You’ve sat through the workshop on giving feedback. The slides were fine, and the breakout-room role-play was awkward. Then, three weeks later, a new manager has to tell a direct report their work is slipping, and they freeze or avoid the conversation altogether.
Watching a good conversation isn’t the same as having one.
One missing ingredient is practice. And calendars rarely make room for enough of it.
Where AI Roleplay is Finding a Place
Much of the visible activity around AI roleplay has emerged in sales training. Second Nature, for example, reports a 34% reduction in ramp time and the ability to scale practice 6x. Quantified reports 40% faster time to readiness and 6x more practice. These are vendor-reported outcomes rather than independent research, so they are best treated as directional.
The use case is expanding beyond sales, and the pattern is worth watching. AI roleplay can create practice opportunities for coaching conversations, difficult feedback, conflict, interviewing and other situations where the behavior matters but a human practice partner isn’t always available.
But the technology isn’t the hard part. The harder question is how you design the practice.
Start with Behaviors, not Soft Skills
“Be a better coach” is difficult to measure. “Asks two open questions before offering advice” is not. For each conversation type, identify three to five observable behaviors. For example:
- Asks inquiry-based questions before offering a view
- Acknowledges the other person’s perspective when they push back
- Sets a clear boundary without blaming
- Summarizes what they heard
- Ends with a specific next-step commitment
This gives the AI something meaningful to evaluate and gives the learner something concrete to improve. It also lets you vary the depth of practice. A new manager may need repeated attempts with reflection after each one. An experienced HR business partner handling sensitive conversations may benefit from AI practice followed by human coaching and calibration.
Design the Conversation, not Just the Character
A useful simulation needs more than a realistic persona. Three elements matter.
1. A scenario arc: context, stakes, trigger, pushback and resolution.
For example, a manager is meeting an employee who has missed several deadlines. The stakes are rising because a client is affected. The employee becomes defensive. The goal is to reach a clear, constructive next step.
2. Choice-driven reactions: the learner’s language changes what happens next. A blaming statement might increase defensiveness; a genuine acknowledgement might create an opening for dialogue.
3. A behavioral rubric: the simulation evaluates the behaviors you actually want to develop.
The basic loop is simple:
That is where AI roleplay can be useful. It creates a practice environment where someone can have the same difficult conversation more than once and try something different. AI roleplay should be viewed as an additional practice environment, not a substitute for human interaction, coaching or real-world experience.
Give the AI a Persona Worth Practising Against
Persona design is where realism starts to matter. A useful persona card should define:
- Role and relationship to the learner
- Goals and constraints
- Emotional state
- Likely objections
- What good sounds like
- What causes the conversation to deteriorate
For example: You are Dana, a senior analyst who reports to the learner. You’re capable but disengaged following a reorganization. You’re skeptical of feedback and may say, “I’ve been doing this for years.” Stay in character. Soften only when the learner asks genuine questions and acknowledges your perspective. Don’t offer solutions unless prompted.
The persona should also have clear escalation rules. If a learner raises a medical issue, harassment concern or legal threat, the simulation should pause and route the situation to the appropriate human or organizational process. AI should not improvise its way through policy-sensitive situations.
Combine AI Feedback with Human Judgment
AI feedback is immediate and available whenever the learner needs it. That makes it useful for repeated practice and micro-coaching.
Human coaches bring something AI feedback can’t reliably provide: contextual judgment, calibration and accountability. Research into AI coaching and AI-mediated roleplay continues to show both promise and limitations, particularly around the nuances of human interaction.
A practical rhythm is:
Ask the AI to make its feedback specific. For example: “Score me 1 to 5 on each rubric dimension. Quote the phrases that influenced the score and give me three alternative ways I could respond.”
Then run the scenario again.
The first attempt establishes a baseline. The second focuses on one micro-skill. The third introduces a variation, such as a different persona mood or stronger resistance. That gives you a way to start testing whether the skill carries into a different situation.
Start with a Focused Pilot
You don’t need a library of 50 scenarios to prove the value! Start with:
- Three to five high-value scenarios
- Clearly defined behavioral rubrics
- SME testing of each persona
- Escalation and privacy rules
- Learner communications
- A measurement plan
Measure what matters: improvement in rubric scores, learner confidence, time to demonstrated competence, manager adoption and quality-assurance results.
Then follow up. A two-week check can provide an early indication of whether the behavior is carrying beyond the simulation.
The Opportunity Isn’t the AI. It’s the Learning Design.
AI roleplay makes practice more accessible. But a realistic conversation alone does not make a learning experience effective. The real work is connecting the right business situation, target behavior, scenario design, rubric, feedback model and governance into one coherent experience.
That is where Mediant Labs can help. We bring learning strategy, instructional design, learning technology and AI expertise together to design practice around the capability an organization actually needs, rather than starting with a tool and looking for a use case.
Because the goal isn’t simply to give people more conversations with AI. It’s to help them handle the real conversation better when it matters.
