A family therapist has a session with a parent and their teenage child, and has to read two people at once, not one. A crisis counselor facilitates a group session where every participant brings a different history to the table. A tax counselor is mid-call when a spouse picks up the other line. A sales rep is in a meeting with three stakeholders who all want something different.
These are, by a wide margin, the hardest conversations someone will ever have to navigate. Multiple people, competing needs, dynamics shifting in real time, no script to fall back on. And until now, they've also been the conversations no one could train for. You could read about de-escalating a family in crisis. You could not practice it — not really — until you were doing it live, with an actual family, where the cost of getting it wrong wasn't a bad grade but a person in front of you.
Multi-party simulations change that. For the first time with ReflexAI, the highest-stakes conversations, which are the ones with the most people, the most volatility, and the least room for error, can be rehearsed before they happen for real.
Every simulation on ReflexAI's Prepare platform, until now, ran one trainee against one AI persona. But the conversations with more than one person in the room are exactly where the real difficulty lives. Multi-party simulations are built for exactly that: the first time those conversations can be practiced using ReflexAI, instead of learned the hard way, on the job, with someone real on the other end.
One conversation, two personas, no script
With multi-party simulations, a simulation author sets up two AI personas, each with their own role, personality, and goals, and drops the trainee into a live voice conversation with both of them at once.
The key word is live. This isn't a scripted back-and-forth where the personas take turns on cue. While the author of the simulation can calibrate how much pushback or assertiveness a persona has, the AI decides who talks when, the same way a real conversation would unfold: one persona might dominate while the other chimes in occasionally, one might defer to the other, one might jump in mid-thought. The trainee has to manage that dynamic in real time, while deciding who to address, handling interruptions, and noticing when one voice in the room needs more attention than it's getting.
That's a fundamentally different skill than handling a single conversational partner, and it's one a lot of our customers have been training around rather than training for, like running two simulations side by side, bolting disclaimers onto sim instructions, stitching together sequential journeys by hand. Multi-party simulations are built to close that gap directly.
Where multi-party conversations show up in high-stakes work
Multi-party training scenarios show up everywhere once you start looking for them: family and couples therapy, group facilitation in behavioral health, mediated financial conversations, multi-stakeholder sales calls, crisis lines where more than one person is on the line at once. It's been one of the most consistently requested capability expansions across our customer base, and it's also core to our partnership with Google.org, which is focused on using multi-party training to scale behavioral health support.
What's included in multi-party simulations
In ReflexAI, teams can now train for the hardest, highest-stakes conversations they'll ever face, in realistic, fully-scored voice simulations:
- Author-controlled dynamics. Control how active or passive each AI persona should be, how much the personas should push back, etc.
- Full, automatic scoring. Multi-party simulations are fully scored through your company’s rubric pipeline.
- Smart persona interactions. AI personas answer each other's questions when it makes sense, instead of waiting for the trainee to direct every exchange. Conversations flow the way they actually would in the room.
- Smooth voice. Significantly low latency, so the back-and-forth feels responsive instead of stilted.
- Spanish support. Multi-party simulations run end-to-end in Spanish, matching the language support already available for one-on-one sims.
What’s next for multi-party training
Currently multi-party simulations support two personas per simulation, but it’s early days yet. We're actively working through the harder design questions that come with more: how personas should hold asymmetric knowledge about each other, how scoring adapts to a multi-speaker transcript, how far we can push turn-taking before it stops feeling natural. Hand-offs are next on the roadmap, where for example, a therapist calls a life line on behalf of their patient, then hands the phone over and drops off the call, leaving the patient to continue alone.
If you want to train for the hardest customer-facing conversations with multiple people in the room before you encounter them live, schedule a demo today.