Problem Solving › Data Analysis

The model that works but discriminates

Il modello che funziona ma discrimina · Data Analysis · Beginner

The case

The scoring model you built to prioritize loan applications works: it cuts defaults by 20% and has been in production six months. But an analysis you ran out of diligence shows it systematically penalizes applicants from two low-income geographic areas, with double rejection rates at equal actual creditworthiness: the model does not use the address, but reconstructs it through proxy variables. The business will not touch a model that earns, legal says no forbidden variable is formally used, and the AI Act is starting to apply to credit scoring systems. Rebuilding the model without the proxies takes three months and, per first estimates, costs a third of its effectiveness. What do you do?

Il modello di scoring che hai costruito per prioritizzare le richieste di finanziamento funziona: riduce le insolvenze del 20% ed è in produzione da sei mesi. Ma un'analisi che hai condotto per scrupolo mostra che penalizza sistematicamente i richiedenti di due aree geografiche a reddito basso, con tassi di rifiuto doppi a parità di merito creditizio effettivo: il modello non usa l'indirizzo, ma lo ricostruisce da variabili proxy. Il business non vuole toccare un modello che rende, il legale dice che formalmente nessuna variabile vietata è usata, e l'AI Act inizia ad applicarsi ai sistemi di credit scoring. Rifare il modello senza le proxy costa tre mesi e, secondo le prime stime, gli fa perdere un terzo della sua efficacia. Cosa fai?

Read the case, write how you would solve it — in free text, no prompt required — and Grace's AI judges whether your solution actually works: solved or not solved, with a 0-100 score and coaching feedback that points at what is still uncovered without revealing the answer. Your first 5 evaluations each day are free.

Try to solve this case →

Frequently asked questions

What is the "The model that works but discriminates" problem-solving case on Grace?

It is a expert-level Data Analysis case based on a real workplace situation. You read the problem, write how you would solve it in free text, and Grace's AI judges whether your solution actually works — or what is still uncovered.

Can I retry if my solution doesn't solve the case?

Yes. The feedback tells you what remains uncovered — without revealing the solution — so you can strengthen your plan and try again. After you solve it, 6 more Data Analysis cases are waiting.

Who is this case for?

Anyone who wants to train structured problem solving: managers, professionals and candidates preparing for case interviews in Data Analysis. The expert difficulty makes it an accessible starting point.

How is my solution evaluated?

Grace's AI checks four things: whether you address the core problem (not just the symptoms), whether your plan is feasible within the constraints of the case, whether you consider risks and consequences, and whether you avoid the typical mistakes. You get a solved/not-solved verdict, a 0-100 score and coaching feedback.

More Data Analysis cases

All 348 problem-solving cases → · Prompt training scenarios →

From our network

HSE Genius — AI for Safety Data Sheets

Extract SDS data, H phrases and ECHA compliance checks in seconds, powered by AI.

Visit hsegenius.com →

From the Agora Intelligence blog

More on agora-intelligence.com →

📱 Download the Android app (beta) iOS coming soon

Say what you mean. Get what you need.

Grace Certified — the AI coach that trains and certifies your prompt engineering — by Agora Intelligence.