Most people who try using AI to prepare a presentation walk away thinking it worked. Then they stand up to deliver it and realise something went wrong somewhere between the AI’s output and their mouth.
The content is usually fine. The structure is often genuinely better than what they’d have come up with alone. But the language is subtly off — too formal in some spots, too casual in others, full of transitions that read well but feel strange to say aloud. “It’s worth noting that” is a perfectly grammatical sentence. Nobody actually says it to a room of people.
This is the specific failure mode AI introduces into presentation prep, and it doesn’t show up in a document review. It only appears when you’re speaking.
Where AI Actually Earns Its Keep Here
Before getting to the failure mode, be clear about what AI is legitimately good at here — because it’s not nothing.
Structure is the strongest one. If you have a mass of notes, half-formed points, and a rough idea of what you want to say, AI is very good at organising that into a logical sequence. Not because it knows your topic better than you, but because it can apply a clear narrative skeleton — problem, stakes, evidence, implication, ask — without getting attached to your original ordering the way you would.
Argument stress-testing is underrated. Paste your pitch into Claude or ChatGPT and ask: “What’s the most obvious objection to this, and does my current structure address it?” You’ll get something useful more often than not. It’s not the same as a real adversarial reviewer, but it’s faster than waiting two weeks to get time with one.
Cutting is also genuinely useful. Tell the model your time limit and ask it to identify which sections are load-bearing versus filler. It won’t always be right, but it’ll flag the right things to interrogate.
The Specific Way It Goes Wrong
The problem is that AI optimises for text that reads well. Presentations aren’t text — they’re speech with slides as a secondary channel. Those are different enough that the optimisation target is actually wrong.
Spoken language has shorter sentences. It repeats key ideas deliberately, because the audience can’t re-read. It uses false starts and self-corrections sometimes, because those signal real thinking rather than rehearsed delivery. It lands on a word and sits there for a second rather than immediately pivoting to the next clause. None of this is what a model produces when asked to “write a presentation script.”
What you get instead is polished prose that, read aloud, sounds like someone reading polished prose. Audiences can tell. It’s not that the words are wrong — it’s that the rhythm belongs to a different medium.
There’s a second issue: AI can’t know what you find natural to say. “Leverage” might be a word you never use. “Fundamentally” might be one you overuse. The model doesn’t know this, so the script it gives you will have verbal tics that aren’t yours. You’ll either stumble over them or unconsciously swap them out mid-sentence, which breaks your train of thought.
A Better Way to Use It
The fix isn’t to stop using AI for this — it’s to stop treating its output as a draft you edit and start treating it as a scaffold you rebuild.
Concretely: use AI to generate the structure and the key points per section, not the sentences. Take that outline and talk through each section out loud, as if explaining it to someone, and record yourself. Transcribe the recording — Whisper-based apps are fast and accurate enough for this. That transcription is now your actual draft, in your actual voice. Then use AI to edit it for clarity and tighten the logic.
That loop — AI for structure, you for language, AI again for tightening — produces something that sounds like you giving a prepared talk rather than reading someone else’s script.
For pitches specifically, there’s one more pass worth doing: give the model your cleaned-up script and ask it to identify every claim an investor or client could immediately challenge that hasn’t been pre-empted. Then decide, for each one, whether to address it in the talk or prepare the answer for Q&A. This is the kind of adversarial prep that’s awkward to ask a colleague to do and easy to ask a model.
The Part Nobody Mentions: AI Makes It Easy to Over-Prepare the Wrong Thing
There’s a subtler problem. Because AI makes the structural work fast, you can end up with a very polished outline for a talk that isn’t quite the right talk — and by the time you realise it, you’ve already invested enough effort that backtracking feels costly.
This is sunk cost operating under a different name. The speed of AI-assisted prep creates an illusion of progress that substitutes for the harder earlier question: is this the right framing for this specific audience on this specific occasion?
That question doesn’t have a shortcut. You have to answer it before you hand anything to a model, or the model just helps you build the wrong thing faster.
One prompt that helps: before generating anything, write two or three sentences describing who is in the room, what they already believe about your topic, and what you need them to do or think differently after. Give the model that context first. The outputs shift noticeably when the model has a specific audience rather than a generic one.
What to Actually Ask For
A few prompt patterns that work better than “help me write a presentation on X”:
- “Here are my raw notes on this topic. Give me three possible structural approaches and explain the trade-off of each.” — Forces you to make a deliberate choice rather than accepting the first structure the model picks.
- “Here’s my current outline. What would someone skeptical of my main claim say after hearing section two?” — Useful for finding the argument gaps before they find you.
- “Here’s a transcript of me talking through this section. Edit it for clarity without changing the sentence rhythm or vocabulary level.” — This constraint matters. Without it, the model will “improve” your language into something more formal and less yours.
The vocabulary constraint in that last one is doing real work. Adding it preserves the spoken quality in a way that “keep it natural” doesn’t, because “natural” is underspecified and the model fills it in with its own defaults — which skew written, not spoken.
If you’ve got a talk coming up, try recording a rough verbal run-through before you touch any AI tool. Use that recording as your input rather than a blank prompt. It’s a slower start, but you’ll end up with something you can actually deliver — which is the only thing that matters once you’re standing in front of the room.
