Search for "AI role play training" and you'll find rankings of tools written by the companies that sell them. Meanwhile, the category is growing fast. When Udemy launched AI role-play in May 2025, creators published more than 3,100 scenarios in four weeks (Udemy). Synthesia raised $200 million and announced a shift from video toward conversational agents that employees can practice with (Synthesia).
The analysts are more cautious. Gartner placed generative AI simulators at the very start of its 2026 Hype Cycle for corporate learning technologies, according to a vendor press release (Digital Terminal). Fosway, the European learning technology analyst, openly asks whether AI role-play actually changes behavior at work (Learning News).
For instructional designers, the question is practical: when should you use AI role-play, when is a well-designed branching scenario still better, and how do you combine them? Here is what the evidence and the most cited voices in learning design say.
EP 03Key takeaways.
- Branching scenarios are scripted, and AI role-play improvises. Each is better for different skills.
- The best studies show that AI practice can help, but none has measured long-term behavior change at work.
- AI helps learning when it makes people think, and harms it when it thinks for them.
- The experts agree on one rule: humans design the decisions and the standard, and AI plays the part.
- A hybrid, with a designed decision structure and AI characters inside it, combines consistency with realism.
Branching scenarios, AI role-play and hybrids: definitions
A branching scenario is a designed story in which the learner makes decisions at set points, and each decision leads to a different consequence. The designer writes every path. That makes it consistent, auditable and easy to compare between learners and over time.
AI role-play training puts the learner in a conversation with an AI character, by text, voice or video, that responds freely within instructions the designer sets. It gives unlimited variation and many repetitions, but less control over what happens and how performance is judged.
A hybrid keeps a designed decision structure (key moments, consequences and a debrief based on your standard) and lets AI characters play the counterpart and vary the dialogue. Devlin Peck, an instructional design educator who also sells an AI role-play product, puts the distinction simply: a branching scenario follows a script, while an AI role-play improvises within guardrails you define (devlinpeck.com).
What the evidence says about AI role-play training
The research base is young, small and mostly outside corporate settings. It is still worth knowing, because critics will ask about it.
| Study | Setting | What it found | Limitation |
|---|---|---|---|
| Rehearsal, Stanford (CHI 2024) | 40 participants practicing conflict | After practicing with an AI simulation, people escalated 67% less in a later conflict conversation and were twice as likely to choose cooperative strategies than after a lecture | Small lab study, not a workplace |
| AI vs. actor patient (JMIR 2025) | Randomized crossover, two UK medical schools | AI simulation produced smaller self-rated skill gains and lower satisfaction (8.09 vs. 9.21) than an actor; the authors still see it as a low-cost way to add practice capacity | Self-rated outcomes |
| AI virtual patient (PMC) | 87 trainee psychotherapists | Improved competence and confidence, but didn't beat video-based training | Single program |
| AI coaching RCT (summary) | 114 senior leaders | Human coaching had an effect; AI coaching did not outperform the control group | Secondary summary of the study |
| AI tutor with and without guardrails (Wharton/PNAS 2025) | About 1,000 high school students | Unrestricted AI improved practice scores by 48%, but students did 17% worse once it was removed; a tutor version that made them think avoided the loss | Students, math, Turkey |
Two patterns stand out. First, AI practice adds capacity but rarely beats practice with skilled people. The defensible position is "AI for scalable, repeatable practice before and between human sessions", not "AI replaces trainers". Second, design matters more than the model. The same technology helped or harmed learning depending on whether it made learners think or gave them answers. A randomized study at Anthropic showed the same effect with software developers: those who delegated coding to AI scored 50% on a comprehension test, versus 67% without AI (Anthropic).
What is missing matters just as much: no published study has measured long-term transfer of AI role-play to workplace behavior. Vendor claims about faster ramp-up times come without control groups. If you pilot AI practice, plan your own before-and-after measurement from day one (see how to measure training effectiveness).
What learning experts recommend
Philippa Hardman: use AI as a practice engine, not a content machine
Dr Philippa Hardman, one of the most widely read writers on AI in learning design, argues that the real opportunity of generative AI is practice, not content. She describes four formats: structured role-play, a decision simulator with compounding consequences, a feedback simulator and an adaptive case study. She also lists the failure modes: characters drift or break role, debriefs default to the AI's norms instead of the organization's, and subject experts must test scenarios before launch. AI practice should supplement human practice, not replace it (Hardman, Substack).
On avatars, she warns that realism is the wrong thing to optimize. Script, pacing, emotion and relevance matter more, and learners who detect AI report lower trust, although she still recommends disclosure (Hardman).
