After 11 years building e-learning software at a major enterprise, here’s what surprised me when we started shipping GenAI features to real users.
The Starting Point Nobody Talks About
Most GenAI integration blog posts start with a clean slate — a greenfield app, a fresh codebase, a blank canvas. Real life is messier.
When we began integrating generative AI into our e-learning authoring tool, we weren’t starting from scratch. We were dealing with a mature enterprise product — millions of users, legacy architecture decisions made years ago, compliance requirements from Fortune 500 customers, and LMS interoperability standards (SCORM, xAPI) that were designed long before anyone imagined AI-generated content.
The challenge wasn’t “how do we call an AI API.” It was “how do we ship AI features into a product that thousands of instructional designers depend on daily, without breaking their workflows or their trust.”
Here’s what I learned.
Lesson 1: The AI Feature Your Users Want Is Not the One You Think
When we first scoped out AI integration, the engineering team gravitated toward the flashy stuff — generate an entire course from a prompt, auto-create assessments, AI-powered slide design.
Then we talked to actual users.
Instructional designers didn’t want AI to replace their expertise. They wanted it to eliminate the tedious parts of their workflow:
- Reformatting content across different output types (responsive HTML5, PDF, SCORM packages)
- Generating alt-text for hundreds of images in accessibility-compliant courses
- Summarizing lengthy SME-provided documents into digestible learning chunks
- Suggesting quiz questions from existing content (not generating courses from nothing)
The takeaway: don’t let engineering excitement drive your AI feature roadmap. Do 10 user interviews before writing a single line of integration code. The highest-impact GenAI features are usually the boring ones.
Lesson 2: Prompt Engineering Is a Product Decision, Not an Engineering Task
We initially treated prompt design as an implementation detail — something the backend team would figure out. That was wrong.
The prompts that generate content for an e-learning tool carry product-level implications:
- Tone and reading level — A course for factory floor workers needs fundamentally different language than one for executives. The prompt must encode this.
- Bloom’s taxonomy alignment — Assessment questions need to target specific cognitive levels (remember, understand, apply, analyze). This isn’t something you bolt on after; it shapes the entire prompt structure.
- Bias and accuracy — Enterprise L&D customers have zero tolerance for hallucinated content in compliance training. We needed guardrails baked into the prompt, not just a disclaimer.
We ended up creating a “prompt design review” step in our feature development process, where product managers, instructional design consultants, and engineers collaboratively refined prompts. It added a week to each feature cycle but prevented months of post-launch fixes.
Lesson 3: LMS Interoperability Makes Everything Harder
Here’s something unique to e-learning that most AI integration guides completely ignore: your AI-generated content doesn’t just live in your tool. It gets packaged as SCORM 1.2, SCORM 2004, or xAPI content and deployed to dozens of different LMS platforms — Moodle, Cornerstone, SAP SuccessFactors, Docebo, you name it.
This creates constraints that don’t exist in other domains:
-
AI-generated interactive elements need to produce valid SCORM API calls. An AI that generates a drag-and-drop interaction also needs to generate the corresponding
cmi.interactionstracking code. - Content structure must map to a navigable SCO (Sharable Content Object) hierarchy. You can’t just dump AI-generated HTML into a package.
- Completion and scoring logic must be deterministic. If AI generates a quiz, the scoring must be unambiguous and LMS-reportable. Probabilistic or fuzzy scoring breaks compliance reporting.
We built a validation layer that sits between AI output and SCORM/xAPI packaging — essentially a “content compiler” that checks structural validity, tracking completeness, and accessibility compliance before anything gets exported.
Lesson 4: Latency Tolerance Varies Wildly by Feature
Not all AI features need to be real-time. Understanding this saved us enormous infrastructure cost.
We categorized our AI features into three latency tiers:
| Tier | Example | Acceptable Latency | Approach |
|---|---|---|---|
| Instant | Smart text suggestions while typing | < 500ms | Edge-cached smaller models, pre-computed suggestions |
| Interactive | Generate quiz from selected content | 2-8 seconds | Standard API call with streaming UI feedback |
| Background | Alt-text generation for 200 images in a course | Minutes | Async job queue, notification on completion |
The mistake teams make is treating every AI feature as “interactive.” Your users will happily wait 5 minutes for bulk alt-text generation if you give them a progress indicator and let them keep working. But they won’t wait 3 seconds for an inline text suggestion.
Lesson 5: Enterprise Customers Ask Questions You Haven’t Considered
Before shipping AI features, be ready to answer these from procurement and IT security teams:
- “Does our training content get used to train the model?” — You need a clear, documented answer. For enterprise e-learning, the content often includes proprietary processes, trade secrets, and PII. Data residency matters.
- “Can we disable AI features entirely?” — Some regulated industries (healthcare, finance) may need to turn off generative features for compliance reasons. Build a kill switch from day one.
- “How do we audit AI-generated content?” — Enterprise L&D teams need to know which parts of a course were AI-generated vs. human-authored, especially for compliance training that faces regulatory review.
- “What happens when the AI model version changes?” — If you regenerate content with a newer model, it might differ from the previously approved version. Version pinning and content fingerprinting become essential.
We ended up building an “AI transparency layer” — metadata tagging that marks every piece of AI-generated or AI-assisted content with the model version, timestamp, and prompt used. This became one of our most valued enterprise features.
Lesson 6: Your Existing Architecture Will Fight You
Our authoring tool was built on a component architecture where content elements (text blocks, images, interactions, quizzes) were discrete, well-defined objects with schemas.
AI doesn’t naturally produce schema-compliant content. A language model generates text — converting that into structured content objects requires a translation layer that’s more complex than the AI call itself.
For example, when AI generates a quiz question, we need:
- The question stem as a specific content object
- Each answer option as a child object with correct/incorrect flagging
- Feedback text for each option as another child object
- SCORM interaction tracking metadata
- Accessibility attributes (ARIA labels, reading order)
We spent roughly 40% of our AI integration effort on this “structured output” problem — and I’d bet most enterprise tools face the same ratio.
Lesson 7: Measure What Matters, Not What’s Easy
It’s tempting to measure AI feature success by adoption rates (X% of users tried the AI quiz generator). That’s vanity.
The metrics that actually mattered for us:
- Time-to-publish — Did AI features reduce the average time from blank project to published course?
- Revision cycles — Are AI-generated elements surviving SME review, or getting deleted/rewritten?
- Accessibility compliance rate — Did AI-generated alt-text actually pass WCAG audits?
- LMS compatibility — Are AI-generated interactions producing valid tracking data across LMS platforms?
After 6 months, our data showed that AI-assisted alt-text generation had an 87% first-pass acceptance rate and cut accessibility compliance time by ~60%. The “generate entire course” feature — the one engineering was most excited about — had a 23% acceptance rate and most outputs were heavily edited.
Build more of what works. Kill what doesn’t, regardless of how technically impressive it is.
The Bottom Line
Integrating GenAI into an enterprise product isn’t a model problem — it’s a systems problem. The AI call is 10% of the work. The other 90% is:
- Understanding your users’ actual pain points
- Respecting existing standards and interoperability requirements
- Building guardrails, audit trails, and kill switches
- Designing the translation layer between AI output and your product’s content model
- Measuring real impact, not usage vanity metrics
If you’re working on AI integration for any mature enterprise product — especially in e-learning, healthcare, or regulated industries — I’d love to hear what you’ve run into. The challenges are more similar than you’d think.
Top comments (0)