Last Tuesday, I watched ChatGPT write a technical job aid in 4 minutes that would have taken me 3 hours to draft from scratch.
Then I spent the next 90 minutes rewriting half of it.
Not because the AI was wrong. It wasn't. The terminology was accurate. The steps were logical. The formatting was clean.
Every single example it generated was fictional. The workflow it described was technically correct but completely missed the part that actually confuses Delta's customer service specialists. And the tone? Perfect for a software manual, terrible for someone handling an angry passenger at 6 AM.
Here's what 18 months of using AI daily for training development at Delta has taught me: AI accelerates brilliantly and replaces catastrophically.
The people who figure out the difference will produce training faster than ever. The ones who don't will flood their organizations with technically accurate content that nobody can actually use.
Note: All examples in this article have been anonymized to protect proprietary information and customer privacy, but the lessons and patterns remain accurate.
AI Writes Job Aids Perfectly While Completely Missing What Confuses People
AI has access to feature documentation, technical specs, and system workflows. You know what it doesn't have? The Slack message from last Thursday where someone said "I know how to do the thing, I just don't know WHEN to do the thing."
I learned this the expensive way.
We rolled out a new rebooking tool for irregular operations. I fed the technical documentation to ChatGPT. It produced a beautiful 3-page job aid explaining every button, every field, every option in the system.
Specialists looked at it and said "This is great, but when do I use the simple rebooking versus the complex rebooking?"
The feature wasn't the problem. The decision point was the problem.
AI knew how the tool worked. It had zero insight into the actual moment of confusion that happens when a specialist has 5 people in line and needs to make a judgment call in 30 seconds.
That insight came from reading support tickets, sitting with specialists during their shift, and noticing the pattern. They weren't confused about the mechanics. They were confused about the strategy.
No amount of prompt engineering fixes this. AI can't attend your team meetings. It can't read your support tickets. It can't notice that your users keep asking the same question that your training never addresses.
That part still requires you.
AI Loves Clean Scenarios Where Everything Works (Reality Is Messier and More Useful)
Here's a ChatGPT-generated scenario from last month:
"Sarah needs to rebook a customer whose flight was cancelled due to weather. She opens the rebooking tool, searches for available flights, selects the best option, and successfully rebooks the customer. The customer thanks her and ends the call satisfied."
Cool. Useless, but cool.
Here's what actually happened last Tuesday according to a support ticket:
"Customer was on a cancelled flight but had already rebooked themselves on a partner airline through our app. They're calling because they want their original bag fee refunded but our system is showing they still have an active booking even though they're not actually flying with us anymore. Do I process the refund first or fix the booking status first?"
One of these scenarios teaches decision-making. The other teaches following perfect instructions in a perfect world.
AI will happily generate 47 variations of the first scenario. It cannot generate the second one because it's never seen your actual support tickets, listened to your call recordings, or watched someone get stuck.
After I started pulling real scenarios from support tickets instead of using AI-generated ones, specialists started saying things like "Oh yeah, I had this exact situation last week." That's when you know your training is actually useful.
Real scenarios are messier. They're less polished. They involve system bugs and edge cases and situations where there's no perfect answer.
They're also the only scenarios worth training on.
What You Must Do Before Handing Anything to AI
AI will build whatever you ask for. It will not tell you what you should ask for.
Should this be a 30-minute eLearning module or a 1-page job aid? AI doesn't know. It will cheerfully build the module while you waste 40 hours on training for a feature that changes in 3 weeks.
I built a framework after watching myself and other designers make this mistake repeatedly. Four questions before touching any AI tool.
Why should they care about this feature?
Not "what does it do" but "what can they do NOW that they couldn't do BEFORE?" If the answer is "it's 2 clicks instead of 4," you don't need training. You need a Slack announcement.
When does this actually matter in their workflow?
Not "what is this feature" but "what's the MOMENT in their day when they'd need this?" If you can't describe the moment, you don't understand the problem well enough to build training.
What's the one thing they try first?
The 30-second version. If they walked away after 30 seconds, what's the one action that would make them feel competent? That's your practice exercise. Everything else is optional.
What's the shelf life of this training?
Experimental features get job aids (2 hours to create, 20 minutes to update). Core features get full modules (40 hours to create but worth it because they last). AI doesn't know the difference.
These decisions are 100% human. They require context about your business, your users, and your roadmap. Get these wrong and AI will very efficiently help you build the wrong thing.
The 30-Minute Draft That Takes 90 Minutes to Polish (And Why That's Still Worth It)
Once I know what I'm building and why, AI becomes dramatically useful.
My actual workflow using ChatGPT and Copilot.
The Prompt That Actually Works
I don't send AI clean requirements. I send it my messy notes from SME interviews, real support tickets, and actual user questions. Then I tell it what format I need.
Example prompt:
"I need a 2-page job aid for customer service specialists handling rebooking scenarios. Target reading level: 8th grade. Tone: conversational but professional, like talking to someone who's busy and stressed. Here are the actual scenarios from this week's support tickets: [paste 3-4 real tickets]. Create a job aid that helps them make the decision quickly."
What I get back in 4 minutes instead of starting with a blank page for 45 minutes: 70% of the way there.
Where I Spend the 90 Minutes
Not rewriting everything. Targeting the parts AI cannot do.
- Replacing generic examples with real support tickets
- Adding the "why this matters" context that comes from knowing how specialists actually work
- Adjusting tone from "technically correct" to "actually how we talk here"
- Adding the edge cases and exceptions that only exist in your specific system
- Removing the perfectly polite corporate-speak that nobody uses when they're handling 5 people in line
Side-by-side example:
AI Version:
"When a customer's flight is cancelled, follow these steps to provide rebooking assistance: 1. Verify the customer's booking. 2. Search for alternative flights. 3. Present options to the customer."
