8 Mistakes People Make with AI Presentations
Most bad AI decks fail in the same handful of ways, which means most bad AI decks are preventable with attention rather than effort. None of the fixes below costs extra time; each one redirects a few minutes from the wrong place to the right one. Each mistake gets its symptom — what it looks like in the finished deck — its cause, and the fix in practice, because knowing the fix and knowing when to apply it are different skills.
Read the list once and you'll start catching these in other people's decks, which is faster than catching them in yours.
1. Prompting without a brief
Symptom: the deck is competent and belongs to no one — it would survive any company swapping the logo. Cause: "make me a presentation about X" gives the AI a topic when it needed a room; with no audience and no goal, it aims at the average of everything ever written about X, and the average of everything is nothing.
Fix: two lines before any tool — who is the audience and what should they do or believe afterwards. Even one line helps: "for the board, who think churn is a support problem" will transform a deck by itself. The 30-minute workflow starts here for a reason, and its brief template has six lines that take four minutes. The micro-habit that makes this stick: write the goal sentence before opening any AI tool, on paper if necessary. If you can't write it, you're not ready to build slides — and no prompt will fix that, because it's not a prompting problem.
2. Accepting the first output
Symptom: the deck feels 80% right and ships at 80% — fine structure, dead titles, one slide that makes a point two other slides already made. Cause: draft one genuinely is decent, and nothing in the process creates the pressure to push. Unlike a blank page, an 80% draft has no natural moment where you stop and fix it; the default is to polish the commas and ship.
The fix: budget three pushbacks per deck and name them before generating, so they're scheduled rather than optional: merge the overlaps, reorder so the strongest number leads, sharpen the vaguest slide. The formula article has the full follow-up list; these three do most of the work in ninety seconds each. What makes the pushes effective is that they're moves — merge, cut, reorder, strengthen — not rewrites. "Make it better" changes nothing; "merge slides 4 and 5" changes the deck. The quality difference between a shipped draft one and a pushed draft two is larger than the difference between using AI and not using it at all.
3. Letting AI pick the slide count
Symptom: sixteen slides for a ten-minute slot, four of them "overview" slides, the recommendation appearing at slide twelve. Cause: unconstrained, AI optimizes for coverage — every facet of the topic gets its slide — and coverage is the opposite of what a slot needs, which is compression toward a decision.
Fix: state the count and the minutes in the prompt: "8 slides for a 12-minute slot; cut anything that doesn't survive." The minutes matter as much as the count, because they force prioritization rather than formatting — a 20-slide deck is defensible for a 40-minute training and absurd for a standup. When the AI cuts under a limit, it makes judgment calls; watch what it cuts, because the disagreements between its priorities and yours are exactly the slides you need to think about. A limit also improves the writing: slides built under constraint claim; slides built for coverage describe. If the deck still arrives over length, the follow-up is mechanical: "cut to N slides and tell me what you removed and why" — the "why" is where you catch it quietly deleting your ask.
4. Presenting numbers you can't source
Symptom: a specific, plausible statistic — "studies show 73% of teams…" — that nobody in the room can place, including you. Cause: AI fills factual silence with invention, fluently, formatted exactly like the real numbers beside it. Nothing in the output marks it; the confidence is identical.
Fix: verify every number against a real source, or replace it with [NEED DATA] and fill it yourself before the deck moves. The human pass has the procedure — claim inventory, triage the load-bearing five, soften or cut the rest. This is the only mistake on the list that can genuinely hurt you, because the other seven make a deck mediocre and this one makes a presenter untrustworthy: get caught once inventing a number, and every real number you present afterwards gets discounted by the people who caught it. The asymmetry is brutal — a deck of invented bullets costs you a meeting; one invented statistic costs you the room. The two-minute check per number is the cheapest insurance in professional presenting. And when someone asks "where's that from?" mid-meeting, "the AI said so" is not an answer; "it's from our CRM export, page 4 of the report" is the whole point of the workflow.
5. Staying generic
Symptom: "a leading company in the industry", "users may struggle with", "significant efficiency gains" — a deck built entirely of sentences that are true in every company and useful in none. Cause: the AI doesn't know your customers, your Q3, or why the last launch underperformed, and it will fill that gap with the industry average rather than leave a hole.
