Decisioning Pipeline
End-to-end flow from input to recommendation
OverviewData & FeaturesModels & AlgorithmsDecisioning PipelineReal-time ArchitectureExplainabilityFeedback LoopMetrics & Impact
Scoring Pipeline
Step 1
Input Collection
Gather customer and coverage data
Customer profile (demographics, location)
Risk attributes (vehicle, property)
Coverage requirements (limits, deductibles)
Agent context (assigned agent, channel)
Step 2
Feature Generation
Transform raw data into model features
Price competitiveness index
Carrier win rate for similar profiles
Risk category embeddings
Payment reliability score
Step 3
Model Scoring
Score each carrier using ML model
Generate prediction for each carrier
Calculate conversion probability
Estimate expected premium
Compute retention likelihood
Step 4
Ranking Engine
Rank carriers by predicted outcome
Sort by conversion probability
Apply secondary sort (retention, LTV)
Handle ties with confidence scores
Generate ranked list with scores
Step 5
Business Rules Overlay
Apply business constraints
Carrier appetite filters
Regulatory compliance checks
Commission tier requirements
Exclusion rules (blacklists)
Step 6
Output Recommendation
Final ranked carrier list
Top 3-5 carriers with scores
Conversion probability per carrier
Expected premium range
Explanation (why recommended)
Hybrid System: ML + Rules
ML Model
Scores carriers based on conversion probability and predicted outcomes
Combined
Business Rules
Enforces compliance, appetite limits, and business constraints
Why hybrid? ML provides accuracy, rules provide guardrails. Rules can override ML when compliance or business logic requires it.
Example: Personal Auto Quote
Input
45yo, CA, 2 vehicles, clean record
Features
Risk score: 72, LTV proxy: $4,200
Scoring
5 carriers scored in 120ms
Ranking
Travelers > Hartford > CNA
Rules
All pass appetite check
Output
Travelers (87% prob, $1,842 est)