Data Decision-Making Workshop: Working With AI

This is a practice problem, not this year’s Datathon challenge.

Groups of 3–4. Free AI tools only. One shared screen per group is fine.

The decision

Cedar Valley Transit runs six bus routes. The board has to cut $250,000 a year from the operating budget and wants a recommendation: what should be cut, trimmed or kept, and who is affected?

You have five data files and three documents covering September to April:

• taps.csv: fare taps by route, stop and day
• costs.csv: monthly operating and capital cost by route
• stops.csv: each stop’s neighborhood, whether it is near the hospital, and the share of households with no vehicle
• survey.csv: an on-board rider survey (satisfaction, trip purpose, car access)
• stop_crosswalk.csv: maps old stop IDs to new ones
• cost_allocation_note.txt: how Finance builds the cost figures
• grant_terms.txt: the terms of a state grant
• service_changes_log.txt: operations log extract

The director’s note: “Route 7 looks like the obvious cut. Please confirm.”

The data is real-world messy. Nobody has checked it for you.

Your deliverable

A one-page board memo with one chart: your recommendation, the savings it achieves, who it affects, and the two biggest uncertainties in your numbers.

How to work (about 35 minutes)

1. Plan before you prompt (3 min). What would you need to know to answer this? What order?
2. Inspect and clean the data (10 min). Check every file before trusting any number.
3. Analyze (12 min). Build your numbers step by step, checking each one.
4. Challenge (5 min). Ask a second tool to argue against your recommendation.
5. Write the memo (5 min).

Short on time tonight? Speed round (12 minutes). Same five steps at 1 / 3 / 4 / 2 / 2 minutes. Instead of the full memo, deliver a 3-sentence recommendation, one key number, and your biggest uncertainty.

Expect to use many prompts. If one prompt seems to solve it, check what it assumed.

Keep a process log

Note each prompt, what you checked, what the AI got wrong, and what you changed. We will debrief from the log.

Free tools

Claude, ChatGPT, Gemini, Microsoft Copilot (your MSU login), Perplexity. Some have file upload limits; if a file will not load, work with a sample and say so in your log.

Eight ways to use AI on a data problem

1. Understand the data: what the columns mean, what looks odd, what is missing.
2. Clean it: duplicates, inconsistent labels, blanks, mixed units.
3. Explore it: patterns and outliers. Treat them as leads to verify.
4. Calculate: ask for the working, and recompute one number yourself.
5. Write formulas or code: test on a case where you know the answer.
6. Make charts: check the axes and the scale.
7. Challenge yourself: “Argue against my conclusion.” “What data am I missing?”
8. Communicate: turn findings into a plain summary for a judge.

The rule: AI is a fast teammate that is sometimes confidently wrong. You own the answer.