change its numbers: each pass re-derives everything, including, as we explained in why AI gives a different number every time, what the metric means.

Why these are hard to catch by reading

A wrong sentence usually reads wrong. A wrong number does not. It has the right format, the right number of decimal places, roughly the right magnitude, and it sits under a well-designed chart in your brand colors. The polish of the slide is exactly what lowers the reader's guard. The only reliable check is to stop reading the deck as prose and start reading it as a list of claims.

How to check an AI-made deck in 10 minutes

This is the routine for a ten-slide deck. All it needs is the source data and a calculator, and it is worth doing whatever made the deck.

1. List every number (2 minutes). Titles, chart labels, axis values that are quoted elsewhere, takeaway bullets, the summary slide. Write each one down with its slide number. Ten slides usually carry thirty to sixty figures.
2. Trace each number to a source row (3 minutes). For each figure, point at the row, the cell or the query result it came from. Anything you cannot trace to your own data or a cited source is treated as invented until shown otherwise. Benchmarks and "industry averages" with no link go first.
3. Recompute every share (1 minute). For each percentage, ask "percent of what?" and check that the denominator is the full total, not just the categories on the chart. If shares on a slide add to exactly 100% but the chart only shows the top few, the denominator is wrong.
4. Check the window on both sides of every comparison (1 minute). For every "up" or "down," confirm both periods are the same length and both are complete, or both are the same number of days into a partial period.
5. Match every takeaway to its chart (2 minutes). Every figure in a bullet or title must be visible in the chart beside it, or directly computable from it. A bullet that mentions a measure the chart does not show gets deleted or sourced.
6. Reconcile the summary slide (1 minute). Every number on the executive summary should appear on a body slide with the same value, the same rounding and the same period. Summary slides are written last and drift most.

What a deck tool has to do differently

The checklist works backwards from the finished deck. A tool can do the same work forwards, and the requirements fall straight out of the six failure modes:

• Every number comes from a calculation run against a dataset, not from the model's reading of it.
• The chart is drawn from exactly the rows that calculation returned.
• The words beside the chart are written from the chart figures, and any sentence quoting a figure that is not in those rows is removed.
• Shares and comparisons are checked for the denominator and the date window.
• A slide that cannot be answered from the data says so, instead of being filled.
• Metrics mean the same thing on every slide, every month.

How Quiriz builds a deck from your data

Quiriz for PowerPoint is built on that list. It is not a design-from-a-prompt tool: it builds slides from your datasets, in your template. You upload spreadsheets or connect a source (HubSpot connects directly, and Google Drive or OneDrive folders refresh on a schedule), pick the data, and say what the deck is for.

Quiriz plans the deck and shows you the plan before it builds anything: each slide's title, the question it will ask, and whether it will be a chart, a table or text. Edit the titles, untick what you do not want, add your own slides, then press Build.

Each slide's question is answered by a query against your data, so the arithmetic is done by the database, not written by a model. The chart is drawn from the rows that query returned, and the two to four takeaways beside it are written from that chart's figures. Takeaway bullets whose figures are not in the answer's rows are dropped before they reach the slide. Checks catch a share computed over the wrong total and the two halves of a comparison using different date windows. The deck opens with an executive summary, written last from what the slides found and checked against them.

When a slide's question cannot be answered from the data, or the data does not match what was asked, the slide says so and why. You get a slide that tells you the data has no margin column, not a slide with a margin figure on it.

Definitions come from Company Context: an admin writes down once what win rate, net sales or pipeline means for your business, and every slide, in every deck, uses those definitions. The same question returns the same number in PowerPoint, Word, Excel and Slack, because every surface asks the same datasets.

And the charts are native, editable PowerPoint charts, not pictures. Right-click any chart, choose Edit Data, and the numbers behind the bars are right there for anyone reviewing the deck to check against the source. Restyle them like any chart you built yourself, and when new data arrives, Refresh all asks every question again and redraws each chart, table, figure and takeaway in place, keeping your formatting.

Where this does not apply

If the deck is a pitch, a brand story, a training module or a keynote, with no business data behind it, none of this matters and a design-first AI deck tool is the right choice. The problem in this article is specific to slides that quote numbers. If a slide quotes a number, that number should come from a query you can trace, and if no tool you use can show you where a number came from, the ten-minute checklist above is the minimum.

Frequently asked questions