AI Course Creator — Case Study · Aman Verma
Aman Verma

Shipped Product design case study · Lyearn

Designing an AI course creator that builds the right course, not just a fast one.

A progressive, context-led AI system that turns trusted company knowledge into structured, reviewable learning, without removing the course creator from the process.

See the flow Jump to impact
Learning Content › Create Course

All Content

+ Create content
Search content
ListGrid
All content typesStatusCategoryCreated byFilter
ContentTypeStatusCategoryLearnersLast updated
Negotiation fundamentalsCoursePublishedSales1,24012 Sep 2026
Customer discovery 101CoursePublishedProduct8609 Sep 2026
Data privacy essentialsCoursePublishedCompliance2,3102 Sep 2026
Running effective 1:1sCourseDraftLeadership—29 Aug 2026
Quarterly planning kitCollectionPublishedOperations54021 Aug 2026
Handling objectionsCoursePublishedSales98018 Aug 2026
Onboarding: your first 30 daysPathPublishedHR3,12014 Aug 2026

How would you like to create your course?

Step 1 of 9

Start · choose how to create the course

Interactive recreation of the shipped product, rebuilt from a screen recording of the live implementation.

The brief · what this project solved

Lyearn’s learning management system (LMS) could deliver training, but L&D teams still built every course by hand. I designed an AI course creator inside the LMS that turns their own documents into a structured course.

  1. 01The problem

    Every course started from a blank page.

    Admins read dense source material, decided what mattered, set learning goals, structured modules and wrote every lesson. Creators who tried outside AI tools still copied, pasted and reformatted it all into the LMS.

  2. 02What I designed

    A guided AI flow inside the LMS.

    Setup captures the context once. The creator approves learning outcomes before AI builds the structure, content and assessments, so each approved step feeds the next.

  3. 03What changed

    Days of authoring became a day or two of review.

    L&D teams moved from writing courses to checking them, and kept control at every stage.

RoleSole designerDiscovery to shipped product
OwnershipProduct strategy + UI/UXAI workflow, interactions, handoff
TeamProduct · AI · EngineeringDaily design and build loop
PlatformEnterprise LMSClient trains 56,000+ learners
01The real problem

“I already have the material. Turning it into a course is the hard part.”

Lyearn already delivered training, but creating it was manual. Admins started from a blank course and had to read dense source material, decide what mattered, define learning goals, structure modules and then write every lesson.

What we heard
“I already have a long product document. To make it training-ready, I still have to read it1, decide what matters2, define what my team should learn3, structure it into modules4, and then create the lessons5.”
L&D manager, enterprise clientRaised in client calls
  1. 1Source materialDocs, SOPs, updates
  2. 2What matters?Identify key concepts
  3. 3What to learn?Define learning goals
  4. 4How to structure?Modules and lessons
  5. 5Create the courseContent + assessments

Most of the effort happened before content creation even began.

02Stakes and success

The client needed hundreds of courses, fast, without losing quality.

A major enterprise client was preparing training for a new card-product rollout. The client trains more than 56,000 learners across 20+ business units, from voice and telesales to KYC and NRI banking. Manual authoring would slow readiness and stretch L&D teams. Competitors were already offering AI authoring, and the client was evaluating alternatives, so closing this gap was critical to retaining the account.

  1. New card-product rolloutTraining had to be ready on time
  2. Hundreds of courses neededFor 56,000+ learners across 20+ teams
  3. Competitors offered AI authoringThe client was evaluating alternatives
  4. Risk of losing the clientThey could move to an LMS that already had AI authoring
03Discovery

AI had already entered the workflow. Just outside the product.

Creators were already drafting with outside AI tools. It was faster, but it meant juggling three places: the source doc, the AI tool and the LMS.

  1. 01Open source docRead the product doc to find what matters.
  2. 02Switch to AI toolLeave the LMS for an outside AI tool.
  3. 03Re-explain contextEach new chat starts cold, so paste the context again.
  4. 04Copy outputCopy the draft the AI wrote.
  5. 05Paste into LMSOpen the course in the LMS and paste it in.
  6. 06Reformat it allThe formatting breaks, so fix it by hand.

