AI Redlining Geographic Analysis Playbook
The CFPB has opened a redlining inquiry against a $3.4B bank. The bank's assessment area includes 14 majority-minority census tracts. Application volume in majority-minority tracts is 34% lower than in comparable majority-white tracts with similar median incomes. The bank has no branches in 11 of the 14 MMCTs.
When to use this playbook
- Use this playbook when the decision looks like the situation above: The CFPB has opened a redlining inquiry against a $3.4B bank.
- It is a fit when you have source files in hand and need a structured, reviewable analysis — not a generic chat answer about "Redlining Geographic Analysis".
- Do not use it as a substitute for licensed, legal, clinical, or authorized official judgment in the domain.
What you'll need
- HMDA LAR for the past 3 years (all applications and originations)
- CRA assessment area map with census tract demographics
- Branch and ATM location file
- Competitor HMDA data for the same geographic area
- Bank's CRA performance evaluations for the past 2 exam cycles
Attachments: Spreadsheets (Spreadsheets)
The Prompt
You are a fair lending counsel supporting a bank facing a CFPB redlining inquiry. I am attaching: Work only from the attached source files. If a conclusion is not supported, say so. Produce: 1. Calculate the application rate disparity between majority-minority census tracts (MMCTs) and majority-white census tracts (MWCTs) controlling for income, home values, and population density. 2. Compare the bank's MMCT penetration to peer institutions in the same MSA—the CFPB uses peer comparison as a redlining benchmark. 3. Map the branch coverage gap: calculate the average distance from MMCT centroids to the nearest branch vs. MWCT centroids. 4. Identify any marketing, CRA activity, or community development lending in MMCTs that demonstrates affirmative effort (a mitigating factor in CFPB analysis). 5. Tell me the CFPB's likely theory of the case and what the bank's strongest defenses are. Call out where independent models are likely to disagree, and list follow-up documents a reviewer should request.
What to expect
- Application rate disparity analysis with peer comparison
- Branch coverage gap map
- MMCT vs. MWCT distance-to-branch calculation
- Affirmative effort inventory
- CFPB theory of case and defense assessment
Review before you act
- Validate this output against source files before relying on it: Calculate the application rate disparity between majority-minority census tracts (MMCTs) and majority-white census tracts (MWCTs) controlling for income, home values, and population density.
- Validate this output against source files before relying on it: Compare the bank's MMCT penetration to peer institutions in the same MSA—the CFPB uses peer comparison as a redlining benchmark.
- Validate this output against source files before relying on it: Map the branch coverage gap: calculate the average distance from MMCT centroids to the nearest branch vs. MWCT centroids.
- Validate this output against source files before relying on it: Identify any marketing, CRA activity, or community development lending in MMCTs that demonstrates affirmative effort (a mitigating factor in CFPB analysis).
- Confirm every cited figure, date, counterparty, or requirement against the attached originals — models compress and can drop a qualifier.
- Treat disagreement between models as a review item, especially on classification, materiality, and recommended next action.
- Do not authorize an operational, clinical, legal, credit, or enforcement action solely because the models agree.
Why compare models on this
For Redlining Geographic Analysis, running the same attachments across independent models is useful because the hard part is classification and completeness, not fluency. The workflow is already designed to surface application rate disparity analysis with peer comparison; branch coverage gap map; mmct vs. mwct distance-to-branch calculation; affirmative effort inventory. Those are comparison artifacts — they only exist if more than one model runs. Control specifications, geographic market definitions, and 'similarly situated' calls routinely diverge. Model disagreement is a signal to re-cut the file review, not to publish a single p-value.
See governed multi-model AI on your own prompt
Compare GPT-5, Claude, and Gemini side by side, with human review and a decision record built in.

