How to Make a Presentation With AI: Structure First, Slides Last

The promise is hard to resist: describe your topic, get a finished deck. And the tools deliver something — slides appear, with headings and stock-photo energy. Then you read them and realize the deck says nothing you would stand behind in a room. Learning how to make a presentation with AI is mostly learning what to do before any slide tool opens.

We build decks weekly — client proposals, strategy readouts, training materials — and AI is in every one of them now. But it enters at specific steps, in a specific order. Done backwards (slides first, thinking later), AI makes presentations faster and worse. Done in order, it makes them faster and noticeably better, because the time it saves goes back into the argument.

The order that works: argument, structure, text, slides

A presentation is an argument delivered in sequence. The slides are packaging. Every failed AI deck we have seen failed at the argument stage — the tool generated packaging for a point that was never defined.

So step one happens in a chat, not a slide tool:

I am preparing a presentation. Before writing anything, interview me:
ask me one question at a time until you can state (1) who exactly is in
the room, (2) what they believe before I speak, (3) the one sentence I
need them to agree with after, and (4) what decision or action follows
if they agree. Then propose the argument as a chain of 4–6 claims,
each one sentence, each building on the previous.

Topic: [your topic]. Audience: [who]. Occasion: [when/why].

Letting the model interview you is the single highest-value move here. Typing a brief from scratch, you forget what the model needs; answering questions, you surface it naturally — constraints, politics in the room, the objection you are afraid of. Ten minutes of this produces a better foundation than an hour of staring at a blank slide.

From argument to slide plan

Once the claim chain stands, convert it:

Turn this argument into a slide plan. One claim per slide. For each slide:
a headline that states the claim as a full sentence (not a topic label),
what evidence or visual supports it, and one line of speaker notes.
Max 12 slides. Mark any slide where I still owe you data.

The full-sentence headline rule is old presentation craft and it disciplines the model wonderfully. «Q3 results» is a topic; «Q3 margin recovered because we cut the two unprofitable channels» is a claim. If the model cannot write the sentence headline, the slide has no point and should not exist — the plan just told you something.

Note the last instruction: slides where data is owed. The model will happily fabricate a supporting statistic if you let it. Numbers come from you, with sources; the model arranges them. We keep that boundary everywhere, for reasons covered in how to fact-check AI answers.

Where each tool fits

Stage	Tool	Why
Argument and structure	ChatGPT or Claude in chat	The interview and claim-chain work is pure reasoning — chat models are strongest here
Slide text and speaker notes	Same chat, same thread	Context from the interview carries over; a fresh tool starts from zero
Visual draft	Gamma or PowerPoint/Google Slides AI features	They turn a structured outline into designed slides quickly
Final design	Your hands, in the deck tool	Brand, emphasis, what to cut — judgment calls the generators consistently miss

Gamma deserves its reputation for the visual-draft stage: paste a structured outline and you get a respectable designed draft in minutes. What it cannot do is rescue an unstructured one — garbage outline, prettier garbage. The slide generators built into PowerPoint and Google Slides behave the same way: quality in, quality out.

Charts are their own lane. If the data lives in a spreadsheet, do the analysis there first — we described that workflow separately — and bring finished numbers to the deck. Asking a slide generator to «visualize our growth» without handing it the figures is how fictional charts are born.

The pass most people skip: making it survive the room

A deck that reads well and a deck that presents well are different artifacts. Two prompts close the gap.

The hostile read: «You are the most skeptical person in this meeting — [describe them: the CFO who hates vendor pitches]. Read this slide plan and attack it: what is unsupported, what is vague, where would you interrupt?» Fix what the attack finds before the real skeptic finds it.

The cut pass: «This deck must work in 10 minutes. Which slides merge, which move to an appendix, which die?» Models are genuinely good at this because they have no emotional attachment to your slides. You do; that is why decks grow.

Where this breaks

Generated images for business decks remain a lottery — fine for a mood, risky for anything technical, and never acceptable for charts of real data. Brand compliance is manual: no generator reliably holds your fonts, colors and logo rules, so budget a formatting pass. Speaker dependence is the quiet failure: a deck the model wrote and you skimmed will betray you in the Q&A, because the room always asks about the one slide you did not really own. The fix is the process above — if you built the argument in the interview step, every slide is yours by construction.

And confidential material follows the same rule as every other upload: strategy decks full of unreleased numbers do not belong in consumer tools — what not to upload, in detail.

In the practicum «AI system in two weeks», presentations are part of the third session — documents, data and visuals: the interview technique live, the claim-chain method on a real deck of yours, and the chart-honesty rules that keep generated slides out of trouble. The program and dates are here: AI system in two weeks.