Skip to Content
ComponentsAI Conversation

AI Conversation Component

An AI Conversation component lets participants talk to an AI-driven persona — a vendor, a stakeholder, an interviewer, a difficult client — as part of the simulation. Conversations can happen by typing, or as a live voice call, and the outcome can feed directly back into the simulation’s metrics and data.

Setting up a persona

Each AI Conversation is built around a persona:

  • Name, role, background, and personality — who the AI is playing and how they should come across.
  • Goals and constraints — what the persona is trying to achieve, and any hard limits on what they’ll agree to (e.g. “cannot offer a discount below $10/unit”).
  • Custom instructions — free-text guidance for how the persona should behave beyond the structured fields above.

Giving the AI context

An AI Conversation can be wired up to see relevant parts of your simulation before it talks to a participant, via context bindings:

  • Pull in a metric’s current (or initial) value.
  • Pull in a data resource — an entire table, a single row, a column, or a single cell.
  • Inject a static value you define directly.

You control when that context is captured:

  • Once at start — the AI sees a snapshot when the conversation begins and doesn’t re-check it mid-conversation.
  • Every turn — the AI re-reads the current simulation state before every reply, so it reacts to changes happening elsewhere in real time.
  • On demand — captured once, similar to “once at start,” for conversations that don’t need live updates.

Conversation modes

Every AI Conversation runs in one of three modes, which determines how it ends:

  • Roleplay — the conversation ends when the persona itself reaches a decision (for example, a negotiation concluding with terms agreed). There’s no manual “end” button — it’s driven entirely by how the roleplay plays out.
  • Evaluation — the goal is a graded outcome, not a decision. The conversation stays open as long as the participant wants to keep talking; they end it themselves with an “End & Evaluate” button when they’re ready to be scored.
  • Hybrid — completes via either path, whichever happens first: the persona reaching a decision, or the participant manually ending it.

Turning the conversation into results

Two independent things can come out of a conversation:

1. Structured outputs (decisions)

You define a set of output keys the AI is allowed to produce once a decision is reached (e.g. dealApproved, finalPrice). Each key can be wired to:

  • A metric — set it, increase/decrease it, multiply it, or apply a percentage change, with optional min/max bounds and change limits so the AI can’t move it further than you allow.
  • A data resource column — update a specific column on a specific row (or whatever row the conversation is already bound to).

This is what lets a conversation directly move the simulation forward — e.g. a successful negotiation actually changes the budget metric, or an interview decision updates a candidate’s status in a data table.

2. Evaluation scoring

If evaluation is enabled, you define a set of criteria (each with a name, a weight, and an optional description clarifying what “doing well” looks like) and a passing score. Once the conversation ends, it’s automatically scored against these criteria and produces:

  • An overall score and pass/fail result.
  • Feedback the participant can see.

That score, and whether it passed, can then be used exactly like any other event in the simulation — to unlock a round or task, adjust a metric, or trigger a notification, scoped to a specific conversation and a specific score threshold you set (e.g. “unlock the next round only if this conversation scores above 85”).

Retries

Each AI Conversation has a configurable number of retries — how many times a team may restart the conversation from scratch after it ends. A fresh attempt wipes the prior transcript and drops the previous evaluation, so it’s judged on its own merits.

Voice / call mode

Optionally, an AI Conversation can support a live voice call instead of (or alongside) typed chat:

  • Participants speak naturally; their speech is transcribed automatically.
  • The persona’s reply is read aloud in one of several available voices.
  • Participants can interrupt the AI mid-reply just like a real call, mute themselves, and check their microphone before starting.

Tips

  • Only give the AI write access to the metrics and data resource columns you actually want it to influence — the AI can’t touch anything outside the outputs you’ve explicitly configured.
  • Use evaluation mode for open-ended practice (e.g. “have a negotiation conversation and see how you did”), and roleplay or hybrid when a specific decision needs to actually happen for the simulation to move forward.
  • Write clear criteria descriptions — they directly guide how consistently the AI grades different participants.
Last updated on