Stop 01 · Data Science FoundationsTopics · 5/9

Data-Driven vs Model-Driven Decisions

In short

Not every question needs machine learning. Knowing when a simple count is enough saves weeks of work and makes you the person people trust.

Some decisions just need clean descriptive data (data-driven); others genuinely need a predictive model's output (model-driven) — knowing which is needed prevents both under- and over-engineering a solution.

The mascot shopkeeper checks a ledger book behind a counter stacked with sacks and jars.

Think of it like this

The kirana store owner

Uncle knows from his register that sweets sell more in Diwali week. That's data-driven: look at the numbers and decide. Predicting exactly how many kilos of 300 items to order every week, for every festival and rain spell, needs a model.

Watch it happen

Which one does this question need?

Data-driven

Which city cancelled most last month?
Answered with SQL or a chart
Done in hours, easy to explain

Model-driven

Which user will cancel next week?
Needs a trained prediction model
Takes weeks, needs monitoring later

Ask 'could a dashboard answer this?' first. Build a model only when you must predict for many cases, often.

The big ideas

What you need to know

Past questions need data

'What happened?' and 'how much?' are usually answered by counting, grouping and charting.

Repeated predictions need models

When you must guess an outcome for thousands of cases every day, a model makes sense.

Models carry a cost

They need data pipelines, testing, monitoring and retraining. A chart needs none of that.

Saying 'no model needed' is a skill

Good data scientists often talk stakeholders out of ML when a simple rule or report works.

See it in code

Walk through the code

sql

1-- Question: which city had the most cancellations last month?2SELECT city,3       COUNT(*) AS cancelled4FROM orders5WHERE status = 'cancelled'6  AND order_date >= DATE '2026-09-01'7  AND order_date <  DATE '2026-10-01'8GROUP BY city9ORDER BY cancelled DESC;

1/3Write the question down first. It keeps the query honest.

A data-driven question answered with one query. No training, no model, and anyone can check the logic.

Save yourself the pain

Common mistakes

Oops

Building a model because the stakeholder said 'use AI'.

Do this instead

Ask what decision they'll make. If a monthly report answers it, start there.

Oops

Refusing models even when predictions are needed for every user daily.

Do this instead

A count of last month can't say who leaves next week. That's when a model earns its cost.

Your mission

Sort ten questions

~15 min

Quick check

Which question genuinely needs a model?

Go deeper

1Recall a decision you've seen made and classify it as needing data or needing a model.
2Learn to ask 'could a simple dashboard answer this?' before reaching for ML.
3Understand that a good data scientist often talks a stakeholder OUT of needing a model.