Design Mistakes AI Still Makes (and Quick Fixes)

October 29, 2026 · 14 min read

AI slide generators get structure right and design mediocre, and the pattern is stable across tools: they summarize when they should decide, decorate when they should emphasize, and fill when they should leave space. The good news is that the gap closes with about 15 minutes of manual fixes, and the fixes are formulaic — the same five every time. They're listed in order of visual impact: fix the first one and the deck already looks half-human; fix all five and nobody in the room will guess where the deck came from.

1. Walls of bullets

AI defaults to six or eight bullets of fifteen-plus words per slide, because a language model summarizes and a summary of a topic is a list. The fix has three moves. First, move detail into speaker notes: every bullet that's really a sentence of explanation belongs in the notes, which is where you'll use it anyway. Second, cut what remains to at most five bullets of eight words or fewer — the constraint isn't arbitrary; eight words forces the AI's text into claims, and claims are what slides are for. Third, reshape: a bullet list describing a sequence becomes a table with two or three rows; a list describing one thing becomes that one thing in big type. A slide that still needs more than that is two slides, and you already know which line splits it.

What this looks like in practice: a slide titled "Channel Performance Analysis" carrying eight bullets about trends, costs, and conversions becomes a slide titled "Purchased lists cost 3× per closed deal" with one table of four numbers — and the seven minutes of prose move to the notes where they become your talking points.

2. No visual hierarchy

When everything on a slide is the same size, nothing gets read; the eye bounces and lands on nothing. Every slide needs one dominant element — the number, the statement, the chart, or the image — and it should be two to three times larger than the surrounding text. The cheapest version of this fix: promote the takeaway title to the biggest thing on the slide and shrink everything else. Twenty-four-point title, forty-eight-point key number, fourteen-point support text is a hierarchy that works from the back row.

Whitespace is part of hierarchy, and AI is terrible at it — it fills every box it draws. Deleting decorative elements (a second image, three icons, a gradient band) does more for hierarchy than adding emphasis does. When in doubt: one element big, everything else smaller, and enough empty space that the big element has room to be big. Ten minutes of hierarchy beats an hour of picking fonts, because fonts were never the problem.

3. Generic imagery

AI-generated stock art of smiling teams pointing at whiteboards marks a deck as machine-made instantly, the way stock photos marked corporate sites in 2010. The audience has seen that exact image in four other decks this quarter. The fix has a strict order: delete first, replace second — an empty slide is easier to fix than a cluttered one, and half the generated images deserve nothing in their place. When you do replace, the ranking is: your own screenshots, real photos of your product or team, a big number on a plain background, nothing at all. A slide with one real screenshot beats a slide with three generated illustrations every time, because real is the thing the deck is for.

The one exception: clearly abstract, conceptual imagery — a subtle texture behind a title, a stylized diagram shape — can stay if it stays out of the way. The rule of thumb is that imagery should never pretend to be information. The moment a generated image is "showing" your customers or your data, it's lying, and someone in the room knows it.

4. Inconsistent spacing and alignment

Generated decks drift: the title sits at 40 points here and 34 there, margins wander between slides, body text starts at a different height on every slide. Individually invisible; collectively it reads as "assembled", which is exactly what it was. The fix is one operation, not twenty: apply your template or slide master after generating, not before — templates fight the generator during creation, but masters clean up after it. Two minutes in the master view fixes the whole deck at once.

Three alignment rules while you're there. Left-align body text — centered text reads measurably slower, and centered bullet lists read like a wedding invitation. Keep the title in the same position on every slide; the audience's eye learns the pattern by slide 3 and stops hunting. Give all content the same margins, which in the master view is literally setting two numbers once.

5. Charts and contrast

AI picks chart types decoratively rather than by the question being asked, so the fix starts with the question. Match the chart to what the slide claims:

The slide's questionThe right chartWhat AI defaults to
How does it change over time?Line chart, one line per seriesBar chart of every period, unlabeled
Which item is bigger?Bar chart, sorted, direct labelsPie chart of eight slices
What is it made of?Stacked bar; donut only for ≤3 slicesExploded 3D pie
Do these relate?Scatter plotTwo lines on twin axes

Then the labels: label lines and bars directly instead of using legends — the reader shouldn't hold a color key in working memory. And refuse 3D effects on principle, because they distort the values they claim to show; a 3D pie chart's front slice always looks bigger than its number.

