AI presentations: inspect the file your audience will actually receive
Review an AI presentation from source material through editable slides to final export. Check claims, charts, references, notes and the recipient’s file.
Article language: English

In this article
In the editor, everything was in place: the number, the reference and the small qualification "percentage of invitees who responded". After export, the recipient sees only an impressive 64 percent. A few words have disappeared, and a much bolder claim has emerged. A presentation should not grow more confident on its way to becoming an attachment.
This is a fictional example of Mera, a company summarising workshop participation. Thirty-two of fifty invitees completed the form, giving a response rate of 64 percent. AI turns a note into slides. The team needs to review their content first, then the file actually sent to the audience. An export that matches the editor cannot repair a mistaken interpretation of the data; a correct source does not establish that nothing was lost while assembling the slide.
Three versions of the material, three different questions
Distinguish the source material, the editable presentation and the final export. The first supplies the basis for claims. The second enables work on slide layout and content. The third is what the selected audience receives. A shared title such as "workshop summary" is not enough. Identify the version of each part and connect them in a short review record.
In its instructions for file-creating agents, Microsoft recommends checking generated material for accuracy before sharing. That guidance concerns a particular product. For Mera, we turn it into our own test: trace a selected sentence in the note to its slide and the corresponding page in the export. We do not treat a generated preview as automatic approval of facts.
Before generation, state what the presentation should enable. The team may want to decide whether to improve workshop invitations. That differs from presenting participant satisfaction or evaluating the entire service. Preserve definitions of numbers, dates and limitations in the input. Reducing the note to striking percentages removes part of the basis against which you will later compare the result.
For Mera, "we received responses from 32 of 50 invitees" does not mean "64 percent of participants recommend the workshops". The second sentence describes different information. If AI has turned a response rate into a recommendation, improving the layout resolves nothing. First reject the unsupported conclusion and restore the number's meaning. Only then check its appearance on the slide.
Mera's calculation is thirty-two divided by fifty, multiplied by one hundred. Check what each count means in the material: invited people, forms started or complete responses. A repeat submission should not become another person without an agreed rule. If the input cannot distinguish these states, the slide must retain the limitation rather than present a confident metric with a different meaning.
Record where important limitations appear. They may be in a title, chart caption, footer or speaker note. Decide which must be seen by someone reading independently. A sentence hidden in the speaker's notes does not automatically travel with a slide sent by email. If the conclusion changes meaning without it, put the necessary explanation in the version being sent.
A decision note (in Polish) should also preserve the question, options and unknowns. A model may shorten them to one suggestion that sounds like an approved decision. Distinguish the author's recommendation from the team's decision. The audience should know whether a slide proposes a change, describes a choice or asks for a decision.
Start review with slides that could change a decision
Not all elements carry the same weight. In our example we select three decisive slides: the response count, commentary on the results and the proposed next step. We then check the rest of the presentation for omissions and connections. The sample helps reveal material problems; it is not a reason to skip the remaining claims before sending.
For every sentence, identify its basis in approved material. Do not add a source simply because it concerns a similar topic. If AI added comparative data from the internet, check it separately or remove it. A list of links at the end does not confirm that a particular chart uses the correct table. Retain enough information to reconstruct a number's meaning and calculation.
On a chart, compare values, axis labels, units and group scope. The same correct percentage can be assigned to the wrong category. A title such as "customer opinions" broadens the scope if the data covers only responses from a selected group of participants. Similar distinctions are explored in charts in a brand's visual language (in Polish). The design must preserve the conditions of comparison.
Assess the source from the intended recipient's perspective too. A link may lead to a file only its author can open. Decide whether the recipient needs access, an approved extract or a sufficient explanation of the number's basis in the presentation itself. Do not grant access to an entire folder just to make one reference work. If an important claim cannot be checked without the source, record that gap as a task before sharing.
Check the role of images too. A generator may illustrate an event with a scene that does not come from those workshops. Do not present it as documentation of the real meeting. Its caption, placement and surrounding content should reflect what the image is. Assess a large image with a small explanation at the size at which it will be viewed, not only zoomed in on the author's computer.
