Descriptive, Diagnostic, Predictive, Prescriptive

• Place any analysis in one of the four maturity levels
• Match methods to maturity level
• Sequence work so prescriptive advice rests on trustworthy description

• Teams over-invest in prediction while their description is still broken.
• Mixing levels confuses expectations ('is this dashboard or is this a recommendation?').
• Each level implies different stakes and different rigor requirements.
• Knowing the ladder tells you where your team's analytics actually is.

First symptoms are recorded (descriptive), tests explain them (diagnostic), prognosis forecasts (predictive), and the treatment plan is chosen (prescriptive). Skipping the diagnosis to prescribe is malpractice in both worlds.

• Descriptive: dashboards and reports compute counts, rates and trends from historical data.
• Diagnostic: drill-downs, segmentation and correlation identify drivers of an observed pattern.
• Predictive: regression, time-series forecasting and classification models estimate future values.
• Prescriptive: optimization, simulation and recommendation logic choose actions that maximize an objective.
• Feedback: measuring outcomes of prescriptions closes the loop and improves the whole ladder.

• A descriptive BI layer reports weekly revenue; diagnostic analysis ties a dip to a pricing change in one region.
• A predictive model forecasts demand weekly so the supply chain pre-orders stock.
• A prescriptive optimization recommends which customers should receive a retention offer, using LTV and predicted churn.

• Validate each level before climbing: a predictive model built on skewed descriptions misleads.
• Label deliverables by level so stakeholders know whether this is a report or a recommendation.
• For prescriptive advice, state the objective function and constraints explicitly.
• Timebox prediction to far-future assumptions you document.

• Calling a descriptive dashboard 'AI' — it is not predictive.
• Going straight to prediction when the diagnostic story is still a mystery.
• Prescribing actions with no measurement plan, so nobody learns if it worked.
• Overfitting the past and presenting it as a forecast of the future.

• Open 'predictive' deliverables and check for a historical description underneath — it often is.
• When prescriptions fail, check the descriptive input they were built on.
• Revalidate forecasts against actuals with error metrics (MAE, MAPE), not vibes.
• Segment predictions to find where the model is confident vs guessing.