The opportunity wasn’t to introduce AI. It was to bring that speed into the workflow, without the friction.

04Questions and answers

Using AI was obvious. Designing the right way to use it wasn’t.

  • Grounding + generationGenerate it all at once or step by step, and how do we keep it tied to trusted sources?
Our answer

Speed, in the right direction

Optimise for an accurate first pass instead of endless prompting, regeneration and wasted credits.

  • Interaction modelHow much prompting should users have to do?
Our answer

Clear steps over open conversation

Ask the key questions once, upfront, then follow a consistent flow so the system always knows its context.

  • Human controlWhat should AI decide, and where should users review?
Our answer

AI suggests, the user decides

Edit, regenerate one item or all of them, and move forward only when a stage is ready.

05Market and options

The market had proven AI could generate courses. The approaches were very different.

Approach Lyearn AIOur approach Sana 360Learning Docebo TalentLMS
Generation styleProgressive, outcomes firstFast source-to-courseOutline first, then courseDraft, then refineFast one-shot
Prompt burdenLowKey questions asked once in setupMediumMediumMediumLow–medium
Human checkpointsAt every stageEdit after generationReview outline firstEdit blocks + structureLighter review

Competitor approaches as reviewed during discovery.

Five ways to generate a course

Select an approach to see where the creator gets a say

Creator checkpoint: you review before moving on

Best for

Trade-off

06The decision

The fastest AI flow wasn’t the best course-creation flow.

Two choices were debated internally. Both added a step upfront, and both removed rework later.

1
Decision 1

How users talk to AI

Chat-led
Structured flow

The team, like much of the industry, leaned towards chat. But every new message is another inference step: context gets buried, users keep explaining themselves, and credit usage climbs.

My callI pushed to ask the important questions once in setup, then carry that context forward.

2
Decision 2

What AI builds first

Skip outcomes
Define outcomes

Going straight to an outline was faster, but AI would structure the course before the learning goals were explicit: plausible, yet misaligned. Outcomes cost an extra step and more AI credits.

My callI proposed defining and reviewing outcomes before AI expands them into structure and content.

A progressive, context-led flow

Each approved step becomes stronger context for the next one. Select a step to see who does what.

AI does

Creator decides

    What the learner gets

    Everything above happens on the admin side. The finished course reaches learners across 20+ business units, on desktop or in the Lyearn mobile app. Over the past year, 23,500+ of them were active on the platform.

    Desktop lesson reader: a hero image, the lesson title, body text, an audio clip, key points and a quote
    Desktop · the lesson reader
    Mobile course page: cover, title, due date and duration, about, skills, instructors, prerequisites and curriculum with a Start Course button
    Mobile · course page
    Lyearn mobile app home: a featured learning path with a Resume Path button, the library, and bottom navigation
    App · home

    The goal wasn’t the fastest possible generation. It was the fastest path to the right course.

    07Designing with AI + engineering

    The constraints became the design.

    A daily loop with product, AI, and engineering connected workflow decisions to implementation.

    1. Constraint 01

      Every generation costs credits

      Design response

      Capture key inputs once and provide targeted regeneration.

    2. Constraint 02

      Large documents take time

      Design response

      A 50 MB upload limit, with multiple smaller documents supported.

    3. Constraint 03

      AI output needs editing

      Design response

      Support both individual changes and stage-wide regeneration with instructions.

    4. Constraint 04

      Creators can lose context

      Design response

      Keep relevant earlier information in side panels; use outcome labels and colours across the structure.

    5. Constraint 05

      Course depth varies

      Design response

      Make knowledge level, weekly time, and duration explicit inputs.

    6. 5 → 5

      Five hard limits, five design decisions. Each constraint became part of the flow instead of a workaround.

    08Key screens

    Each screen answers one of those limits.

    Course structure · try it

    Keep every lesson traceable to a goal.

    With many outcomes and lessons, users could easily lose track. The structure keeps earlier decisions in view. Select an outcome to see which lessons cover it.

