Lately I've been watching a lot of people use AI to build slides. Maybe it's because I'm spending my days building Huanjian, but I've started seeing patterns. Most of them are mistakes I've made myself.
Not the catastrophic kind. Just the small, recurring friction points that turn a decent deck into a forgettable one. Here are five of them, grouped loosely.
The garbage in problem
The most common mistake is treating AI like a mind reader. Someone on my team — let's call him James — typed "make a pitch deck for my startup" into a tool last month. He got back 28 slides about market sizing, mission statements, and generic org charts. It was useless.
The AI doesn't know you lost your biggest customer last quarter. It doesn't know your competitor just pivoted. It only knows the average of all the pitch decks it scraped from the internet. If you feed it a vague prompt, you will get an average deck. Not bad. Just average. And average in a pitch meeting is indistinguishable from bad.
The fix is brutally specific prompting. Don't ask for a deck. Ask for an argument. "I need 5 slides that convince a seed-stage investor that our logistics startup is worth a meeting because we've already signed 3 beta customers." That's a different request entirely.
Design by roulette
Another thing I see people do is treat AI-generated design as sacred. The tool picks a template, generates some icons, and the user accepts it all without question. You end up with the swooshes and drop shadows. The stock photo of a conference room. The clipart arrow that doesn't point at anything relevant.
While building Huanjian we realized that design isn't about making things pretty. It's about making things disappear. If the audience is looking at the template, they're not listening to you. The best presentations I've seen from our beta users are visually boring. Clean text. One image that actually means something. No decoration.
The mistake is letting the AI decide what "looks professional." Professional is usually just clean. The tool should help you remove clutter, not add more.
Trusting the hallucination
This is the one that scares me. I watched a demo where the AI generated a slide claiming "85% of customers prefer subscription models." The presenter left it in. Where did that statistic come from? Nobody checked. The AI just made it up, and it sounded credible.
AI hallucinates with confidence. It will invent case studies. It will misattribute quotes. It will generate plausible-sounding data that is completely wrong. If you don't verify the facts, you are not presenting. You are performing fiction.
The fix is not complicated: edit. Treat every AI output as a first draft from an intern who is very confident and very wrong. Check the numbers. Read the quotes. Ask yourself "does this pass the sniff test?"
I'm still figuring out the best balance myself. Some days I let the AI run wild because I'm lazy. Some days I overcorrect and rewrite everything. The tools are getting better, but the fundamental job hasn't changed. You have to know what you think before you ask the machine to dress it up. That's the hard part. That's the part I'm still learning.