I Built an AI Role Play With No AI (and Claude Was a Very Good Sport About It) 🤖🚫

Last week at an ATD conference, I watched a demo where learners talked to an AI avatar client, asked questions, and got real feedback at the end.

My first thought: I need this.

My second thought: my budget cannot afford this. 💸

So I did what any reasonable designer would do. I opened Claude and asked it to help me build an AI simulation... without any AI.

Asking an AI to fake being an AI is a little like asking a chef to make you a microwave dinner. I half expected some pushback. Instead it said “totally doable” and then flagged everything wrong with my mockup and outline. Haha. Rude. Also correct. 😅

Meet Jordan Blake

Jordan is an Operations Manager whose routing errors have tripled, and she wants a 45 minute mandatory course for 300 people. By Friday. You know her. We’ve all met Jordan. 🙃

The learner’s job is to run a discovery conversation and figure out if training is really the answer. (Spoiler: an IT update in September might be involved. Shhh.)

So how do you fake a smart conversation?

A big table. Like, a really big table. Every topic Jordan knows about (the errors, what changed, who’s affected, the data) has a long list of trigger words. When the learner types, the code scores their message against every topic and picks the best match. It’s the same basic idea as Copilot pulling from resources, minus the part where it actually understands you. 😂

Then we layered on the stuff that makes it feel real:

• Follow ups. Say “tell me more” or “why?” and Jordan goes deeper on whatever you just asked.
• Two questions at once. “Who’s affected and when did it start?” gets both answers.
• Typos. “Has anythign chnaged?” still works. Bless.
• Personality. Small talk, thank yous, and yes, if you ask “Are you a bot?” she tells you she’s a very real manager with a very real deadline.

The right panel does the coaching

• Jordan’s Mood shifts live, from Open to Defensive.
• Meeting Clock counts down every turn, so rambling costs you.
• Case Notes fill in as you uncover clues, so learners can see their discovery building.

At the end, a scorecard rates Discovery, Root Cause, Relationship, and Recommendation, with specific coaching for each one.

The honest part

It’s not AI, and I tell learners that on the Tips page. Jordan responds to key words, and if she misses, the suggested questions keep things moving. I’d rather be upfront than have someone think the “AI” is broken. Trust matters, even with a fake client. 😄

Want to build your own?

• Start with the real performance problem, then work backward to the clues.
• Write triggers the way people actually talk, then test with someone who didn’t build it.
• Let the twists react to learner choices, not turn numbers.
• Built in a Rise 360 Code Block, uploaded as a zip, zero API keys required. 🎉

Give Jordan a try and tell me what you’d ask her first!

Is Training Really the Answer?