    Lessons

    Every generated lesson has a reason to exist.

    From the shipped product

    Screens captured from the live course creator. Select any screen to view it larger.

    09Feedback → iteration

    Tested with the knowledge they already worked with.

    10participants1 L&D manager + 9 team members

    After a team walkthrough, participants tried the demo using their existing knowledge-base content and shared feedback.

    Feedback 01Paraphrased request

    Change the whole set.

    Revising outputs individually was tedious when the same change applied to an entire stage.

    Design response

    Bulk regeneration

    One common instruction revises all outputs in the current stage, using prior-step context.

    Feedback 02Paraphrased request

    Let me review it all later.

    The team also wanted a way to generate the whole course at once.

    Design response

    Full-course auto-generation

    An optional path generates the complete course for review and editing afterward.

    Feedback 03Paraphrased request

    Only use the parts that matter.

    Teams uploaded their existing knowledge pages, but some of those pages weren’t relevant to the course they were building.

    Design response

    Instructions for each document

    Every uploaded PDF gets its own comment field. Creators tell the AI how to use that file, such as which pages to skip, and the AI follows it.

    After trying these additions, the team reported being satisfied with the updated workflow.

    Context side panels and colour coding were designer-led improvements. These findings are qualitative feedback, not measured usability gains.

    10Impact

    Course creation went from days of work to a day or two.

    4–5 days1–2 days

    Time to create one in-depth course · project-reported result after rollout

    One in-depth course, across a working week

    1–2 days per course. The striped part is the range.

    What the 10 testers saw

    One row per participant: time to build a course before and after, and what they said.

    Per-person values are illustrative.

    Before (manual)After (AI Course Creator)
    1. P1L&D manager 5 days1.5 days “Reviewing outcomes first meant far less rework later.”
    2. P2Team member 4.5 days1 days “Setup asked the right questions once, so I stopped re-prompting.”
    3. P3Team member 5 days2 days “Regenerating a whole stage with one instruction saved a lot of clicks.”
    4. P4Team member 4 days1 days “The draft already followed our own documents.”
    5. P5Team member 4.5 days1.5 days “I mostly reviewed and tweaked instead of writing from scratch.”
    6. P6Team member 5 days2 days “The colour-coded outcomes made gaps easy to spot.”
    7. P7Team member 4 days1 days “Skipping irrelevant PDF pages kept the content focused.”
    8. P8Team member 4.5 days1.5 days “Auto-generate gave me a full draft to review in one go.”
    9. P9Team member 5 days2 days “Editing one lesson without touching the rest felt safe.”
    10. P10Team member 4 days1.5 days “Assessments matched the outcomes without extra work.”

    Average in this illustration: 4.5 days → 1.5 days per course, in line with the reported team result of 4–5 days → 1–2 days. Rows and quotes are placeholders to be replaced with real study data.

    Minor edits

    Uploaded content is the source of truth, and outcomes are set first, so structure and content start out accurate.

    Adopted

    The client now uses the AI Course Creator for new courses.

    Business value

    Shipping it closed the AI gap and helped retain a key enterprise client.

    Context: the courses created here serve a client audience of 56,000+ learners. Saving around three days per course matters at that scale.

    For L&D teams, the biggest change was relief: what took most of a week now takes a day or two.

    11Reflection

    The model is only one part of the system.

    The real product work is shaping context, review and recovery.

    01
    Learning

    Context is a UX problem

    Better output came from designing better inputs and carrying them forward, not from asking users to write better prompts.

    02
    Learning

    AI speed can amplify mistakes

    The earlier a wrong direction is caught, the less has to be regenerated later.

    03
    Learning

    Control needs multiple levels

    Users need to change one item, a whole section or the whole course, depending on what went wrong.

    What’s next

    Reduce review burden

    Help creators see what actually needs attention, instead of reviewing everything equally.

    More case studies

    More from the platform

    Design to code, MGSA, Career paths and Org structure have their own pages. Knowledge base is being written up.