Pulse: Our first AI as a Meeting Participant

The first time we tried speaking to an AI instead of typing, several things jumped out immediately:

• The awkward lag while it tried to guess whether we were done talking
• The way any new sound was treated as an interruption
• The total inability to distinguish between a human speaking, a child in the background, or a dog barking

These aren't just implementation bugs. They're symptoms of a deeper assumption baked into most AI systems:

AI is treated as a simple, turn-based dialog partner.

You type/speak.
You hit enter/pause.
The AI responds.

That abstraction works just well enough for text — but it breaks down completely when you move into real human conversations, especially spoken ones with multiple participants.

Meetings are not dialogs.
They are living systems.

People talk over each other.
They pause rhetorically.
They agree implicitly.
They circle decisions without naming them.
They miss key moments — not because they aren't smart, but because cognition is already overloaded.

We don't believe the solution is better prompting. We believe the solution is changing the role of AI entirely. Instead of an AI you "talk to," we envision an AI that:

• Listens continuously
• Tracks what's happening now
• Understands conversational dynamics
• Knows when not to speak
• Intervenes briefly, intentionally, and safely

In other words:
AI as a meeting participant.

Not a replacement for humans.
Not a dominant voice.
But a quiet, attentive presence that helps the group think more clearly.

This vision comes to life in JoinIn Pulse, our first product. Pulse joins your meeting the same way a human does — as a participant.

Once there, it listens in real time and maintains a continuously updated understanding of the conversation:

• A live, speaker-attributed transcript
• A sense of what topic is being discussed right now
• Signals that a decision may be forming
• Open questions that haven't been resolved

Crucially, Pulse does not require constant prompting. It maintains a running internal model of the meeting so that help is available the moment you need it.

One of the most common failures in meetings isn't lack of intelligence — it's loss of context. You step away for a minute, get pulled into another tab, or rejoin just as the conversation pivots.

Pulse addresses this with Catch-Up Mode. Pulse is always summarizing — continuously updating a concise view of the last 60–180 seconds. If it's not updating frequently enough, you can manually trigger a refresh:

• A short, bounded recap (never a wall of text)
• What's currently being discussed
• Any decision that seems to be forming
• Any open question still unresolved
• A single sentence you could say right now to re-engage constructively

This isn't a meeting summary — it's a re-entry ramp designed for speed, clarity, and action.

Even when you're fully present, meetings often stall at transition points: "Are we aligned?" "Who owns this?" "Is this actually decided?" "Should we park this?"

Pulse helps here with Next-Turn Mode. Rather than generating paragraphs, Pulse proposes two candidate sentences you could say next — each intentionally brief and labeled with its conversational intent:

• Clarify
• Align
• Decide
• Delegate
• Park

You're always in control. You can copy the sentence, post it to chat, or say it yourself. Pulse doesn't replace your voice — it sharpens it.

Meetings frequently hinge on facts: "Didn't we decide this last quarter?" "What does the data actually say?" "Is that number still accurate?" Pulse watches for moments where a fast lookup could unblock progress, and can proactively perform meeting-aware search:

• Short, grounded answers
• Framed explicitly for the current discussion
• Accompanied by a single "meeting-ready" sentence you can say or post
• Answers right now, not action times or transcript analysis.

This is intentionally metered and bounded — because the goal isn't to flood the meeting with information, but to unblock it.

The hardest part of our vision is also the most important: When should an AI be allowed to speak? Pulse is silent by default. When it does speak, it does so under strict, user-controlled guardrails:

• One sentence only
• No more than once every 45–90 seconds
• Only when it has high confidence the intervention fits an allowed intent
• Immediate, non-negotiable STOP NOW control

Users can explicitly grant Pulse a short Release to Speak window — similar to letting a junior teammate jump in briefly. During that window, Pulse may speak without another click, but always within the agreed constraints. This isn't about autonomy — it's about earned trust.

Pulse is designed to get better — visibly and responsibly. Every suggestion, post, or spoken intervention can be rated:

• 👍 Good
• 👎 Not good
• 🚩 Flag

With lightweight, structured feedback like: Wrong tone, Missed context, Too long, Bad timing, or Risky to say.

This feedback isn't buried. It's a core part of the experience — because building an AI that participates in human conversations requires continuous calibration.

Meetings are where decisions happen — or fail to. They are cognitively demanding, socially complex, and chronically underserved by today's AI tools. We believe the future of human-computer interaction isn't about better prompts or faster responses. It's about AI that understands context, respects social dynamics, and helps groups think together.

Pulse is our first step toward that future. And this is only the beginning.