A deck built from a spreadsheet has to survive being questioned

There are three, and each fails differently.

Paste as a picture. Copy the chart, then Paste Special → Picture (Enhanced Metafile) or PNG. The slide now holds an image. It renders identically on every machine, it cannot be accidentally edited, and it will never prompt anyone with a dialog box. It also cannot be corrected in the room. When someone spots that a series is mislabelled, or asks to see the same chart without the outlier, you have nothing to work with but a screenshot. And because the image carries no data, nothing about it can be verified after the fact.

Paste with a link to the source. PowerPoint keeps a reference to the workbook and refreshes the chart from it. This is the option that sounds correct and causes the most damage. The link is a file path. It resolves on your machine and on nobody else's, so the recipient gets either a security prompt about updating links, an error, or, worst, a chart that quietly renders the last cached state while looking live. If the workbook moves, is renamed, or lives on a drive the recipient cannot reach, the slide is now showing a number with no provenance at all. Use linked charts only when the deck and the workbook stay in one place and one pair of hands.

Rebuild the chart in PowerPoint. Insert → Chart gives you a native chart backed by a small embedded worksheet. It travels well, it is editable in the room, and it is fully under your control. Its failure mode is drift: the embedded sheet is a copy taken at one moment, and when the source figures are revised nobody updates both. Six weeks later the deck and the model disagree and neither knows it.

There is no option without a failure mode. Pick by audience. Circulating widely to people who will not open the underlying data: picture. Presenting live to people who will push back: rebuilt native chart, with the workbook open in another window. Keeping a deck in sync inside one team on one shared drive: linked, and accept the maintenance.

The same trade runs through decks built from a BI tool. Exporting visuals from a Power BI report gives you pictures with no data behind them, and every question in the meeting becomes an action item. There is a longer treatment of that particular conversion if that is the path you are on.

Fewer than you have. A deck presenting an analysis needs about six slides in the body, and each one has a job.

The finding. One slide, one sentence, one number. Not "Q3 performance review" but "Returns rose from 4.1% to 9.3% between June and August, driven almost entirely by one product line". If you cannot write that sentence, the analysis is not finished and the deck will not save it.

The evidence. The chart that shows the finding is true. One chart. If it takes three charts, you have three findings and should say so.

The mechanism. Why the number moved. This is where a waterfall chart earns its place, because it is the only common form that decomposes a change into its contributing parts and shows the arithmetic of how you got from the old figure to the new one. It converts "revenue fell by £180k" into four bars that add up, and it pre-empts the question about which part of the business is responsible.

The counter-check. What you tested that would have shown the finding to be false, and what happened. This slide is unusual and it is the one that buys you credibility for the rest of the deck. How to tell a real finding covers what belongs on it.

The decision. What you want the room to do, with the options and what each costs. A deck that ends at the finding transfers the work back to the audience, and they will not do it.

The provenance slide. File name, sheet, date range, row count, exclusions and metric definitions. Four lines of small text. This is the slide that ends arguments.

Everything else goes in the appendix.

Being able to say where the number came from.

Not "the finance export", but: this came from orders_2026_q3.csv, 41,882 rows, order dates from 1 June to 31 August, revenue net of refunds and excluding VAT, 214 test orders removed, 63 rows dropped because the order date would not parse.

That sentence does three things at once. It answers the question. It shows the analysis was done carefully enough that the answer was to hand. And it converts the room from auditing you to discussing the finding, which is the entire purpose of the meeting.

Keep it written down. On the provenance slide if the audience is technical, in your notes if not. The failure is not that people cannot remember row counts; it is that they never recorded them, so the number does not exist to be recalled.

Chart form matters here too, and for the same reason. A truncated axis, an inconsistent time interval or a dual axis chosen to make two lines cross will be noticed by exactly one person in the room, and that person will then doubt everything else. How to spot a misleading chart is written for the reader, but it is more useful read as a checklist for the author.

The rule is simple: the body carries what you will say, the appendix carries what you will be asked.

Body: the finding, the evidence, the mechanism, the counter-check, the decision, the provenance.

Appendix: the segment breakdowns nobody asked for yet, the alternative definitions of the metric, the sensitivity of the result to the exclusion rules, the same chart on a longer time axis, the method notes, and the full detail table. Anything you built while working that did not make the argument goes here rather than in the bin, because the value of the appendix is that you can jump to a slide instead of saying you will follow up.

A good ratio is six body slides to fifteen appendix slides. It looks lopsided and it is exactly right. Nobody is annoyed by an appendix they do not read. Everybody notices the question you could not answer.

If the audience will not ask questions, none of it applies.

A monthly update where the numbers are routine, a status slide inside a longer deck, a chart for a document that will be skimmed once, paste it as a picture and move on. Adding a provenance slide to a deck nobody interrogates is work spent for no return, and the honest advice is to skip it.

The judgement is about what happens if a figure turns out to be wrong. If the answer is that someone corrects it next month, paste the picture. If the answer is that a decision was already made on it, build the deck properly. Most people can tell which situation they are in and talk themselves out of it anyway, because building properly takes longer and the meeting is on Thursday.

Data Analysis App returns an editable PowerPoint deck alongside the cleaned data and the dashboard, with the row counts, date ranges and exclusions already written down rather than left for you to reconstruct. The mechanism behind the figures is the point: a language model may decide what to compute, but it never decides what the answer is, and every figure is computed twice by two independent engines with anything the two disagree about dropped rather than published. The published benchmark shows what that does and does not catch, including where it misses. There are sample outputs under examples.