Bizocity Score at AT&T — Case Summary 📞📊

Overview

• Case: Bizocity score at AT&T Long Distance.
• Problem: AT&T Long Distance needed an efficient way to identify and target non-AT&T phone numbers likely to be business customers for customer acquisition/marketing.
• Solution proposed by Bell Labs (Corina Cortes & Daryl Pregibon): build a daily behavioral scoring system (“Bizocity”) using call detail data (CDD).

Business Problem (As-Is) ❗

• Marketing goal: acquire new customers (prospecting) within limited budget → must prioritize high-value prospects.
• Existing data source: purchased phone directories (costly, unreliable, lacked business/residence flags).
• Manual approach: telemarketers called numbers to classify business vs. residential → very low response rates and inefficient.
• Two core issues:
  • Data sourcing: directories incomplete/unreliable.
  • Targeting/classification: inability to identify which numbers are business vs. residential and which prospects are valuable.

Insight & Data Source 🗂️

• Bell Labs observed AT&T’s own network logs (Call Detail Data, CDD) already contain many non-AT&T numbers:
  • CDD fields: caller ID, receiver ID, start time, end time (thus duration).
  • CDDs are legally retained and capture who called whom, when, and for how long.
• Over time AT&T sees millions of distinct numbers (65M/day, 300M/month), giving broad coverage of non-AT&T prospects.

Analytics Solution — Bizocity Score 🔢✨

• Goal: compute a daily score (probability 0–1) indicating likelihood the phone number is used for business.
• Key behavioral indicators derived from CDD:
  • Time of day (business hours vs. non-business hours) ⏰
  • Call duration (longer calls as proxy for monetary/value / business activity) ⏳
  • Proportion of calls to known businesses (calls made to businesses → business-like behavior) 🏢
• Modeling approach:
  • Build a regression / classification model (e.g., logistic regression, decision trees, etc.) using these indicators.
  • Score every telephone number every day; scores updated as new data arrives (weighted average with recency bias).

Prospect Profiling (RFM-style) 🧾

• Each number receives a small profile composed of:
  • Recency: days since last seen
  • Frequency: average daily appearances / time between appearances
  • Monetary proxy: average daily minutes (duration)
  • Bizocity score
• Use combined filters (bizocity + R/F/M) to prioritize valuable business prospects, not just likely-business numbers.

Business Value & Implementation Notes 💡

• Benefits:
  • Eliminates costly directory purchase and low-response telemarketing classification.
  • Produces ranked prospect lists so marketing spends budget on highest-value targets.
• Operational considerations:
  • Real-time / near-real-time scoring requires infrastructure:
    • CDD ingestion, aggregation pipelines, daily retraining/updating of models.
    • Storage for profiles and historical aggregates; integration with marketing systems for campaign targeting.
• Key lesson: translate business problem into analytics problem; combine domain knowledge + available data + statistical modeling to produce actionable, operational solutions.

Key Takeaways ✔️

• Use in-house behavioral data (CDD) instead of third-party directories to solve data-source problems. 🛠️
• Behavioral indicators (time, duration, call targets) can predict business usage — produce a daily Bizocity score for prospecting. 📈
• Combine scoring with RFM-like profiling (recency, frequency, duration) to prioritize valuable prospects. 🎯
• Analytics must be tied to business objectives and operationalized (infrastructure, retraining, integration) to create real value.