Fix: the specificity pass from the editing article — every generic example gets replaced with a real customer, project, or screenshot, or gets deleted. There is no third option; "sort of specific" reads as vague with extra steps. Ten minutes of replacements turns a template-scented deck into one that could only have come from your team, and it's the highest-value ten minutes in the edit. The deeper point: generic content isn't just boring, it's unconvincing, because specificity is how a room verifies that you know what you're talking about. "We reduced onboarding time significantly" is a claim; "onboarding went from 11 days to 4" is evidence, and rooms decide on evidence. If you don't have the real example, that's not a deck problem — it's a homework problem, and now you know the homework.
6. Discovering export problems at the deadline
Symptom: fonts substituted, the corporate master missing, images compressed into mush — discovered the night before, in the room where the deck must run. Cause: the deck was built in a web tool, the deliverable is a .pptx in the company template, and "export" was assumed to work because the button exists. Export fidelity varies between tools more than any feature list suggests, and it fails at the boundary — fonts, masters, notes, image compression — where nobody tests until it's too late.
Fix: the day-one export test from the tools article: a two-slide test deck, exported to the format you actually deliver, opened in the destination environment — company template, client machine, conference laptop. Check five things: fonts survived, the master applied, images didn't degrade, speaker notes are still attached, nothing reflowed into overlap. Ten minutes, once per tool, ever. The test also tells you which tool to standardize on before the team builds a habit on one that can't deliver its output. This is the most boring mistake on the list and the only one that produces a purely wasted evening.
7. Skipping the outline
Symptom: an hour spent dragging slides, re-typing titles, and re-balancing layouts — the deck is structurally wrong and the tool is the wrong place to learn that. Cause: the generated deck looks "almost done", which locks structural decisions into the most expensive editing surface: every merge costs minutes plus layout cleanup, every reorder means re-flowing three slides.
Fix: one outline round in a chat first — brief in, slide-by-slide outline out, two pushbacks applied — then generate slides from the finished outline. The outline is the cheapest editing surface you will ever have for this deck: merges cost five seconds, reorders cost nothing, and the title-read test tells you whether the argument works before a single slide exists. Skipping it doesn't save the ten minutes; it moves the same structural work into a tool that charges fifty. The tell that you skipped it: you know your deck's problem is the order, and you're fixing it by dragging thumbnails with a rising sense of futility.
8. Using AI to sound smarter
Symptom: "leverage synergies across stakeholder touchpoints to unlock seamless value delivery" — a deck that reads like consulting-deck parody, and rooms trust it less for reasons they can't articulate. Cause: AI's training rewards impressive prose, and left alone it inflates: plain verbs become abstract nouns, claims grow qualifiers, everything becomes "robust" and "end-to-end". Sounding smart is what the model optimizes for by default; being clear has to be requested.
Fix: two constraints in every prompt — "plain language; explain any term a customer wouldn't use" and a banned-phrase list of your office's favorite inflated words (synergy, seamless, leverage, robust). Then the rewrite test on any slide that survived them: would you say this sentence out loud, to this room, in this tone? Spoken language is the fluency test for written presentation language, because the room hears you talk normally and then reads the slide; the gap registers as inauthenticity. The deeper irony worth remembering: the whole point of AI-assisted decks is clarity at speed. Using the machine to obscure — even accidentally — is paying for the tool and eating the cost it was supposed to remove.
Why the same eight repeat across whole teams
These mistakes look individual, but they reproduce at team scale through two mechanisms. The first is demo-driven adoption: someone sees a beautiful generated deck at a conference or in a vendor demo, and the team adopts the tool's default workflow — one-shot generation — without anyone deciding to. The defaults spread; the judgment doesn't. The second is deadline asymmetry: the deck is due Thursday, the fixes cost minutes, and skipping them costs nothing today — the cost arrives later, in the room, attributed to "the deck" rather than to the skipped step. Both mechanisms are cultural, and both have the same antidote: make the right sequence the normal sequence, visibly, a few times, until it's just how decks get made here.
Team norms that prevent all eight
Four norms cover the list, and none requires management pressure to install — just one person running them visibly. The brief is required: decks start from a filled brief file; after a month nobody presents without one, because the improvement is visible to everyone, including the presenter. The numbers policy: any statistic in a deck carries its source in the notes — a one-line norm that kills mistake 4 at the source and gives every reviewer an easy check. The export standard: the team's generator is whatever passed the export test, and the two-slide test file lives in the shared drive for anyone adding a new tool. The review ritual: decks get a named reviewer and the three questions from the editing article — fifteen minutes of reviewer time that catches mistakes 2, 4, and 5 for everyone. Teams that run these four norms get the same quality floor regardless of who's presenting, which is the actual goal: not one person's good decks, but nobody's bad ones.
