A designer on an L&D team does five distinct jobs that often get blurred together: visual identity, layout, image creation, motion or video, and quality control. The teams that produce professional-grade training without a designer are not using one tool to do all five. They are running a stack of four or five tools, each pointed at a single job, with a clear decision rule for when each one is the right answer.
LinkedIn’s 2025 Workplace Learning Report finds about 47% of L&D functions have generative AI fully embedded, and content authoring is consistently the top use case. The teams that have actually replaced or supplemented designer time, rather than just generating slightly nicer slides, have learned the decision framework below the hard way. This is what it looks like.
Job 1: Visual identity (set this once, everywhere)
Before any module gets designed, lock the visual identity in a one-page spec: two primary colours with hex codes, two secondary colours, one heading font, one body font, the rule for image style (photographic, illustrated, mixed), and the rule for icons (filled, outlined, two-tone). Most L&D teams already have this from marketing. If yours doesn’t, build it once in Canva’s Brand Kit and treat it as the source of truth.
The reason to do this first is that every downstream AI tool needs to be told the identity. AI image generation without a colour spec produces images that drift from your brand. AI presentation tools without a font spec default to whatever they default to. Five minutes of identity setup at the start saves hours of cleanup at the end.
Job 2: Layout and slide structure
The default L&D failure mode is to generate a script in ChatGPT and then paste it into PowerPoint slide-by-slide. The output is a deck that has the right words in the wrong shape: dense text blocks, no visual hierarchy, no rhythm.
The fix is to ask AI to design the layout grid first, then fill it. Tome, Beautiful.ai, and Gamma all do this in one step: paste your written content, get back a deck with reasonable layouts already applied. The output is rarely final, but it is usually 80% there, which is the difference between three hours of cleanup and twenty minutes.
If your stack is Canva-led, the prompt that gets you a usable layout brief is:
“You are a presentation designer. I am converting this 800-word training script into an eight-slide module. For each slide, design the layout, not the text. Specify: the dominant element (single visual, side-by-side comparison, three-column, full-bleed image with overlay, etc.), the supporting element (subhead, callout, bullet list of three items max), the visual hierarchy (what the learner reads first, second, third), and the colour-block strategy (which of our two primary colours dominates and why). Output as a numbered list, one slide per row, with a one-line layout brief I can hand to a designer or build myself in Canva. Our two primaries are [HEX 1] and [HEX 2]. SCRIPT: [PASTE].”
The output gives you eight specific layout instructions. Build each slide against its instruction. The result feels designed because it was, even though no designer touched it.
Job 3: Image creation
The wrong tool here produces obviously wrong output. Three sub-jobs, three different tools.
Stock-style on-brand photography: use Adobe Firefly or Canva Magic Media. Both produce defensible corporate photography. The failure mode is obviously fake people: extra fingers, melted faces, ages off-brief. Generate four variants, pick the one that survives a ten-second believability check.
Conceptual diagrams and process illustrations: use Napkin. It turns written process descriptions into clean diagrams without the baggage of generated photography. Right answer for “the four stages of a feedback conversation” or “how a complaint moves through our process.”
Custom abstract or branded artwork: use Midjourney or DALL-E 3 in ChatGPT. Highest fidelity, more iteration. Decision rule: realistic people = Firefly; diagrams = Napkin; stylised = Midjourney.
Job 4: Video without a film crew
The single tool that has changed L&D production in the last 18 months is AI video generation. Synthesia, HeyGen, and Vyond each fit a different job. Synthesia and HeyGen produce realistic AI presenters reading a script in your brand voice. Vyond produces animated explainer-style video where a presenter would be the wrong choice.
The script-writing prompt that works for AI video is shorter and more spoken than a written script:
“Write a 60-second script for a Synthesia AI presenter explaining [CONCEPT, e.g. why we run quarterly performance check-ins]. Audience: [AUDIENCE, e.g. new managers]. Tone: spoken, second-person, conversational. Open with a single sentence that names a problem the audience recognises. Body: three quick beats, each one a single point that takes 12 to 15 seconds to say aloud. Close with a one-sentence prompt for what the learner should do next. Use contractions. No lists. No headings. Word count: between 130 and 160 words to land at 60 seconds at a normal speaking pace.”
Generate the video, drop it into your authoring tool (Articulate Rise or Sana) as a section opener or a worked example. The cost per minute of usable video drops from “we need budget approval” to “we’ll have it tomorrow.”
Job 5: Quality control (the step everyone skips)
The reason AI-designed training often looks like AI-designed training is the missing fifth job. Real designers run a final pass for the kind of inconsistencies most people don’t notice consciously: padding, font rendering on the wrong slide, an icon set where one icon is from a different family. Take screenshots of every slide, paste them into a multimodal model, and ask it to flag every inconsistency against the brand spec from Job 1. Better still: keep your slides in Tome or Gamma, where a single regenerate-with-spec call applies the brand spec uniformly.
The counterintuitive part: you don’t replace the designer, you absorb the role
The L&D teams running this stack successfully are not making the same training their designer used to make, faster. They are making different training: shorter modules, more video, more diagrams, less generic stock imagery, more bespoke layout. The output is recognisably more contemporary than the in-house designer was producing five years ago, partly because the stack rewards a contemporary visual vocabulary.
The role that survives is the role of the person who decides: which job is which, which tool fits, where the brand spec lives, and what gets through quality control. That role is no longer “designer.” It is closer to “L&D producer,” and it usually sits with the same person who used to write the modules. They are now also the person who runs the design.
The failure mode: tool sprawl
The stack above is five tools across four or five jobs. The temptation is to add a sixth, an eighth, a twelfth, every time a new AI design tool launches. The teams that ship don’t. They pick one tool per job, learn it deeply, and only swap if the replacement is a clear step-change. A team running ChatGPT, Canva, Firefly, Synthesia, and Tome can produce nearly any module a corporate L&D function needs. A team running fifteen tools none of them know well produces less, slower, with more drift. Pick the stack. Run it for six months. Change one tool a year.
The training a small L&D team can ship today, without a designer on staff, is genuinely better than the training a small designer-equipped team was shipping in 2020. The output gain is real. The discipline is in keeping the stack small enough to actually master.