Descriptive
"What happened?"
Foundational
Summarizes historical data into reports, dashboards, and KPIs. No model is "scoring" anything — it's aggregation, counting, and visualization. Output is almost always consumed by a human decision-maker. This is where most organizations start and where many business questions live permanently.
Typical deployment modes
Dashboard / Human
Batch (Scheduled Reports)
Pipeline (Pre-Model)
Real-world examples
Retail
Weekly sales report by region
Batch ETL aggregates POS data nightly → scheduled report emailed to regional managers every Monday.
Healthcare
Hospital census dashboard
Real-time dashboard shows current bed occupancy, admissions, and discharges. Nurse managers check it hourly to manage staffing.
Finance
Monthly P&L summary
Batch pipeline rolls up transactions into financial statements. CFO reviews in a dashboard with drill-down to cost centers.
Diagnostic
"Why did it happen?"
Explanatory
Investigates causes behind observed patterns — root-cause analysis, drill-down exploration, and statistical testing. More interactive than descriptive: analysts use ad-hoc queries and visual exploration to find the "why." Often triggered when a descriptive metric moves unexpectedly. Output is still human-consumed, but the tools are more exploratory.
Typical deployment modes
Dashboard (with Drill-Down)
Batch (Ad-Hoc Analysis)
In-Database (SQL Queries)
Real-world examples
Manufacturing
Root-cause analysis of defect spike
Quality engineer drills into a dashboard filtering by shift, machine, and raw material batch to isolate what changed when defects tripled.
Marketing
Why did Q3 campaign underperform?
Analyst runs in-database queries segmenting conversion rates by channel, audience, and creative. Presents findings in a one-time report to the CMO.
HR
Investigating turnover drivers
HR runs a regression to explain turnover by department, tenure, and manager rating. Results presented in a leadership deck — advisory, not automated.
Predictive
"What will happen?"
This Course ★
Uses historical patterns to forecast future outcomes or classify new cases. This is where the full range of deployment types appears — because predictions can be consumed by humans (dashboards), by other software systems (APIs), or by automated workflows (real-time blocking). The 72 use cases below are almost all predictive.
Typical deployment modes
Batch
Real-Time
In-Database
Edge / Embedded
Hybrid
Dashboard / Human
Real-world examples
Telecom
Customer churn prediction
Batch scoring weekly → top risk scores pushed to CRM → retention team calls the most likely churners first.
Banking
Fraud detection at swipe
Real-time API scores each transaction in milliseconds → high scores trigger step-up auth or hard block.
Manufacturing
Equipment failure prediction
Edge model on IoT sensor detects vibration anomalies locally → alerts maintenance before failure occurs.
Prescriptive
"What should we do?"
Advanced / Frontier
Goes beyond prediction to recommend or automate the best action. Typically wraps a predictive model inside an optimization or decision engine. A prescriptive system doesn't just say "this customer will churn" — it says "offer this customer a 20% discount on a 12-month contract, delivered via email on Tuesday, because that combination maximizes expected retained revenue." This is the frontier of analytics maturity and where the most business value lives.
Typical deployment modes
Real-Time (Decision Engines)
Batch (Optimization)
Hybrid (Predict + Optimize)
Dashboard (Recommend → Human Approves)
Real-world examples
eCommerce
Dynamic pricing engine
Real-time system predicts demand elasticity per product → optimization algorithm sets the price that maximizes margin within competitive constraints. Fully automated.
Logistics
Route optimization for delivery fleet
Batch forecast predicts package volumes by zone → optimization solver assigns trucks to routes minimizing fuel + time. Dispatcher reviews and approves.
Healthcare
Treatment recommendation system
Predictive model estimates patient risk → prescriptive layer recommends care pathway (e.g., telehealth vs. in-person). Clinician reviews and decides — human-in-the-loop is mandatory here.
Finance
Portfolio rebalancing
Batch optimization runs quarterly: forecasts returns, applies risk constraints, recommends trades. Advisor reviews before execution.
Key insight for interviews: Most real-world analytics systems blend multiple levels. A fraud detection system is predictive (scores the transaction) but becomes prescriptive when it automatically decides to block, allow, or request step-up authentication. Descriptive dashboards monitor whether the predictive models are still performing. The analytics maturity isn't a ladder you climb once — it's layers you stack and maintain simultaneously.