Beyond the Hype: A Practical, “Human-in-the-Loop” Workflow for AI

Introduction: The Trough of Disillusionment
If you scroll through LinkedIn for more than five minutes, you are bombarded with the same promise: “Create a full e-learning course in 30 seconds with AI!”

As Instructional Designers, we look at these claims with a healthy dose of skepticism. We know that “content” is not the same as “instruction.” Generating a wall of text is easy; designing a learning experience that changes behavior, adheres to adult learning principles, and aligns with business goals is hard.

We have moved past the initial shock of Generative AI and are now sitting squarely in what the Gartner Hype Cycle calls the “Trough of Disillusionment.” The shiny toy phase is over. Now, we have to do the actual work. We are asking the pragmatic questions: How does this actually fit into my ADDIE or SAM workflow? How do I use this without compromising data privacy? Does this actually save me time, or am I just spending hours prompt-engineering?

The truth is, AI isn’t going to replace the Instructional Designer, but it is fundamentally changing the drudgery of our work. The most successful IDs in 2026 will not be the ones who let AI do everything, but the ones who treat AI as a tireless, albeit occasionally hallucinating, Junior Assistant.

Below is a practical, stage-by-stage workflow for integrating Large Language Models (LLMs) like ChatGPT, Claude, or Gemini into the real-world instructional design process—not to generate the final product, but to amplify the human architect behind it.

Phase 1: Analysis & The SME Bottleneck

One of the most persistent pain points in our industry is the Subject Matter Expert (SME) disconnect. SMEs are brilliant, but they often struggle to un-complicate their knowledge. They suffer from the “Curse of Knowledge”—they cannot remember what it’s like not to know the material.

Historically, we would spend hours reading dense technical manuals just to prepare for a kickoff meeting. This is where AI shines as a Synthesis Engine.

The Workflow:

Instead of going into a kickoff meeting cold, use an LLM to build a mental model of the topic.

1. The “Explain Like I’m 5” Strategy: Paste a sample of the technical content (scrubbed of proprietary data) into the AI and ask: “Explain this concept to me using three different metaphors: one related to cooking, one related to driving a car, and one related to sports.”

• Why this works: You aren’t using this content for the course yet. You are using it to arm yourself with analogies so you can speak the SME’s language during the interview.

2. The Devil’s Advocate Generator: SMEs often focus on the “Happy Path”—when everything goes right. Before your meeting, ask the AI: “I am interviewing an expert about [Topic X]. What are 10 common misconceptions beginners have about this? What are the edge cases where this process usually breaks?”

• The Result: You go into the meeting asking targeted questions about failure states (“What happens if the system goes offline?”) rather than just nodding along.

Human Check: AI cannot understand the political landscape of your organization. It doesn’t know that “Department A” hates “Department B,” or that a certain term is taboo in your company culture. Use AI to understand the topic, but rely on your intuition to navigate the people.

Phase 2: Design & The “Blank Page” Syndrome

The Design phase is where “Writer’s Block” strikes hardest. Staring at a blank Storyboard or a blinking cursor is daunting.

Many IDs make the mistake of asking AI to “Write a course on Leadership.” The result is always generic, corporate drivel. The secret is granular, iterative prompting that aligns with Bloom’s Taxonomy.

The Workflow:

1. Objectives that Actually Measure: We all fall into the trap of using fuzzy verbs like “understand” or “know.”

• The Prompt: “I have a learning objective that says ‘The learner will understand data privacy.’ Rewrite this into 5 distinct, measurable performance objectives using Bloom’s Taxonomy levels of Application and Analysis. Focus on observable behaviors.”

• The Pivot: The AI will suggest things like “Identify three security risks in a provided email simulation” or “Differentiate between PII and non-PII data sets.” This is instantly more usable.

2. Scenario Ideation: Writing realistic dialogue is difficult. We often end up with robotic characters saying things like, “Hello Bob, surely we must follow the compliance protocol!”

• The Prompt: “Create a character persona named ‘Sarah,’ a stressed middle-manager who is skeptical of new software. Write a dialogue between her and an IT support specialist where she resists the new security update. Make her tone frustrated but professional.”

