What you’ll be able to do
You’ll go from a rough brief, the kind you scribble on a sticky note or half-remember from a meeting, to an actual deck with a story running through it. Not thirty slides of bullet points nobody reads. A deck that says something, in the right order, with evidence to back it up.
That’s the goal here. Not pretty. Clear.
Before you start
Right, before you open anything, you need three things sorted in your head. What are you actually trying to say. Who’s in the room. And what do you want them to do differently once you’ve finished talking.
If you can’t answer those three, no AI tool is going to save you. It’ll just help you produce a confident-looking deck about nothing in particular, which is worse than no deck at all, in a way, because everyone nods along and then forgets it by lunch.
I’d also say: know roughly how long you’re speaking for. Ten minutes needs ten slides, not thirty. That’s not a rule I invented for effect, it’s just what actually works when you’re stood in front of people.
Get set up
There are a few genuinely useful tools for this right now, and they split by what you’re starting from.
If you’re starting from a document, a report, some notes, a Word file full of half-formed ideas, then Copilot in PowerPoint is built for that. It drafts a deck straight from a Word document, but you need a Microsoft 365 Copilot licence, so check that before you get excited.
If you’re on Google Workspace, Gemini in Google Slides does something similar: generates slides and images, on eligible Workspace plans. Again, check eligibility first, it’s easy to assume you’ve got access when you don’t.
If you’re starting from nothing, just a vague feeling and a deadline, then Gamma is usually the first stop. You give it a prompt or a document and it builds a full deck from that, structure and design included. It’s quick, and it’s a genuinely good way to get from zero to a first draft fast.
Canva’s Magic Design and its presentation tools are worth knowing too, especially if design matters more to you than structure, you get a lot of visual polish there. And Beautiful.ai handles auto-layout well, so your slides don’t end up looking like a ransom note even if you’re not a designer.
Here’s the thing though, and it’s the bit people skip: none of these tools are good at working out your story for you. They’re good at laying it out once you’ve worked it out. So the actual first move, regardless of which tool you end up using, is to write the story in ChatGPT or Claude before you touch any slide-making tool at all.
Try it yourself
Start with the story, not the slides. Here’s a prompt that forces the structure before any design happens.
I need to give a presentation about [topic] to [audience].
The goal is for them to [specific outcome, e.g. approve the budget,
understand the new process, back the proposal].
Help me write the story first, one line per slide, in this format:
- Point: what this slide argues
- Evidence: what backs it up
- So what: why the audience should care
Aim for 8-10 slides total. Don't write slide text yet, just the story.
Once you’ve got that back, read it properly. Does it actually flow, or does it just list facts in a row? If a slide’s “so what” is weak, that’s the one to cut, not pad out.
When the story holds together, then you turn it into slides. If you’re using Gamma, Copilot, or Gemini, you can often paste the story straight in as your brief. If you want ChatGPT or Claude to draft the actual slide text first (useful if you want to tweak wording before it hits a design tool), try this:
Using the story below, write the actual slide content for each slide:
a short headline, then no more than 3 bullet points of supporting text.
Keep language plain. No jargon. No filler words like "leverage" or
"synergy". Flag any slide where you think the evidence is thin.
[paste your one-line-per-slide story here]
That last instruction matters, get the AI to flag weak evidence itself. It’s often better at spotting a gap than you are, because you already know what you meant to say, whereas the AI’s only working from what’s on the page.
Last job: images. Don’t just bung in stock photos because the tool suggested them. If you need a specific image, a diagram, a chart concept, a mood shot, go through the create-an-image guide on this site properly rather than accepting whatever the deck tool auto-generates. It’ll usually be more on-brief. But if you’re happy with the tool’s own image generation (Gemini in Slides and Gamma both do this), a decent prompt looks like this:
Generate an image for slide [number] that shows [concrete description,
e.g. "a simple flowchart with three boxes: request, review, approval"].
Style: clean, minimal, no text in the image itself, works on a white
background.
Keep the description concrete. “Something professional” gets you nothing useful. “A flowchart with three boxes” gets you an image you can actually use.
Check the result
Before anyone else sees this deck, run a few checks yourself.
First, the story test. Read just the headlines, slide by slide, nothing else. Does it make sense as a story on its own? If you need the bullet points to explain the headline, the headline’s not doing its job.
Second, the so-what test. For every slide, ask: if I deleted this slide, would the argument fall apart? If not, cut it. Ten slides that matter beats thirty slides where half are padding.
Third, check every number. Every single one. AI tools can get figures wrong, misquote a source, or just make something up that sounds plausible. If a slide says “revenue grew 40%”, go find where that number actually came from and check it matches. Same with any chart, don’t trust that the chart shape reflects the data unless you’ve checked it against the source yourself.
Fourth, and this one’s genuinely useful: rehearse with the assistant as a sceptical audience. Try something like asking what a sceptical CFO would ask on slide four. It’ll often surface the exact gap a real person in the room would spot, which gives you a chance to fix it before it’s a problem live.
If it doesn’t work
Sometimes what comes back is just bland. Generic headlines, bullet points that repeat the same idea in different words, or a structure that technically follows your brief but says nothing. When that happens, go back to the story step. The slides are only ever as good as the one-line story behind them, so if the output’s weak, the story was probably weak or vague, and no amount of asking the tool to “make it punchier” will fix that.
If the tool keeps inventing numbers or facts you didn’t give it, stop and be explicit: tell it not to add any statistic, figure, or claim that you haven’t provided yourself. Some tools do this more than others, so treat every number in a first draft as unverified until you’ve checked it.
Now, the bit that actually matters more than any of the above: don’t upload confidential decks, real client names, financial figures, unreleased product details, anything under an NDA, to a tool whose data terms you haven’t actually read. Not skimmed. Read. Some tools use uploaded content to improve their models unless you’ve turned that off, some don’t retain it at all, and the settings genuinely differ between the free and paid tiers of the same tool. If you’re not sure, don’t upload it, and instead work with placeholder names and numbers until you’re ready to swap in the real ones by hand.
Keep exploring
If you’re pulling data into a deck, “How to get AI to make a spreadsheet” covers getting the numbers sorted before they end up on a slide. If your brief started life as a long report, “How to get AI to summarise a document” will get you to a usable story faster than reading the whole thing twice. And for the images bit above, “How to get AI to create an image” goes into proper detail on prompting for visuals that actually match what you need.
Sources and review notes
- Gamma: https://gamma.app/ · Beautiful.ai: https://www.beautiful.ai/ · Canva: https://www.canva.com/
- Microsoft 365 Copilot (PowerPoint): https://support.microsoft.com/en-us/microsoft-365-copilot/
- Gemini in Google Workspace: https://workspace.google.com/products/gemini/
Review date: 15 September 2026 — every source above was opened and checked against the text on that date.