Contrast is the second half of this fix, and the failure shows up only in the room: thin light fonts on gradient backgrounds, mid-gray text on white, colors that survive a laptop screen and die on a projector. The test is the back row — dark text on light backgrounds, nothing smaller than 18 points for anything meant to be read, and the deck must survive a bad projector, because the bad projector is the one you'll get. If you present in the room where the deck will actually be shown, spend two minutes projecting it; if you can't, assume the worst screen in the company.

Two before-and-after walkthroughs

Concrete beats abstract, so here are two slides from the workflow article's deck, before and after. First, the wall of bullets: "Channel Performance Analysis" carried eight bullets — trends, costs, conversion rates, three hedged sentences, and a conclusion hiding at position six. After: the title promoted to the takeaway ("Purchased lists cost 3× per closed deal"), the eight bullets became a four-row table (channel / volume / conversion / cost per deal), one bullet survived as a single line under the table, and the rest moved to the notes. Same information, ten seconds to read, and the argument now states itself.

Second, the chart fix: the draft had a pie chart of "lead sources" — six slices, three under 8%, a legend, and 3D. The slide's actual question was "where does volume come from versus where do deals come from," so the fix was a two-bar comparison: volume share next to closed-deal share, direct labels, the mismatch between the two bars doing the arguing. The pie was deleted, and the slide got its first real hierarchy — the gap between the bars is the point, and now the design says so.

Numbers as design

The cheapest "design" in a business deck isn't visual at all: it's one big number. A slide whose dominant element is "$412 per closed deal — triple target" needs no imagery, no icons, and no template fight; the number is the design, and formatting does the rest. Three rules make numbers carry slides: round honestly ("$2.1M" on the slide, "$2,087,431" in the appendix — precision belongs where people check it), state the comparison in the same breath as the value ("triple target," "leads up 22%, cost up 40%"), and never show a bare number without its baseline, because a number without a reference point is trivia and rooms don't decide on trivia. When the deck has one slide per argument and each argument leans on one number, the big-number pattern does half the hierarchy work automatically.

Typography in two minutes

Font decisions consume disproportionate time for their visual return, so here is the whole subject at working-deck scale. Two fonts maximum: one for titles, one for everything else — or zero decisions, by using whatever the corporate template ships with. The size scale that works on a projector: titles 28–36 points, body text 18–20, sources and footnotes 14, nothing smaller ever, because the back row exists. Bold is hierarchy, not emphasis-by-volume: bold the number in a sentence, not the sentence. And the one thing typography cannot fix is text volume — 200 words in 14-point type is still 200 words. Shrink the words first; the font never fixes them. If a deck seems to need a typography article, it actually needs the bullets fix from section one.

Slide titles as the second design system

Typography, spacing, and charts are the visible design system; the titles are the invisible one. A deck where every title sits in the same position, at the same size, and reads as a chain of takeaways looks designed even in a plain template — because design, in a work deck, is mostly rhythm and consistency. Two habits make titles carry their weight: keep them to one or two lines (a three-line title is a sentence that hasn't been edited), and make them parallel in voice — all statements, or all questions, but not both alternating. When the title system works, most decoration can be deleted from the slides, because the titles were doing the structural design all along. This is also why the title-read test doubles as a design check: a deck whose titles read as an argument feels coherent at a glance, and coherence is what "designed" means in a conference room.

Four layout patterns that cover most slides

You don't need layout creativity; you need four patterns applied without apology. Statement slide: one sentence, big, centered in space — for the slide that changes the room's mind; it's usually underused because it feels wasteful, which is exactly why it works when it appears. Number slide: one figure at 60+ points, a one-line caption beneath, nothing else — for every argument that leans on a figure. Table slide: three to five rows, direct labels, generous row spacing — for comparisons; tables are the most honest pattern because they show what you're not claiming. Two-column proof: claim on the left, evidence (screenshot, chart, quote) on the right — the workhorse for "here is why." Map every slide in the deck to one of the four and the deck looks designed, because pattern consistency reads as intentionality. The generated decks that look "off" are usually five patterns deep in one deck; the fix is consolidation, not more design skill.