Final Version:
"Cancelled flight? Quick decision: If they're traveling in the next 4 hours, call the rebooking hotline (they're faster for urgent situations). If it's later than 4 hours, you can handle it. Here's how..."
One version is correct. The other version helps someone make a decision in 30 seconds while someone's yelling at them about missing their daughter's graduation.
Time breakdown: 4 minutes for AI draft + 90 minutes for human polish = 94 minutes total versus 3 hours from scratch. I'm saving about 90 minutes per piece of content.
But only if I know which 90 minutes to spend.
The Quality Control Checklist I Actually Use (Because AI Makes Subtle Mistakes)
AI will confidently write things that sound correct but aren't.
Not usually. But often enough that skipping quality control is how you teach people the wrong thing and don't discover it until someone sends a confused Slack message 3 weeks later.
Before anything goes live:
The "Would They Actually Say This?" Test
Read it out loud. If it sounds like a software manual, rewrite it. Specialists don't say "initiate the rebooking process." They say "start the rebooking."
The Real Example Check
Every scenario in the training needs to be real or based on real situations. If I made up an example to make the content flow better, I delete it and find a real one.
The Confusion Audit
The question I'm always asking: "Does this answer the questions users actually have, or just the questions I think they should have?" If I'm not sure, I send it to someone who does the job and ask where they'd get stuck.
The Technical Accuracy Review
SME reviews are not optional. I've caught AI confidently explaining workflows that were accurate 6 months ago but changed in the last update.
The Shelf Life Reality Check
Did I build this for the shelf life it actually has? If this feature changes next month, can I update this in 20 minutes or do I need to rebuild the whole thing?
Last month this process caught a job aid where AI had confidently explained how to handle a situation that's technically possible in the system but violates company policy. The steps were accurate. The advice would have gotten someone written up.
That's why the quality control can't be automated.
The Tools and Prompts That Work (Not the Ones That Look Good in Blog Posts)
I use ChatGPT and Copilot. Not because they're the best. Because they integrate with tools I'm already using and Delta's IT department has approved them.
I tried Claude. I tried Jasper. I tried tools specifically marketed for training development.
What I learned: The tool matters less than the prompt quality and your willingness to edit.
Prompts I Actually Use
For job aids:
"Create a 1-page job aid for [specific audience] handling [specific task]. Include: the decision points where people get stuck, the 30-second version, and what to do when [common exception]. Use real scenarios from these support tickets: [paste tickets]. Format: scannable headers, bullet points, no full paragraphs."
For video scripts:
"Write a 2-minute video script explaining [feature] to [audience]. Tone: conversational and fast-paced. Structure: problem (15 seconds), solution overview (30 seconds), show it in action (1 minute), where to get help (15 seconds). Avoid corporate jargon."
For interview guides:
"I'm interviewing a subject matter expert about [topic]. Generate 10 questions that will help me understand: what users struggle with, what mistakes they make, what questions they ask repeatedly, and what context they need to make decisions. Focus on the gaps, not the happy path."
The pattern: I'm specific about format, audience, tone, and purpose. I provide real examples. I tell AI what to avoid.
Generic prompts get generic output.
What I don't waste time on:
Trying to make AI's first draft perfect. It won't be. I use it for the structure and the first pass, then I make it useful.
The Uncomfortable Truth About Time Savings
Am I saving time? Yes.
Am I saving as much time as I thought I would? Not even close.
The math I expected: AI generates content in minutes instead of hours equals massive time savings.
The math that actually happens: AI generates drafts in minutes, quality control takes longer than I estimated, some things are faster to write myself than to fix AI's output.
My honest current ratio:
About 60% of what I create uses AI for first drafts. 40% I still write from scratch because by the time I explain what I need to AI and then fix what it gives me, I could have written it.
When AI actually saves time:
- Reformatting existing content (take this 10-page document and turn it into a 2-page job aid)
- Generating multiple variations of the same thing (manager version and specialist version of the same training)
- Creating structure and outlines when I'm staring at a blank page
- First drafts of straightforward procedural content
When I skip AI entirely:
- Anything requiring deep knowledge of our specific user struggles
- Strategic decisions about what format to use
- Content where tone and voice really matter
- Situations where explaining what I want takes longer than just writing it
The real value isn't eliminating the work. It's eliminating the blank page problem and the first-draft paralysis.
I'm not spending 3 hours writing anymore. I'm spending 30 minutes editing. That's a real improvement.
But I'm not spending 30 minutes total. I'm spending 2 hours on quality control, SME reviews, and making sure I'm not teaching people something that's technically correct but practically useless.
AI Is a Tool, Not a Strategy
The question isn't "Should we use AI for training development?"
The question is "What should we use AI for and what should we keep doing ourselves?"
After 18 months of daily use, here's what I know.
AI accelerates drafting. It doesn't replace thinking.
AI handles structure. It doesn't handle strategy.
AI knows features. It doesn't know your users.
The skill that matters now isn't prompt engineering. It's knowing what to automate and what requires human judgment.
Start small.
Pick one type of content where you spend a lot of time on first drafts (job aids, scripts, outlines) and experiment with AI doing the first pass while you do the polish.
Build your quality control checklist before you scale. Figure out what mistakes AI makes in your specific context and how to catch them.
And remember: The goal isn't to eliminate your role. It's to spend less time on the parts that don't require your expertise so you have more time for the parts that do.
The real question is whether you're going to figure this out deliberately or get pushed into it while your competitors are already moving faster.