This is how a hard conversation gets rehearsed before it happens.
Frank puts a voice counterpart on the other side: one that answers back, pushes, and can end the meeting. This page walks through what happens in a session, where the scenario comes from, how the counterpart behaves, and what each person gets at the end.
What happens in a session.
Three parts. The scenario is written beforehand from your company’s context, the conversation happens live and by voice, and the report arrives at the end, only for the person who practiced.
- 01Before: the scenario is your company’s
- Frank reads the documents you already have, such as the culture handbook, the discount policy or the sales playbook, and proposes conversations with your products, your objections and your numbers inside them. You approve the ones that go in.
- 02During: live voice roleplay
- You talk to a counterpart that answers back, pushes, leaves the silence open and can end the meeting. There is no coach and no metrics on screen: while the session lasts, Frank is just the other person.
- 03After: a private report
- Written in words, with the key moments marked on the timeline and what to practice next time. The person who practiced sees it. The organization sees aggregated progress.
A session runs 5 to 10 minutes, works in the browser with nothing to install, and is available in Spanish and English.
See the conversations it trainsProgram, story, session.
Frank isn’t a course made of loose modules. A program groups stories, a story is practiced across several sessions, and the counterpart arrives at each session carrying what happened in the last one.
- 01Program
- What the leader sets up for the coming weeks. It groups stories around one goal: commercial negotiation, feedback to the team, conflict handling.
- 02Story
- A situation that moves forward chapter by chapter and can branch depending on what happened: if you held the price, the next session is the technical close; if you gave up margin, it’s winning the term back.
- 03Session
- One voice conversation, start to finish. The counterpart doesn’t start from zero: it remembers what you told it, what you promised and what was left open.
How a program is put together, from which documents and in how long, is explained on the programs page.
See how a program is builtA counterpart with character and memory.
It carries an emotional register that shifts with the conversation. It isn’t a script with branches: it’s a stance that moves depending on what you say and how you say it.
Give it two solid arguments in a row and it softens its stance.
Once it has conceded a point, it can’t harden again without a new reason.
It leaves the silence open and doesn’t fill it for you.
It comes in over your last syllable when the scenario calls for it.
It can warn it’s about to close and, if you don’t react, end the meeting.
Nature of the scenario
Every scenario has a nature. It isn’t a scale: they are different conversations, and a person can hold one well and not the other.
- Routine
- A monthly one-on-one, a check-in with a client who is doing fine. The counterpart cooperates, but gives nothing away.
- Tense
- A client pushing for a discount, a team receiving bad news. The counterpart holds the objection.
- Conflict
- A cofounder questioning the direction, a dismissal. The counterpart can raise its voice and cut the meeting.
Latency is part of the exercise.
If the AI lags, the pressure breaks and the exercise stops working. That’s why Frank’s voice engine is its own persistent, low-latency service, not a function that spins up with every message.
- < 1.5 s
- Sonic 3
- ES / EN
- Browser
Corporate sign-on and integrations sit in the Enterprise plan and are scoped with you in the demo. See plans
The report arrives at the end, and belongs to the person who practiced.
Private, written in words, with the key moments marked on the timeline and a concrete recommendation of what to practice. It doesn’t tell you whether you were right. It tells you how you said it.
You held the first part well. When price came up you sped up and didn’t slow down again until the close. Next time, try leaving a silence before giving the number.
- Key moments
- Where it interrupted you, where you took your time coming back, where you sped up. Each one with its timestamp, so you can listen to it again.
- Compared with yourself
- The reference is your own baseline, not an average or an external ideal. An accent or an unusual timbre isn’t noise: it’s the starting point.
- Reviewed before it’s shown
- A validator agent reviews the text before you see it: that it describes what happened, that it doesn’t judge the person, and that it doesn’t claim more than the audio allows.
- Re-assessment on a fresh scenario
- When it’s time to measure progress, the re-assessment runs on a comparable scenario you haven’t seen. Never on the one you just practiced.
- If the audio isn’t enough
- When the signal doesn’t allow a confident measurement, the report says so instead of reporting a doubtful number.
What it measures, and what it doesn’t.
Observable conversational behavior and descriptive acoustic signals, compared against that same person’s baseline. No diagnosis of emotions, no prediction of job performance, no automated employment decision. That limit is a design decision, not a temporary restriction.