• The Human Polish: The AI provides the skeleton. Your job is to go in and add the “soul”—the company-specific acronyms, the specific stressors your learners face, and the natural cadence of speech.

Phase 3: Development & The “Grunt Work”

This is the phase where AI offers the highest Return on Investment (ROI) regarding time. Development involves a significant amount of repetitive, low-cognitive-load tasks that eat up our day.

1. The Distractor Dilemma

Writing a multiple-choice question is easy. Writing three plausible, non-obvious wrong answers (distractors) is incredibly hard. Usually, we end up with one right answer and three silly ones that learners can guess immediately.

• The Workflow: Feed the correct answer and the learning context to the AI.

• The Prompt: “I am writing a multiple-choice question about [Topic]. The correct answer is [X]. Generate 4 plausible distractors that represent common misconceptions a novice would make. Explain why a learner might choose each distractor.”

• The Benefit: This not only gives you the quiz options but also writes the feedback layers for you. When a learner clicks the wrong answer, you can paste the “why they might choose this” explanation into the feedback box.

2. The Technical Assistant (JavaScript & xAPI)

For IDs working in tools like Articulate Storyline or Adobe Captivate, we often want to do things the tool doesn’t support natively—like printing a certificate with the current date, or sending a specific xAPI statement to an LRS.

• The Workflow: You don’t need to be a coder anymore.
• The Prompt: “Write a JavaScript code snippet for Articulate Storyline that gets the current date and stores it in a variable called ‘SystemDate’. Then, write the code to print the window.”
• The Result: You get clean, copy-pasteable code. (Always test this, as AI can occasionally use outdated syntax, but it gets you 90% of the way there).

3. Accessibility at Scale

Writing Alt-Text for hundreds of images is tedious but ethically and legally mandatory.

• The Workflow: Upload your image to a multimodal AI.
• The Prompt: “Describe this image for screen reader software. Focus on the educational context of the chart, specifically the trend showing the dip in sales in Q3.”
• The Result: A specific, context-aware description that you can paste directly into your authoring tool.

Phase 4: Evaluation & The Synthetic User

We rarely have the budget for extensive User Acceptance Testing (UAT) with real learners before a pilot. AI can serve as a “Synthetic User” to stress-test your content.

The Workflow:

Paste your video script or course text into the AI.

1. The Clarity Check:

• Prompt: “Analyze this text for reading level. Highlight any sentences that are passive voice or use jargon that a 10th-grade reading level wouldn’t understand. Suggest simplifications.”

2. The Tone Check:

• Prompt: “Read this script. Does the tone sound condescending, authoritative, or collaborative? I am aiming for a ‘peer-to-peer’ coaching tone. Point out where I missed the mark.”

The Ethical Guardrails: What AI Cannot Do

While the workflow above sounds utopian, it relies on one critical factor: The Human Instructional Designer.

There are things AI simply cannot do, and these are the skills we must double down on to remain relevant:

1. Contextual Empathy: AI does not know that your sales team is currently burned out from a merger. It doesn’t know that the last training program was a disaster and trust is low. Only you can frame the training in a way that respects that emotional reality.

2. Accuracy Verification: AI hallucinates. It will confidently invent facts, citations, and policies. If you put unverified AI content into a compliance course, you are creating a liability for your organization. The ID’s role is shifting from “Creator” to “Editor-in-Chief.”

3. Data Privacy: We must never paste proprietary company data, PII (Personally Identifiable Information), or trade secrets into a public LLM. The workflow described above requires data sanitization—using generic terms in prompts and adding the specific details back in later manually.

Conclusion: The Future is Hybrid

The “Hype” told us AI would do it all. The “Reality” is that AI allows us to skip the blank page and the grunt work, freeing us up to focus on the high-value tasks: strategy, stakeholder management, and creative problem-solving.

We are no longer just slide-builders. We are Learning Architects powered by AI.

The practical workflow for 2026 isn’t about automating the instructional design process; it’s about augmenting the instructional designer. It’s about using the tool to clear the weeds so we can focus on planting the garden.

So, the next time you stare at a blank project file, don’t ask AI to “do it for you.” Ask it to brainstorm with you, debate with you, and code for you. Then, take that raw material and build something that only a human could design.