TOOL-008 — Product Adoption Pattern Recognizer

Tier 3 Specialist Tool · Stateless · Recognizes feature-by-feature adoption patterns for an account · Different shape than TOOL-004 — this is feature engagement, not consumption volume

Tier 3 · Tool Specced · v31 Post-sales · Activation Haiku

Purpose

Reads per-account feature engagement telemetry (which product features are used, by how many users, how often, in what sequence) and classifies the account's adoption pattern: deeply integrated / surface-only / siloed-by-team / declining / activating. Different from TOOL-004 (consumption volume forecasting) — this tool is about which features are used, not how much. Used by AGT-501 customer health (as activation dimension augment) and AGT-902 Account Brain for "is this account actually getting value?" diagnostic queries.

Closes the post-sales gap on feature-level adoption that the v26 evaluation flagged as missing. Today AGT-501 has a seat utilization dimension (how many seats are active) but no feature-engagement dimension (which features those seats are using). That distinction matters for predicting churn/expansion much earlier than seat metrics alone.

Input schema

{
  "account_id": "uuid",
  "feature_engagement_telemetry": {
    "trailing_window_days": 90,
    "feature_usage": [
      {
        "feature_id": "string",
        "feature_category": "core" | "advanced" | "integration" | "admin" | "experimental",
        "first_use_at": "ISO 8601 | null",
        "last_use_at": "ISO 8601 | null",
        "use_event_count_in_window": 0,
        "distinct_users_in_window": 0,
        "users_pct_of_active": 0.0          // % of account's active users using this feature
      }
    ]
  },
  "account_context": {
    "contract_start_date": "ISO 8601",
    "active_seats": 0,
    "licensed_seats": 0,
    "primary_use_case": "string",
    "ideal_feature_set_for_use_case": ["feature_id"]   // optional —
                                                       // expected features per ICP segment
  },
  "comparison_baseline": {                  // optional — expected pattern per cohort
    "cohort_definition": "string",
    "cohort_typical_feature_breadth": 0,    // # features used by median similar account
    "cohort_typical_users_per_feature_pct": 0.0
  }
}

Output schema

{
  "tool_call_id": "uuid",
  "adoption_pattern": {
    "primary_pattern": "deeply_integrated" | "surface_only" |
                       "siloed_by_team" | "declining" | "activating",
    "secondary_pattern": "string | null",   // when account fits two patterns
    "pattern_confidence": "high" | "medium" | "low"
  },
  "adoption_metrics": {
    "feature_breadth": 0,                    // # distinct features used in window
    "feature_breadth_vs_cohort": 0.0,        // ratio vs baseline
    "core_feature_adoption_pct": 0.0,        // % of core features used
    "advanced_feature_adoption_pct": 0.0,    // % of advanced features used
    "integration_feature_count": 0,
    "feature_concentration_index": 0.0,      // 0-1, low = broad use, high = concentrated
    "newly_adopted_features_in_window": 0,
    "abandoned_features_in_window": 0        // features that lost users
  },
  "key_observations": [
    {
      "observation": "string",
      "supporting_metric": "string",
      "interpretation": "string"
    }
  ],
  "implications_for_caller": {
    "expansion_signal": "strong" | "moderate" | "none",
    "churn_signal": "strong" | "moderate" | "none",
    "intervention_recommendation": "string | null"
  },
  "ungrounded_assumptions": ["string"],
  "data_completeness": "high" | "medium" | "low",
  "tool_metadata": {
    "model": "claude-haiku-4-5",
    "input_tokens": 0, "output_tokens": 0,
    "cost_usd_estimate": 0.0,
    "latency_ms": 0
  }
}

Adoption pattern taxonomy

Pattern	What it looks like	Implication
deeply_integrated	Broad feature breadth + high users-per-feature + multiple integration features active. Account uses the product as system-of-record for the use case.	Strong renewal signal. Strong expansion candidate. Low churn risk.
surface_only	Narrow breadth + only core features + low users-per-feature. Account uses the product but hasn't gone deep.	At-risk on renewal — competitor swap-out is easier. Activation play candidate — can grow into deeply_integrated.
siloed_by_team	Decent breadth but concentrated — one team uses many features, other teams use few or none. High feature_concentration_index.	Cross-team expansion signal. Risky on champion loss — if the active team's champion leaves, account drops to surface_only fast.
declining	Abandoned features > newly adopted features in window. Users-per-feature trending down.	Strong churn signal. Intervention candidate. AGT-503 expansion plays should be suppressed.
activating	Newly-onboarded account with rapidly expanding feature breadth + new users adopting. Trajectory positive.	Onboarding success signal. Activation investment paying off. AGT-501 health should reflect upward trend.

Called by

Caller	Invocation context
AGT-501 Customer Health Monitor	Daily batch on per-account basis. Output augments AGT-501's behavioral dimensions with a "feature engagement" dimension (separate from seat utilization).
AGT-902 Account Brain	"Is this account actually getting value?" queries. Brain integrates pattern + key_observations into BrainAnalysisLog narrative.
AGT-503 Expansion Trigger	Augments expansion qualification — deeply_integrated and siloed_by_team accounts get higher expansion-readiness signal; declining accounts get expansion suppressed.
AGT-601 Onboarding Orchestrator	For activating accounts — tracks onboarding success via feature breadth growth.

Design principles

Pattern recognition, not score generation. Output is a characterization, not a single number. AGT-501 owns score composition; this tool informs one dimension of that score.
Cohort-relative when possible. Feature breadth means little in absolute terms ("they use 8 features") — meaningful only relative to expected breadth for similar accounts. The comparison_baseline input enables this; tool degrades gracefully when baseline absent.
Numerical work in code, characterization in LLM. Same pattern as TOOL-004 and TOOL-005. Concentration index, breadth ratios, abandonment counts — all computed deterministically. LLM produces the pattern label + observations + interpretation.
Distinguish "siloed_by_team" from "concentrated usage." A small account where 3 users use many features is not siloed — it's appropriately scoped. Siloed_by_team requires multi-team account context (active_seats > some threshold). Tool is honest when account size doesn't permit the distinction.
Onboarding-aware. An account 30 days into contract with low feature breadth is not "surface_only" — it's "activating" or pre-pattern. Tool factors contract_start_date.

Cost ceiling

Constraint	Value
Per-call input budget	10K tokens (feature telemetry can be sparse for small accounts, dense for large)
Per-call output budget	1.5K tokens
Default model	Haiku (numerical + classification)
Per-call cost estimate	~$0.01
Monthly cap	$300/mo (~15,000 calls)
Frequency	Daily batch from AGT-501 across active accounts. Higher volume than other tools but Haiku keeps cost contained.

Eval criteria

Criterion	Pass threshold
Schema compliance	100% (hard)
Pattern classification accuracy	20 retrospective accounts with known retrospective adoption profiles. ≥ 75% pattern match.
Churn signal calibration	10 accounts that retrospectively churned. % where TOOL-008 flagged churn_signal as moderate or strong ≥ 12mo before churn. ≥ 60%.
Onboarding-aware classification	5 accounts in first 60 days of contract. % correctly labeled "activating" rather than "surface_only" — 100% (hard, eval-enforced).
Cohort-baseline degradation	5 cases without baseline data. Tool produces output with `data_completeness = 'medium'` or `'low'`; does not fabricate cohort comparisons. 100% (hard).
P95 latency	≤ 2s