How Learning and Development Teams Can Use AI to Design Training That Captures Attention

Take any annual compliance module that the average employee opens, skips through, and gets a 60% on the quiz at the end. The content is fine. The compliance is real. The reason it loses attention is not the topic, it is the shape of the lesson. AI does not make boring content interesting by force, but it does change the shape of a lesson cheaply enough that you can run six structural variants in a morning instead of one variant per quarter.

This post walks through a worked before-and-after of a single 25-minute module. Same source material, same compliance objective, same audience. Two completely different outputs. One captures attention, one does not. The difference is in the prompts.

Before: the lesson that lost the room

The original module on workplace cybersecurity opens with a definitions slide (“phishing is…”), moves to a list of red flags, runs a generic example involving an invented employee called “Sarah,” ends with a five-question multiple-choice quiz. Twelve minutes of content stretched to twenty-five. Completion rate around 70%, average time-on-task slightly higher than minimum, quiz pass rate around 60%.

The failure mode is structural, not informational. Definitions front-loaded. No tension in the first 90 seconds. Examples that nobody recognises as their own job. Quiz that tests recall, not application. Every L&D team has shipped one of these. LinkedIn’s 2025 Workplace Learning Report finds 47% of L&D functions have generative AI fully embedded, and the most common production use case is exactly this kind of structural redesign of legacy content.

The redesign prompt: shape first, content second

The mistake when using AI to “improve” an existing module is to ask for a rewrite. The model rewrites at the sentence level. The problem is at the structural level. Ask for shape first.

“You are an experienced instructional designer redesigning a 25-minute module on [TOPIC, e.g. workplace cybersecurity for non-technical employees]. The audience is [DESCRIBE, e.g. office workers in a 600-person professional services firm, mostly desk-based, mixed seniority]. The current module opens with definitions, lists red flags, and ends with a recall quiz. Engagement is poor. Redesign the lesson SHAPE only, not the content. Propose three different structural arcs the lesson could take. For each arc, give: the opening 90 seconds (what the learner sees first, what tension or question hooks them), the spine (how the content unfolds and where the difficulty escalates), the practice mechanic (what the learner DOES, not just reads), and the closer (what the learner takes away into next Tuesday’s actual work). Do not write the lesson yet. Just the three arcs.”

The output is three structural options. Pick one. Run the next prompt against it.

After: the same content, redesigned around a scenario spine

The arc that won for this compliance module is a scenario spine. The lesson opens not with a definition but with a 30-second video of a realistic-looking email landing in an inbox. The learner sees the email exactly as it would appear in their actual client. The first interaction is a single question: would you click. Roughly 40% of learners click in the average cohort. The lesson now has the room’s attention, because half the room just got a quiet jolt.

Generating that opener used to mean a video budget. With Synthesia or HeyGen you produce a 30-second AI-presenter explainer in under twenty minutes, and with a screen-recorded mock-up of the email, you have the cold open at no production cost.

“Write a 90-second script for an opening scene in a cybersecurity training module. The setup: an employee at [COMPANY ARCHETYPE] receives an email that looks legitimate but is a sophisticated phishing attempt. The script should not yet say it’s a phishing attempt. Instead, it should describe the email exactly as it appears: the sender, the subject line, the time it arrived, the apparent context. End with a single decision point: click the link, or don’t. The visual style is [STYLE, e.g. plain screen recording of an inbox, no narration]. Tone: realistic, low-key, no music sting.”

That script becomes the cold open. The body of the lesson is the same content as before (red flags, anatomy of a phishing attack, the firm’s reporting process). The closer is a “now you” practice mechanic, not a recall quiz. Three new emails, the learner labels each one as “click,” “ignore,” or “report,” and the lesson explains why each label was correct.

The practice mechanic prompt: where attention actually compounds

A lesson holds attention if the practice is harder than passively watching. Ground the practice in realistic edge cases.

“Generate 12 practice scenarios for a cybersecurity training module. Audience: [AUDIENCE]. Scenarios should range from clearly safe to clearly malicious, with most in the realistic middle ground (subtly off, technically plausible, requires judgement). For each: the email as the learner would see it (sender, subject, body, max 60 words), the correct action (click, ignore, report), and a one-sentence explanation that teaches one transferable principle. Vary tactics so no two test the same red flag.”

The lesson uses three of the twelve. The other nine become next quarter’s refresher.

The result, and the honest read on why

The redesigned module runs the same 25 minutes. Same source content. The cold-open hook lands the room. The scenario spine builds tension and self-awareness. The practice mechanic forces the application the recall quiz never required. Completion rates and pass rates both move. More importantly, post-training reporting of suspicious emails to IT actually goes up, because the lesson trained the behaviour, not the definition.

The counterintuitive part: the AI did not make the content better. The content is virtually the same. AI changed how cheap it was to consider three structural arcs instead of one, and then to produce the production-grade artefacts (the script, the practice bank, the AI-presenter video) that the original lesson could not afford.

The failure mode: AI-redesigned content that’s all opener and no spine

The mistake L&D teams make when they get excited about AI for engagement is to over-invest in the cold open and under-invest in the spine. The first 90 seconds become a beautifully produced cinematic moment. The next 23 minutes are the same definition-list-quiz they always were. Engagement spikes for a minute, then flatlines. Completion stays where it was.

The fix is to apply the same prompt structure to every section of the lesson, not just the opener. Each section needs its own micro-arc: a question that pulls the learner forward, a piece of new information, a small practice or reflection. Articulate Rise makes this trivial structurally because it forces section-by-section authoring. The AI provides the inputs, Rise enforces the rhythm.

What this stack looks like in production

Three tools cover most of what the redesign needs. ChatGPT or Claude for shape design, scripts, and practice banks. Synthesia or HeyGen for any video element that used to need filming. Articulate Rise or Sana as the authoring layer. The full redesign of a 25-minute legacy module sits at around four to six hours, against the two to three weeks the original probably took.

The L&D teams whose modules employees actually finish are not the ones with the cleverest prompts. They are the ones who use the cheapness of AI generation to think harder about the shape of the lesson before any content gets written. Attention is captured at the spine. The cold open just signals that the lesson knew you were busy.