Here's a story I didn't expect to be telling a year ago. We shipped four AI coaching personas — Duke the hype man, Annie the straight shooter, George the comedian, Marie the professor. We stress-tested the tonal range in 30+ customer trials. We thought we were done.
We were wrong.
The first call came from Victoria, a VP of Sales at a mid-market SaaS company in Austin. "I love George," she said. "My reps love George. But my reps love him because he sounds exactly like me. And you didn't know that."
The realization
The thing about four personas is that they're still a menu. A rep picks one and — hopefully — finds the best fit among four choices designed by three founders in Costa Rica and Seattle. That's better than a generic corporate coach. It's still a compromise.
What Victoria was asking for, without quite saying it, was this: "I've spent eight years developing a coaching voice. It works for my team. I want the AI to sound like me — specifically me — not like the closest-of-four approximation."
Multiply that by every manager who has ever built a real coaching practice, and the roadmap writes itself.
Persona selection is a menu. Persona authoring is a platform.
What we built
The Voice & Vibe Studio is a manager-authoring surface that takes three kinds of source material and produces a custom AI coach:
- Slack exports — channel posts and DMs from the manager's coaching context. Ideally 12+ weeks of history.
- 1:1 notes — whatever the manager writes before, during, and after weekly 1:1s. Even bullet points help.
- Coaching recordings — transcribed, ideally 10+ sessions. These carry the most tonal signal.
Everything stays in the tenant. We don't train shared models on Victoria's Slack threads. The persona is hers, it's isolated to her org, and she can redact any contributor's material at any time.
Extracting the voice
The interesting engineering problem isn't training a fine-tune — that part is now pretty boring. The interesting problem is what to extract.
A manager's coaching voice has, in our research, four dimensions that matter:
Dimension 1: Tone baseline
Directness vs. warmth vs. playfulness. Measured through sentence length, hedging rate, emoji frequency, and opening/closing patterns. Victoria's baseline: "Direct but warm. Opens with acknowledgment, closes with accountability." George (our Comedian) would never open with acknowledgment.
Dimension 2: Characteristic vocabulary
Every manager has phrases that show up 10× more in their writing than baseline. Victoria's: "let's name it", "what's the real constraint", "I want to see the receipts", "small, fast, honest". These aren't jargon — they're signals of identity.
Dimension 3: Escalation pattern
How does the manager move from soft nudge to direct ask? Victoria does "three soft nudges → one direct message → calendar hold." Different managers do "immediate public call-out" or "never escalates, just re-explains." The AI needs to mirror this — not invent its own.
Dimension 4: Feedback motif
The recipe the manager uses when giving hard feedback. Victoria's: "Name the thing → suggest the move → offer the partnership." That three-beat structure is what makes her tough messages land without bruising.
Why this is the category move
Here's the strategic argument, stripped down:
- Every other AI coaching product is a menu. You pick one of N.
- Voice & Vibe is a platform. You build one of infinity.
- Menus commoditize. Platforms compound.
The more a customer uses Voice & Vibe, the more switching cost accumulates — because the coach is theirs, not ours. Every week the manager refines it, and every week the model of her coaching practice gets richer. Taking that away at renewal time is taking away her team's muscle memory.
That's not defensibility by lock-in. That's defensibility by genuine value.
What Victoria said, three months later
"My reps don't know where Victoria ends and the AI begins. That's either the creepiest thing I've ever done or the most powerful. I think it's both. I think that's the point."
Yep. That's the point.