Experiment: AI Roleplay Training Simulator for De-Escalation, Built Without a Developer

What question did we test?

We tested whether a learning team could build a custom AI roleplay training simulator for de-escalation without a developer and without buying a roleplay platform. The goal was a practice tool that reflects the team’s own scenario, tone, branding, and learning objectives, built in hours of design work rather than a software project.

Most AI roleplay platforms are general-purpose. They work, but you often adapt your scenario to the platform instead of the other way around. We wanted to know if the newer AI app builders had reached the point where an instructional designer could describe the experience in plain language and get a usable simulator back.

The test case was a de-escalation conversation: a frustrated traveler whose flights have been changed and cancelled. It is a common customer service moment, emotionally loaded, and hard to practice with slides.

How was the experiment set up?

The setup was a fictional travel brand, Horizon, and one scenario: a customer affected by flight changes and cancellations. We used Google AI Studio to generate the simulator, Google Cloud for the API connection, and Articulate Rise and Storyline as the delivery shell. Tested in February 2026.

Before building, we did the design work. That step later became its own tool, SIM Creator, which Erick Prospero introduced in The AI SIM Blueprint (June 2026). It asks the questions an instructional designer would ask: what observable behaviors matter, what success looks like, and what failure looks like.

What did we build?

We built a browser-based simulator where the learner talks to an emotionally responsive AI customer. The persona tests empathy and validation: it reacts to how the learner responds, not just to whether they state the right policy. The interface carries Horizon branding, and the finished app sits inside a Rise or Storyline course.

9:29

The workflow had eight steps, all shown in the video:

1. Describe the foundation of the simulator in natural language prompts.
2. Create the AI persona and its emotional responses.
3. Write system instructions that guide realistic customer behavior.
4. Apply branding and visual design without custom code.
5. Let the AI generate the full application logic.
6. Test and refine for realism and learning value.
7. Connect to Google Cloud with an API key.
8. Publish and embed the simulator in Articulate Rise or Storyline.

What happened when we tested it?

The build worked: a non-developer produced a branded, responsive de-escalation simulator and embedded it in standard authoring tools. The bigger finding was that design decisions, not technology, determined quality. Where the system instructions were vague, the character drifted. Where behaviors were defined clearly, the practice felt real.

What worked

• Natural language prompting was enough to produce a working app and interface.
• The persona could be tuned to reward validation and empathy, which is the core of de-escalation.
• Branding and visual design were handled without code.
• Embedding in Rise and Storyline put the practice where learners already are.

What didn’t, or needed work

• The first versions needed several rounds of testing and refinement before the customer behaved believably [VERIFY: number of iterations].
• An API key and Google Cloud connection add a step that some teams will need IT to approve.
• Without clear success and failure behaviors defined up front, the simulator becomes a chat with no learning target. That lesson is why we built SIM Creator.
• Learner outcome data was not part of this experiment [VERIFY: pilot or learner results, if any].

Should you use AI roleplay training? The verdict

AI roleplay training is worth building when the skill is a conversation, the behaviors are observable, and learners need repeated, safe practice. It is a poor fit when the content is mostly knowledge, when every response must be scripted and approved, or when your organization cannot approve an external AI connection.

If scripted control matters more than open-ended practice, look at our Google Veo branching scenario experiment, where every path is written in advance.

How can you try AI roleplay training yourself?

Start with one conversation that goes wrong often in your organization, define what good looks like, and build a small prototype before scaling. A single scenario with one persona is enough to learn whether AI roleplay fits your learners, your content, and your IT environment.

1. Pick one scenario. Choose a moment with real emotional stakes, like a frustrated customer or a worried patient.
2. Define behaviors, not topics. Write two or three actions that show success (acknowledge, validate, offer options) and two that show failure (defend policy first, interrupt).
3. Write the persona’s rules. Describe the character’s situation, starting emotion, and what makes them calmer or angrier.
4. Prompt the app. Describe the simulator in plain language in an AI app builder such as Google AI Studio.
5. Test with practitioners. Have experienced staff try to break the character, then tighten the instructions.
6. Check the plumbing. Confirm API, data, and embedding rules with IT before learners see it.
7. Embed and observe. Place it inside a Rise or Storyline course and watch what learners actually do.

For more finished examples, browse our AI roleplay and simulation examples.

What’s next for this experiment?

The next step is pairing the AI roleplay with realistic video of the same customer, so learners see the person they are talking to. We have already produced that video side in our realistic AI video for customer service training experiment, using the same Horizon scenario.

We also want to test scoring: can the simulator give useful feedback on the specific behaviors defined in SIM Creator, and does that feedback match what a human coach would say [VERIFY: planned timing for scoring test]? We’ll publish what we find in the Lab, including the parts that don’t work.

If you are weighing roleplay against other soft skills approaches, our test of AI for soft skills video production covers the animation side of the same problem.