What 800 meetings taught our team intelligence agent

The Data Nobody Looks At

Every company records meetings. Almost none of them analyze the recordings.

This is one of the great paradoxes of modern business. We obsess over quantitative metrics, revenue, churn rate, velocity, NPS, while sitting on a goldmine of qualitative data that goes completely unexamined. Meeting recordings are treated as a compliance artifact or a “just in case” backup, not as a data source.

At Moonxi, we’ve been feeding meeting transcripts into Apollo Space’s Team Intelligence agent since mid-2025. By March 2026, it has processed 812 meetings across our internal team and client engagements. The aggregate runtime of those meetings is approximately 680 hours, 85 full working days of human conversation.

No human could analyze 680 hours of conversation. But an agent can. And the patterns it found challenge several assumptions we had about how teams work, how clients churn, and where money gets wasted.

Pattern 1: The Silence Ratio and Overstaffing

The first pattern the agent identified was what we now call the silence ratio: the percentage of a meeting’s duration where a given participant neither speaks nor is directly addressed.

In a well-functioning meeting, every participant has a silence ratio below 70%. They’re actively engaged, asking questions, providing input, reacting to others. Even quiet participants tend to contribute substantively at least 30% of the time.

The agent flagged something interesting: in meetings with high silence ratios (one or more participants above 85%), the project was overstaffed 73% of the time.

Here’s what this looks like in practice. Take a hypothetical weekly sync with six participants. Four are actively engaged. Two sit in silence for the entire meeting, occasionally unmuting to say “sounds good” or “I agree.” Those two people aren’t lazy or disengaged, they simply don’t have enough relevant context or responsibility to contribute. They’re in the meeting because someone put them on the invite, not because the project needs their input.

Across 812 meetings, the agent identified 34 instances of persistent high silence ratios (appearing in 3+ consecutive meetings for the same participant). Of those 34:

• 25 (73%) correlated with teams that were later identified as overstaffed
• 6 (18%) correlated with participants who were about to leave the project voluntarily
• 3 (9%) were misclassified (participants who were legitimately in a listening/learning role)

The financial implication is significant. The average overstaffing signal appeared 4-6 weeks before a manager identified the issue through traditional project management reviews. At a blended cost of $85/hour per engineer, 4 weeks of unnecessary staffing costs roughly $13,600 per person.

Pattern 2: Sentiment Shift as a Churn Predictor

This finding changed how we think about client relationships.

The Team Intelligence agent tracks sentiment across meetings using a composite signal: word choice, speaking pace, interruption frequency, and topic avoidance patterns. It doesn’t measure whether someone is “happy” or “sad”, it measures directional shifts from their baseline.

When a client’s sentiment shifts negative by more than 15% from their 30-day rolling average, that client is at elevated churn risk. The agent found this signal to be predictive 6-8 weeks before churn events.

Let’s walk through an anonymized example.

Client Alpha had weekly check-ins with our team. For months, the sentiment signal was stable, fluctuating normally within a band. In mid-July, the agent detected a shift.

The shift wasn’t dramatic. Client Alpha’s project lead didn’t start yelling or expressing obvious frustration. The changes were subtle:

• Question frequency increased 40%. The client started asking more clarifying questions, suggesting decreasing confidence in the team’s direction.
• Future tense usage decreased. Phrases like “when we launch” and “in the next phase” were replaced by “if this works” and “depending on results.” The language shifted from certainty to conditionality.
• Meeting duration shortened. Meetings that used to run 45 minutes started ending at 30. Not because there was less to discuss, because the client was less engaged.
• Topic avoidance appeared. The client stopped bringing up budget discussions and renewal timing. When you’re planning to leave, you don’t talk about the future.

The agent flagged this pattern on July 22. The Post-Sale Health agent elevated the risk level. Our account manager was alerted.

We had a candid conversation with Client Alpha in early August. They were, in fact, evaluating a competitor. The issues were addressable, they wanted faster turnaround on certain deliverables and more proactive communication about timeline changes. We adjusted. They stayed.

Without the agent’s early detection, we would have learned about the problem when Client Alpha sent the cancellation email. By then, recovery rates are below 20% (per Gainsight’s 2024 Customer Success Benchmark report). With 6-8 weeks of lead time, recovery rates jump to above 60%.

Pattern 3: Meeting Frequency as a Distress Signal

This one is almost embarrassingly simple, but nobody tracks it.

When a project is healthy, meeting cadence is stable. Teams have their regular check-ins, and ad-hoc meetings are rare. When a project enters distress, meeting frequency spikes.

The agent quantified this: a 40% or greater increase in meeting frequency for a given project, sustained over two weeks, correlates with project distress 78% of the time.

“Distress” in this context means one or more of: missed deadlines, scope creep, budget overruns, team conflict, or client escalation.

The mechanism is intuitive. When things go wrong, people call meetings. They need to “align.” They need to “discuss.” They need to “get on the same page.” The meetings themselves are a symptom, not a cure, but they’re a highly visible symptom that traditional project management tools completely ignore.

Here’s the data from our 812-meeting dataset:

Meeting Frequency Change	Projects in Distress	Correlation
0-20% increase	12%	Baseline noise
20-40% increase	34%	Mild concern
40-60% increase	78%	Strong signal
60%+ increase	91%	Near-certain distress

The agent now tracks meeting frequency as a first-class metric and alerts when a project crosses the 40% threshold. This gives managers 1-3 weeks of lead time to intervene before the distress manifests in deliverables.

Pattern 4: Decision Velocity Decay

Here’s a pattern that surprised us: the speed at which teams make decisions degrades over time, and the rate of degradation predicts project outcomes.

