🤖⚖️

The Algorithm's Blind Spot

A bank's Gen-AI credit engine has been running for four quarters.
Every dashboard is green. Then one rejected applicant goes public.

📋 Mission

Vaibhav Bank's SmartSanction engine pre-approves MSME loans using Gen-AI. Loan book growth is up 35%. NPA is down. Every headline metric is green — and an emergency Board meeting has been called anyway.

You will be assigned a role, investigate the evidence, sit through the crisis meeting, vote on what happens to SmartSanction, and uncover the number that was buried in a report nobody read.

The number that explains everything: 0.51
Figure out what it is. Figure out what it means.

🎭 Your Role Briefing
⏱️ Game Flow
Phase 1: Investigation (20 min) — unlock evidence, run the calculator
Phase 2: Board Meeting (25 min) — ask questions, debate
Phase 3: Vote on what happens to SmartSanction
Phase 4: Reveal what really happened
🎯 Remember
Stay in character during the Board meeting.
Figure out what 0.51 means — before the room does.
Everyone's dashboard is telling the truth. That's exactly the problem.
🔍 Investigation Phase
Unlock evidence. Find what the dashboard is hiding.
20:00
TIME REMAINING
🧮 Adverse Impact Ratio Calculator — The Number That Explains Everything

📁 Evidence Cards — Click to unlock

Each card reveals a piece of the story. The full picture only emerges when all cards are open.

⚖️ Who Is Most Responsible?

Rank each party. You can update this in the Board Meeting.

🏛️ Emergency Board Meeting
Every role must speak. Every claim must be backed by a number.
25:00
BOARD TIME LEFT
+35%
Loan Book Growth YoY
78%
Aggregate Approval Rate
2.1%
Aggregate NPA
0.51×
Adverse Impact Ratio (hidden)
Every number on the top row is real and accurate. The number in the last box was never on the monthly MIS pack. That is the entire case.

Cross-Examination Questions

0/8 addressed
🗳️ Cast Your Vote
The Board must decide what happens to SmartSanction — starting Monday.

📊 Live Class Vote (simulated)

Cast your vote first to unlock the reveal.

🤖
SMARTSANCTION
Live 4 quarters · 1 zone → 6 zones · 35% loan growth · Adverse Impact Ratio 0.51× · buried since Quarter 2

The Growth vs Governance Debate — Resolved

Zonal Business Head's Claim
35% loan book growth, best year in a decade
Measured: aggregate volume. Segment impact: not considered.
Analyst / CRCO's Claim
Adverse Impact Ratio of 0.51× — severe disparity against Tier-3 women-led MSMEs
Measured: distributional fairness across segments.
The actual answer: Both were completely true, at the same time — measuring different things with different lenses.

Aggregate approval rate: 78%. Aggregate NPA: 2.1%. Both genuinely excellent. Segment approval rate for Tier-3 women-led MSMEs: 41% vs 81% for the reference group — an Adverse Impact Ratio of 0.51×, far below the 0.80 four-fifths threshold used as a standard adverse-impact heuristic.

An aggregate metric being excellent does not mean a segment-level metric is safe. A dashboard can be telling the truth at 30,000 feet and still be hiding a serious, measurable harm one layer down.

This is exactly what decision-grade MIS — diagnostic, not just descriptive — is built to catch.

The Data Analyst who found this wrote it up correctly, seven months before it went viral. Her Q2 report footnote said the disparity was there. Her manager called it "probably noise, small sample size" and moved on.

She did nothing wrong by ordinary process standards. She filed the finding. It was accurate. It was ignored.

What was missing was not her diligence. It was an escalation path that didn't depend on one manager's judgment call.

If that footnote had reached the CRCO directly — or if a standing rule required any AI Adverse Impact Ratio below 0.80 to trigger automatic review, regardless of what any one person thought of the sample size — 1,400 Tier-3 MSME applicants would have been treated fairly three quarters sooner.

The 5-Step Check Every Executive Should Run on an AI-Assisted Process:
1. Ask for the segment-level breakdown, not just the aggregate number.
2. Calculate the Adverse Impact Ratio (focus ÷ reference approval rate); flag anything below 0.80.
3. Check the human-override rate — below ~10% means "decision support" is functioning as autonomous decision-making.
4. Confirm the governance classification has been reviewed since go-live, not only at launch.
5. If a junior team member's red flag gets dismissed, ask what independent channel exists to escalate it anyway.
That five-step check takes 10 minutes.
It would have caught this three quarters earlier.
📖 Faculty Debrief
Discussion questions, answer keys, course connections.

📋 Your Full Decision Record

🧮 Adverse Impact Ratio Lab — Live Calculations

Play Again with a Different Role

The Analyst's experience is very different from the Zonal Business Head's.