Bi Usage Monitor

Usage Pattern Monitor

Tracks product usage at the account level, detects meaningful changes in behaviour, and classifies trend direction against segment benchmarks. Part of the Book Intelligence foundation layer -- feeds data into the Health Score Engine, Risk Signal Detector, and Expansion Signal Detector.

This is a detection and surfacing skill. It observes, computes, and alerts. It never contacts customers or initiates any external action.

When to Run

• Scheduled: Daily for all accounts. Generates alerts only when a meaningful pattern change is detected -- not on every run
• On-demand: When a human or downstream skill requests current usage profile for a specific account
• Triggered: When product analytics reports a significant event (e.g., usage spike or feature-gating hit)

Core Execution Logic

Step 1: Pull Current Period Usage Data

For each account, pull the following from the product analytics platform:

Step 2: Compare to Account Baseline

For each metric, compute the account's own 90-day rolling average as the baseline. This captures the account's normal behaviour, independent of segment benchmarks.

Calculate the delta: current_period_value - 90_day_rolling_average

Express as a percentage change: (delta / 90_day_rolling_average) * 100

Step 3: Compare to Segment Benchmark

Pull segment benchmarks from the shared benchmark dataset (see references/usage-benchmarks.md). Segment is defined by tier, company size band, and contract age cohort.

For each metric, classify the account's position relative to the segment:

• Above 75th percentile: Strong adopter
• Between median and 75th: Healthy adopter
• Between 25th and median: Below-average adopter
• Below 25th percentile: Underperforming

Step 4: Classify Trend Direction

For each metric, classify the trend based on the account's own baseline:

Step 5: Detect Pattern Types

Scan for these specific patterns, which carry different urgency and interpretation:

Gradual Decline

• Definition: 3+ consecutive declining measurement periods (daily or weekly depending on metric)
• Significance: Sustained disengagement. Not a blip -- something structural has changed
• Alert urgency: High

Sudden Drop

• Definition: >25% single-period decline from prior period
• Significance: An event happened -- product issue, stakeholder change, competitive displacement, or data error
• Alert urgency: Critical (verify data integrity first)

Adoption Plateau

• Definition: Usage stable but below segment median for 60+ consecutive days
• Significance: The account has settled into underutilisation. They are using the product but not realising full value
• Alert urgency: Medium (feeds into Product Adoption pillar)

Breakout Growth

• Definition: >30% above 90-day baseline sustained for 2+ measurement periods
• Significance: Expansion signal. New use case, new team onboarding, or champion-driven adoption push
• Alert urgency: Low (opportunity, not risk). Route to Expansion Signal Detector

Seasonal Pattern

• Definition: Usage pattern correlates with a known seasonal cycle (e.g., Q4 dip in certain industries)
• Significance: Not a risk signal -- suppress alert if pattern matches historical seasonal data
• Handling: Compare to same period prior year if data is available. See references/seasonal-patterns.md

Step 6: Correlate with Known Events

For declining patterns, check for correlating events:

• Support tickets opened in the same window (product issues)
• Stakeholder changes in CRM (champion departure)
• Product releases or changes (new version causing friction)
• Known outages or incidents

Present correlations as hypotheses, not conclusions. The human interprets causation.

Step 7: Generate Output

Produce a per-account usage profile and, if a pattern is detected, an alert.

Output Format

Per-account usage profile:

{
  "account_id": "string",
  "period": "2026-03-10",
  "metrics": {
    "dau_mau_ratio": { "value": 0.42, "baseline_90d": 0.38, "delta_pct": 10.5, "segment_position": "above_median", "trend": "growing" },
    "feature_adoption_breadth": { "value": 0.65, "baseline_90d": 0.61, "delta_pct": 6.6, "segment_position": "above_75th", "trend": "stable" },
    "session_depth": { "value": 12.3, "baseline_90d": 14.1, "delta_pct": -12.8, "segment_position": "below_median", "trend": "declining" },
    "key_workflow_completions": { "value": 847, "baseline_90d": 820, "delta_pct": 3.3, "segment_position": "above_median", "trend": "stable" },
    "licence_utilisation": { "value": 0.78, "baseline_90d": 0.72, "delta_pct": 8.3, "segment_position": "above_median", "trend": "growing" }
  },
  "patterns_detected": [],
  "overall_usage_trend": "stable"
}

Pattern alert (generated only when a pattern is detected):

{
  "account_id": "string",
  "alert_type": "gradual_decline",
  "severity": "high",
  "evidence": {
    "metric": "session_depth",
    "current_value": 12.3,
    "baseline_90d": 14.1,
    "consecutive_declining_periods": 4,
    "segment_position": "below_median"
  },
  "correlations": [
    { "type": "support_ticket", "detail": "P2 ticket opened 12 days ago re: performance issues" }
  ],
  "suggested_investigation": "Session depth declining for 4 consecutive weeks. Possible product performance issue -- correlates with open P2 ticket. Verify with customer whether performance is impacting usage."
}

Portfolio usage report (weekly):

• Accounts by overall usage trend: growing / stable / declining / insufficient data
• Top 5 largest usage declines with pattern classification
• Top 5 largest usage increases (expansion signals)
• Accounts with adoption plateaus (below median, stable for 60+ days)

Handoff to Human

See references/human-decision-guide.md for interpretation framework.

Confidence and Limitations

• High confidence for pattern detection given clean data with 90+ days of history
• Medium confidence for causal correlation (why usage changed). Agent presents correlations as hypotheses, never conclusions
• Low confidence for accounts with <30 days of history -- insufficient baseline. Flag as "new account, limited usage history"
• Cannot distinguish intentional usage reduction (customer rightsizing) from disengagement
• Seasonal patterns require at least 1 year of historical data to detect reliably

Dependencies

Required:

• Product analytics API (Mixpanel, Amplitude, Pendo, or data warehouse)
• Segment benchmark dataset (shared with Health Score Engine)
• Account metadata from CRM (tier, size band, contract age)

Downstream consumers:

• bi-health-score (usage component input)
• bi-risk-detector (pattern alerts as risk signals)
• bi-expansion-detector (breakout growth as expansion signal)
• bi-account-brief (usage profile for account context)

References

• references/usage-benchmarks.md -- Metric definitions and benchmark thresholds by segment
• references/seasonal-patterns.md -- Known seasonal patterns and suppression logic
• references/human-decision-guide.md -- How CSMs should interpret usage alerts