The one-image slide

When a slide genuinely benefits from an image — your product, your team, the customer's site — it works under one layout: image dominant, one caption line, nothing else. The failure mode is the collage: four images at equal size, each cropped oddly, none dominant, all fighting the text. Choose the single image that proves the slide's claim and let it fill 60% of the slide; the other candidates are either other slides or deleted. Screenshots get one extra rule: crop to the region that matters and zoom — a full-app screenshot at slide scale is unreadable, and an unreadable screenshot is decoration pretending to be evidence. If it can't be cropped to something readable, it doesn't belong on a slide; link the live version, or move it to the appendix where people can lean in.

The decoration inventory

Generated decks arrive with a standard set of decorations, and each earns or loses its place by one question: does it carry meaning? Arrows: keep when they show sequence or causation, delete when they point at nothing. Icons: keep when they aid scanning of a repeated structure — four sections, four icons — delete when they're clip-art standing next to text. Progress bars and page numbers: keep page numbers, delete the fancy progress trackers nobody perceives. Divider slides: one style, used only between major sections of a 20-plus-slide deck. Boxes and shapes around text: delete unless the box groups things that belong together — most generated boxes group nothing, they just fence. The inventory pass takes three minutes with that single rule, and the deck loses maybe fifteen elements, after which the real content finally has room. Decoration doesn't add meaning; it dilutes the meaning that was already there.

Dark decks, print, and handouts

Two delivery modes change the rules, and generated decks don't notice. On a projector or video call: dark text on a light background wins, because cheap projectors crush dark themes into mud and video compression eats subtle grays. On paper: contrast is easier, but the real difference is purpose — a handout is a document, not a talk, so it needs the sentences the live deck deliberately dropped. The classic mistake is exporting the talk deck as the leave-behind: slides with three words and a big number read as cryptic in an inbox. If a leave-behind is expected, generate it as its own asset — "expand this deck into a two-page self-standing summary" — which takes two minutes and inherits the master's numbers. Two artifacts, two designs, one source: it's the repurposing rule applied one level down.

The projector test, in the room where it matters

Every rule above compresses to one field check: view the deck on the worst screen it will ever face. The old projector in the big meeting room, the video call at 720p, the conference laptop with the missing font. Ten minutes before the real meeting — or at the end of any day when you're already in that room — flip through the deck there. What breaks on bad screens: thin fonts, mid-gray text, red-green chart pairs, subtle gradients, anything smaller than 18 points. What survives: dark text on light, one dominant element, big numbers, direct labels. Fix what broke, and the deck is bulletproof everywhere else by definition. The teams with consistently clean decks aren't better designers; they just always look at the deck the way the room will.

The 15-minute order of operations

Run the fixes in impact order, because a half-finished pass should still leave the deck better: bullets first (4 minutes — everything else builds on the shortened text), hierarchy second (4), imagery third (3 — mostly deleting), master and alignment fourth (2), charts and contrast fifth (2). Doing them out of order wastes work: fixing charts before cutting bullets means re-fixing them when the slide splits in two.

One habit that makes the fixes faster every time: after your second or third deck, write your five recurring corrections into the generator's custom instructions or into your standing prompt constraints — "bullets under 12 words", "one dominant element per slide", "no decorative images", "charts: line for trends, bar for comparison, no pies over 3 slices". The formula already carries the text-side rules; this is the design side of the same trick. The fixes don't disappear, but they shrink from a pass to a glance.

When to stop

The good-enough line, stated as three checks: the deck is consistent (one master, aligned titles, uniform margins), it's readable from the back row, and every slide has one idea with one dominant element. Beyond that, you're polishing pixels, and the marginal hour buys nothing the room will notice. Design's job in a work deck is to not be noticed for the wrong reasons; the noticed-for-the-right-reasons part is your argument, and that lives in the edit pass, not the theme.

Your own screenshots and your own numbers will do more for credibility than any template ever built — which is why this last mile stays manual even as generators improve. The machine gets you a consistent, structured draft in minutes; the fifteen minutes of fixes are how it stops looking like one.