The agent measures decision velocity, the number of decisions made per meeting hour. In healthy projects, this number stays relatively stable or increases as the team gets more comfortable. In troubled projects, it declines.

Across our dataset:

• Healthy projects averaged 2.8 decisions per meeting hour throughout their lifecycle
• Projects that eventually failed averaged 2.6 decisions per hour in their first month, declining to 1.4 by month three
• The sharpest predictor: if decision velocity drops below 50% of its initial rate, the project has a 72% chance of missing its deadline

What causes decision velocity decay? The agent identified several correlated factors:

• Stakeholder proliferation. As more people join meetings, decisions slow down. Each additional decision-maker beyond three increases meeting-to-decision time by roughly 20% (consistent with research by Bain & Company showing that for each person added beyond seven in a decision-making group, decision effectiveness drops by 10%).
• Ambiguity accumulation. When early decisions are deferred rather than resolved, the ambiguity compounds. Later meetings spend more time revisiting unresolved topics from earlier meetings.
• Risk aversion growth. As projects progress and stakes increase, teams become more cautious. Decisions that would have been made in five minutes during the kickoff take 30 minutes by month three.

The Team Intelligence agent now flags decision velocity decay as a project health indicator. When velocity drops below a configurable threshold, it alerts the project lead and suggests specific actions: reduce meeting attendees, resolve deferred decisions, or escalate blockers.

Pattern 5: The Budget Conversation Disappearance

This is the most counterintuitive finding: when a client stops talking about budget, you’re about to have a budget problem.

In healthy client relationships, budget comes up naturally. “Are we tracking to budget?” “We have room to add this feature within the current scope.” “Let’s discuss the investment for Phase 2.” These are signs of a client who is engaged with the financial relationship and planning to continue it.

When budget discussions disappear from meetings, it means one of two things:

1. The client has decided the budget is a sunk cost and is mentally detaching from the engagement
2. The client is having internal budget conversations that don’t include you, usually about cutting or replacing your services

The agent identified 11 instances of budget conversation disappearance across our dataset. In 8 of those 11 (73%), the client either reduced scope or terminated the engagement within 90 days.

The mechanism is psychological. Budget conversations are forward-looking, they imply continuity. When a client stops having them, it’s because they’ve stopped assuming continuity.

Pattern 6: Cross-Team Communication Gaps

The final pattern is structural rather than sentiment-based.

The Team Intelligence agent maps communication networks: who talks to whom, how frequently, and through what channels. It identified that when a person who previously bridged two teams reduces their cross-team communication by 50% or more, those teams begin to diverge in priorities within 2-3 weeks.

This is the “key person” risk that every organization knows about but nobody monitors. When the person who connects engineering to sales, or product to operations, disengages (due to burnout, role change, or departure), the teams they connected start operating in silos.

In our dataset, 6 instances of key-person communication reduction led to 5 instances of measurable team divergence (conflicting priorities, duplicated work, or misaligned timelines).

The agent now tracks communication topology and flags when key bridges weaken. This gives leadership time to either re-engage the bridge person or establish alternative communication pathways before the teams diverge.

What We Do With These Patterns

Identifying patterns is step one. Acting on them is step two. Here’s how Apollo Space’s agent ecosystem uses these signals:

Churn prevention loop. Sentiment shift detected by Team Intelligence → Risk elevation by Post-Sale Health agent → Account manager alerted with context and recommended actions → Follow-up meeting scheduled → Meeting Digest agent tracks whether the recovery conversation produces commitments → Team Intelligence monitors whether sentiment recovers.

Project health loop. Meeting frequency spike or decision velocity decay detected by Team Intelligence → Project lead alerted → Agent suggests specific interventions (reduce attendees, resolve deferred decisions) → Subsequent meetings are analyzed to confirm whether interventions worked.

Staffing optimization loop. Silence ratio signal detected → Overstaffing assessment triggered → Agent cross-references with project budget and timeline → Recommendation generated (reallocate person to higher-impact project) → Manager reviews and decides.

These loops run continuously. They don’t wait for quarterly reviews or annual retrospectives. The signals are captured in real time and acted on in days, not months.

The Ethics of Meeting Intelligence

We take data ethics seriously, and this section is necessary.

Meeting intelligence raises legitimate concerns. People should know their conversations are being analyzed. The patterns should inform team-level and project-level decisions, not individual performance reviews. The agent should surface insights, not surveillance.

Apollo Space’s approach:

• Transparency. All participants are informed that meetings are recorded and analyzed for operational patterns.
• Aggregation. Insights are presented at the team and project level. Individual speaking patterns are not shared with managers unless the individual opts in.
• No behavioral scoring. We don’t rate individuals. The silence ratio pattern, for example, is used to assess project staffing, not to evaluate quiet participants.
• Data retention policies. Transcripts are processed and then the raw data is retained only as long as needed. Derived patterns are anonymized.

The goal is organizational health, not individual surveillance. The distinction matters, and we enforce it in the product.

800 Meetings Later

The most important thing 800 meetings taught us is this: meetings are a data source, not a time sink.

The problem was never that we had too many meetings. The problem was that we treated meetings as ephemeral events, things that happened and then were forgotten, unless someone happened to take good notes.

Every meeting contains signals about project health, client satisfaction, team dynamics, and operational efficiency. Those signals were always there. We just couldn’t see them because no human can process 680 hours of conversation and identify patterns across hundreds of data points.

The Team Intelligence agent can. And the patterns it finds, churn warnings, overstaffing signals, distress indicators, decision decay, are the kind of insights that used to require a seasoned manager with decades of experience and a gut feeling.

Now they’re quantified, trackable, and actionable.

That’s not replacing management judgment. It’s augmenting it with data that was previously invisible.