For reviewers: the five questions that catch the list
If you're reviewing someone's AI-assisted deck, five questions cover all eight mistakes. "Who is this for, and what should they do after?" — catches 1 and 3, because a vague answer means the deck meanders. "Which of these numbers can you source right now?" — catches 4. "Which slide did you push back on hardest?" — catches 2; "none" means draft one shipped. "What did you replace the generic examples with?" — catches 5. "Did the export survive the actual room?" — catches 6. Ask them as a reviewer, not an interrogator: a good deck answers all five in a minute, and the silence on any one of them shows exactly where the missing minutes go before the meeting, which is the most useful feedback a draft can receive.
Three things that look like mistakes but aren't
Before the norms calcify into dogma, three exceptions worth defending. A long deck is not mistake 3 when the slot is long: a 45-slide training for a 90-minute workshop is correct sizing, not padding — the mistake is only count-against-slot mismatch. An agenda slide is not always a crime: in a 40-slide deck presented across two hours, or a deck that will circulate async, the agenda is navigation, and banning it outright applies a conference-talk rule to a reference document. Generated imagery is not automatically mistake 5: an abstract background texture or a conceptual illustration that claims nothing is fine — the mistake is imagery pretending to be your customers, your product, or your data. The pattern in all three: the rules on this list exist to serve the room, and when the room is different, the rules bend. Mistakes are mismatches, not checkboxes — the audit question is never "did I follow all eight rules" but "does this deck fit this room."
The cost accounting of a bad deck
It helps to price the mistakes, because the fixes are minutes and the costs are meetings. An unsourced number that gets challenged: the decision stalls, the follow-up meeting costs an hour of eight people, and a trust discount applies to your next deck — call it two days of calendar across the room. Generic content that fails to move the room: the decision defers, another hour-long meeting gets scheduled, and the presenter spends a weekend redoing what three pushbacks would have fixed. Export failure discovered live: twenty minutes of a full room watching someone troubleshoot fonts, which is a meeting tax that everyone present pays at once. Now the fix costs, for comparison: a two-line brief, four minutes; three pushbacks, three minutes; the export test, ten minutes once per tool; the number check, two minutes per load-bearing figure. The arithmetic is lopsided in every case — which is exactly why these are attention mistakes, and why the sixty-second audit below pays for itself the first time it catches anything at all.
When the deck deserves to fail
One honest caveat to close the list: sometimes a bad deck is the symptom of a decision that hasn't been made, and polishing it is the wrong response entirely. If the brief can't be written — no audience, no goal, nothing you actually recommend — the deck isn't ready, and no workflow will save it; what's missing is a decision, and the fastest path to the deck is deciding, not prompting. Teams sometimes generate decks precisely to avoid this moment: the deck becomes a way to defer the recommendation by presenting "options" without a position. AI makes that escape cheaper — it will happily produce the options deck — but the room always sends it back with the same question: "which one do you recommend?" The brief line forces the position early, while changing it is still free. That's the deepest reason the brief comes first in every workflow on this site: not prompt quality — forcing the one decision that makes all the other decisions possible.
The one mistake behind the seven others
Notice what the list has in common: seven of the eight are one mistake wearing different clothes — accepting machine output without a human decision at the point where the decision was needed. No brief means the audience decision wasn't made. No pushbacks means the structure decision wasn't made. Unverified numbers mean the sourcing decision wasn't made; generic content means the specificity decisions weren't made. The AI does not make these decisions, and it never will — it produces options at the speed of light and waits. The workflows on this site are, in the end, just schedules for making those decisions cheaply, in the cheapest editing surface available. Once you see the list that way, catching your own version of any mistake becomes a one-second check: which decision was skipped here? Make it, and the deck moves.
The 60-second self-audit
Before any AI-assisted deck ships, five questions, one minute total:
- Could someone else have written this deck? (If yes: mistake 1 or 5.)
- Is every number one I could source out loud? (Mistake 4.)
- Did this deck get at least three pushbacks? (Mistake 2.)
- Would the export survive the actual room it will run in? (Mistake 6.)
- Read the titles only: is it an argument or a table of contents? (Mistakes 3 and 7.)
Any "no" points at a specific fix, and every fix is minutes. That's the pattern across all eight: they're not effort mistakes, they're attention mistakes — each one is a few minutes spent at the moment it matters instead of an evening spent repairing what skipped it. If you remember only one of them, remember #4: it's the only one that can actually hurt you. Print the list, tape it near the monitor, and let everyone else keep their agenda slides.