When AI should stay quiet in a meeting (and how we decide)

The default for meeting AI is silence

A good meeting AI talks less than a careful junior teammate would. The whole job is to add value without breaking the conversation, and breaking the conversation is the easiest mistake an AI can make. Every spoken sentence is a tiny interrupt that pulls four or five people out of the thought they were holding.

So we treat silence as the default and speech as the exception. Most products do the opposite: they try to look helpful by saying something every few minutes, and the room slowly learns to tune them out. Once an AI is tuned out, even its good contributions get ignored, and the whole tool turns into background noise.

The five moments AI should stay quiet

These are the moments where speaking, even with a correct answer, costs more than it gives. We hard-code most of these into the default behavior and let teams tune the rest.

1. Emotional beats

If someone just shared bad news, pushed back on a peer, or said something that landed heavy, the room needs a beat. The AI sits this one out. No "I noticed tension, want me to summarize the disagreement?" That's not helpful, that's a robot reading the room and getting it wrong.

Detection here is rough but workable: a sharp tone change, a long pause, an unusually short reply from someone who normally talks more. When two of those land together, the AI shuts up for the next two to three minutes regardless of what else is happening.

2. Mid-sentence

A human teammate would never finish someone else's sentence. The AI shouldn't either. We wait for a real handoff: a clear end-of-thought, a direct question, or a quiet beat where someone is visibly looking for input. If the pause is under one second, that's almost always a person reaching for the next word, not an opening.

This sounds obvious. It's the single most common failure mode in voice AI today. The model finishes processing what it heard, the latency budget says "now," and it talks over the speaker. The fix is treating short pauses as part of the speaker's turn, not an invitation.

3. Free brainstorm

When the team is in divergent mode (throwing out half-formed ideas, building on each other, riffing) the AI is the worst participant in the room. It can't follow the implicit threads, it can't read the half-jokes, and any "helpful" contribution pulls the group back into linear thinking. So during brainstorms we drop the AI to listen-only mode.

How we detect it: short turns, lots of switching speakers, lots of "what if" and "yeah and what about." When that pattern shows up, the AI stops offering things and just logs the ideas as they fly past. It can resurface them in the synthesis phase, when the team actually wants structure.

4. Sensitive topics

HR, legal, personal stuff, comp, performance, anything that names a specific person in a critical way. The AI listens, transcribes if the team allows it, and never volunteers a thought. There is no upside to a model offering an opinion on whether someone should be on a PIP. The downside is everything from a weird vibe to an actual legal problem.

We keep a simple block list for this: trigger phrases, name-plus-verb patterns ("we should let Sarah"), and any meeting tagged in the calendar as 1:1, HR, legal, or review. When any of those match, the AI defaults to deep silence: no proactive contributions, just quiet logging if consent is set.

5. Low confidence

If the AI's confidence in what it's about to say is below the team's threshold, it stays quiet. A wrong "fact" stated confidently in a meeting can survive in the team's memory for weeks, and the cleanup cost is enormous. We'd rather say nothing than say something we'd have to walk back.

The threshold is something teams should set, not a single global number. A sales call probably wants a higher bar (the cost of an embarrassing mistake is high). An internal design review can tolerate more guesses. The system exposes the knob; the team picks the setting.

The three moments AI should speak

Speak only when the room is already reaching for what you have. Otherwise, listen. Your value isn't measured by airtime, it's measured by how often you save the team from a re-do.

The flip side of the quiet rules: there are clear moments where the AI absolutely should jump in, because not jumping in costs the team something real.

1. A fact the team is visibly looking for

Someone says "what was our churn last quarter?" and three people start scrolling Notion. That's the moment. The AI says the number, cites the source, and the meeting moves on. Total cost: two seconds. Total saved: a five-minute detour while everyone hunts for a doc.

The signal here is specific: a direct question that has a knowable answer, plus visible search behavior (typing, scrolling, "let me find it"). Without the search behavior, the AI waits. The team might be using the question rhetorically.

2. A decision that needs a clarifying question

If the team is about to commit to something and a critical piece is missing (no owner, no deadline, no scope), the AI surfaces the gap once and then stops. "Just to confirm, who's owning this one?" That's it. Not a paragraph, not a recap, one targeted question.

This is where most "helpful" AI overshoots. It tries to summarize the decision, list pros and cons, suggest alternatives. The team didn't ask for any of that. They asked for a missing piece, implicitly, by almost moving on without it.

3. An action item nobody named

The conversation has shifted but a clear action item just got buried. The AI captures it and reads it back once, naming the owner if it's obvious from context. "Putting this on the list: Maya to send the updated pricing deck by Friday, correct?" Confirm, log, move on.

This is the highest-value speech moment in most meetings. It's the difference between a meeting that decided something and a meeting that vaguely talked about deciding something. The AI doesn't even need to be right every time. As long as it surfaces the candidate, the team can correct it in one beat.

The rule we use to draw the line

The whole framework collapses into one question: would a careful new hire speak here? If yes, the AI can. If no, the AI stays quiet. New hires don't interrupt to summarize. They don't volunteer trivia. They don't offer opinions on personnel decisions. They ask a clarifying question when a decision is about to ship half-baked. They name the missing action item when nobody else did. They look up the number when the room is hunting for it.

That heuristic is doing a lot of work. It's why we resist the temptation to give the AI a "personality" or have it offer constant micro-contributions. Personality is a distraction. The job is to be useful and forgettable.

What teams should actually configure

We default to the rules above, but the right answer depends on the meeting type. A few knobs worth exposing to the team:

• Confidence threshold: how sure the AI needs to be before stating a fact. Higher for client calls, lower for internal exploration.
• Minimum silence window: how long a pause must be before the AI counts it as an opening. Default one to two seconds. Bump it for cultures that talk over each other less.
• Topic blocks: meeting types or keywords where the AI is fully silent. HR and legal are the universal ones. Add your own.
• Speak budget: a soft cap on how often the AI can speak in a meeting. We default to once every five to ten minutes, and let teams tighten it.
• Per-person mute: any participant can silence the AI for the next stretch of conversation. The AI keeps logging, just doesn't speak.

Most teams set these once and forget them. The product's job is to make those defaults sane enough that the configuration step is optional.

The quiet AI still does most of the work

Here's the part that's easy to miss: when a meeting AI is well-behaved, it looks like it's barely doing anything during the call. The room runs smoothly, decisions land, and there's maybe one or two spoken moments where the AI saved a beat. Then the meeting ends and the artifacts show up.

Full transcript with decisions highlighted. Action items with owners and dates, in your team's tracker. Research notes for the topic that came up halfway through. A short recap that goes to the people who couldn't make it. The work was happening the whole time. It just wasn't loud.

This is the part most demos hide, because silence doesn't demo well. But it's the part that actually matters to a real team running real meetings. (For how this fits into the live experience, see the use cases.)

Where relly draws the line

relly ships with these rules as the default. Silence is the baseline. Speech is reserved for the three moments above. Every speak event is logged so teams can see when and why the AI jumped in, and tighten the rules if they want it to talk less. Most teams end up loosening the budget after a week, once they trust the silence isn't laziness.

If you want to see what a well-behaved meeting AI feels like, early access is open. The first time the AI sits quietly through an emotional moment that any other tool would have stepped into is the moment most teams understand the design choice.