Mike Taylor: AI needs an expert in the loop, not just a human
Learning designer and Friday Finds newsletter author Mike Taylor argues that AI without direction simply produces bad training faster. He asks L&D to act like a film director, with AI as the crew. In his words: "AI raises the floor on training production. It does nothing for the ceiling." (Friday Finds). His preferred example of well-directed AI is a feedback-conversation scenario followed by a group debrief.
Christy Tucker and Cathy Moore: write consequences, not "correct!"
Branching-scenario specialist Christy Tucker uses AI as a co-author, especially for drafting plausible wrong choices and their consequences, which is the slowest part of scenario writing. She is not yet convinced by fully AI-generated scenarios (YouTube; Training Magazine Network). Cathy Moore, creator of action mapping, has long argued that a good scenario continues the story after a choice instead of announcing "correct" or "incorrect" (YouTube). Both points apply equally to AI role-play: realism comes from the situation and the consequence, not the rendering.
Julie Dirksen and Clark Quinn: who does the thinking?
In October 2026, Julie Dirksen hosted a three-part series on the Mind Tools L&D Podcast about who judges work that AI contributed to, what "human in the loop" really means and how to stop AI doing your thinking (Mind Tools). Clark Quinn warns that generative AI produces average answers and wants it used as support, not as a solution (Learnlets). Gartner reached a similar conclusion for 2027: training must actively counter "cognitive surrender", where people stop checking what AI tells them (HR Dive).
Donald H Taylor and Ross Stevenson: practice is where AI goes beyond content
Donald H Taylor, chair of Learning Technologies, told the Slovak podcast e-learning žije! in March 2026 that conversation practice is the first area where AI moves beyond content. It used to require an expensive human, and it is especially strong in customer support, although results depend on the underlying data (YouTube). Ross Stevenson of Steal These Thoughts calls practice a core pillar of L&D's AI use. He also warns that most teams won't build an interactive video coach themselves, and that output quality isn't guaranteed (DataCamp, YouTube).
When to use AI role-play, branching scenarios or a hybrid
| If you need to… | Use | Why |
|---|---|---|
| Train a procedure with one right way (safety steps, a policy rule) | Branching scenario | Consistent paths, clear standard, auditable record |
| Compare learners or measure before and after | Branching scenario or hybrid | The same decision points make results comparable |
| Give many repetitions of an open conversation (objections, small talk, de-escalation) | AI role-play | Unlimited variation and volume |
| Practice a high-stakes conversation where your standard matters (performance review, complaint with a refund limit, gift during a tender) | Hybrid | Designed decisions and consequences, natural dialogue from AI characters |
| Build fast with scarce subject expert time | Hybrid | The expert approves 5 decision points, AI drafts the dialogue |
| Document understanding for compliance | Branching or hybrid | Recorded decisions against a defined standard (see compliance guide) |
A rule of thumb: the more your organization's standard matters to the right answer, the more structure you need. The more the skill depends on reacting in the moment, the more AI improvisation helps.
A design checklist: 6 questions your critics will ask
- Does the AI character stay in role? Test for drift: does the "angry customer" start coaching the learner after three turns?
- Does the debrief use your standard? Feedback should come from your policy, your sales method or your values, not from an AI model's generic advice.
- Who reviewed it? A subject expert should play every path, or a sample of AI conversations, before launch.
- Does the learner do the thinking? Avoid hints that give away the answer. Ask for the decision before showing feedback.
- What is the evidence of transfer? Plan a before-and-after measurement on the same scenario, plus one behavior indicator the business owner tracks.
- Is it clear that the characters are AI? Under Article 50 of the EU AI Act, people must be informed when they interact with an AI system, from 2 August 2026 (AI Act Art. 50). Also note that the AI Act prohibits emotion recognition in the workplace (Art. 5), so don't score learners' faces or voices for emotion.
How to pilot AI role-play training in 6 weeks
- Week 1: Pick one conversation that costs your business money, such as a refund escalation, a renewal objection or a first feedback conversation for new managers. Agree one indicator with the process owner.
- Week 2: Map 3 to 5 key decisions with a subject expert. Let AI draft the dialogue and the wrong options, and have the expert choose the realistic ones.
- Week 3: Run the scenario as a pre-test with the pilot group.
- Weeks 3–5: Let people practice as often as they like, alongside your normal human sessions.
- Week 6: Run the same scenario again and compare first-attempt decisions. Review the indicator with the process owner.
A modest, honestly measured effect from your own pilot will persuade your stakeholders more than any vendor case study.