In the editable file, make an ordinary change the team will need: a longer title, an extra row or a new value. Check that text fits its box and the label remains with the correct chart. A brand presentation template (in Polish) should support such changes too. A first slide generated for a demonstration does not prove the whole structure is usable.
Distinguish actual editability from the file extension alone. A chart in the presentation may be an image and text may be part of a graphic. Establish what the team will later need to change. If values need updating, the material should include an agreed, maintainable basis for those changes. Otherwise every update will require the entire slide to be generated again and its meaning checked again.

Do not accept corrections solely because an assistant replies "done". Open the changed slide and inspect the specified element. For numbers, compare the title, legend and commentary too. A model may correct a value in one place and leave the old value elsewhere. Keep a change list so the approver knows what needs rereading. After reordering slides, check references to their numbers. "Details on slide seven" may now point to another topic even though the chart and source remain correct.
Aeneia Audit
Describe how your team uses AI when preparing materials. You can discuss the situation with Aeneia Audit and agree whether an audit is appropriate and what it should cover.
Discuss the scope of an AI audit
Review the export in the form in which someone will read it
Microsoft's instructions for saving PowerPoint to PDF show separate paths and options for Windows, macOS and the web version. Do not assume every export route produces identical results. Record the program, version and settings relevant to the material being shared.
Open the finished export outside the editor and go through it at normal reading size. Check page count, order and whether the intended slides are included. A hidden working slide may contain an unapproved proposal; a missing slide with conditions may change how a recommendation is understood. Establish what should be included, then compare the result with that scope.
Compare text and appearance on the three decisive slides. In Mera's example, "percentage of invitees who responded" must remain with the result so readers know what 64 percent means. Look for clipped sentence endings, unreadable characters and displaced labels. Record what is missing from the export instead of saying only "the PDF looks bad". This directs the correction to a particular slide and transition between versions.
Follow the links needed to verify claims. A visible source title and an active link have different roles; establish what the intended use requires. Microsoft documents differences in hyperlink behaviour between the export routes it describes. Check your own output file rather than extending one observation to every program and version.
Review information meant only for the team. A working comment on negotiations or note about an unapproved option may not belong in the partner's package. Record what may be shared and check the actual contents of the resulting file. Do not assume everything invisible during a slideshow has been removed from the editable file. If both formats are shared, review both.
If the basis of a number was in speaker notes, check what happens to it in the chosen export. Distinguish material for a live presentation from material for independent reading. The first may rely partly on the speaker's explanation. The second should include necessary conditions without expecting the reader to guess missing speech. Place those conditions in the material so they remain when the attachment is forwarded.
Check accessibility for the planned way of reading. Can important values be read outside the chart image itself? Does text order reflect the slide's meaning? Does the source have a comprehensible name? These checks require appropriate tools and the actual file, not just viewing thumbnails. State the scope checked; a correct appearance is not complete confirmation that the document is accessible.
When reducing file size for sending, compare the readability of small labels and charts. Compression makes sense if the recipient can still read the information needed for a decision. If data files are attached to the presentation, check the package scope too. A visually matching PDF does not justify sending an entire spreadsheet the recipient does not need.
Consider the format question alongside the choice between PDF and HTML (in Polish). A presentation intended for a public library may have different needs from a one-off talk. If the final material is an active page, review must also consider how AI-generated HTML and SVG are opened (in Polish). The word "presentation" does not settle every consequence of format.
Fix the source of the error and inspect the result again
When something differs, identify the stage. A wrong denominator in the input requires a return to the data. A correct number with a wrong slide label requires a content correction. A clipped qualification in the export requires a layout or file-saving fix. These may look similar to the recipient, but the same instruction will not fix all three.
Keep the previous version and a list of affected locations. After correction, produce a new export from the right presentation and check the changed elements and their surroundings. Do not fix only the PDF while leaving the original error in the file the team will use next week. The next update could restore it without any new decision.
You can discuss your team’s use of AI with Aeneia Audit and agree on the scope before any work begins. State whether findings cover selected slides or the whole material being shared. At Mera, 64 percent must remain the form response rate, however much visual polish was added between the note and the attachment.








