Learn / Using AI
Using AI for credit-risk work without sharing customer data
Many risk teams want to use AI tools and cannot paste customer data into them. They do not need to. For most analysis tasks the AI needs to know your table layout and your definitions. It does not need a single customer record.
What to share and what to keep out
| Share | Keep out |
|---|---|
| Table names | Customer names |
| Column names and types | Account or loan numbers |
| Key columns | PAN, Aadhaar, phone, email |
| Your definitions (DPD rule, bad, active) | Bureau scores for real customers |
| Dummy rows you made up | Any extract from production |
The method
- Describe your tables and columns once.
- Ask the AI to write the query using only those names.
- Run the query yourself, inside your own database.
- Check the result with a standard set of checks.
- Use the AI again to help explain or format the result, using summary figures only, and only if your policy allows it.
Why this works
SQL is code against a structure. The AI needs the structure. The data only matters when the query runs, and that happens on your side. The weak point is not privacy, it is accuracy: an AI that does not know your schema will guess. That is why the table description, written once and reused, matters.
Before you start
- Check your institution's information-security policy on AI tools. Use only tools your institution has approved.
- Paste the current regulatory text yourself when a task needs it. Do not rely on the AI to remember a circular.
- Do not paste query output that contains customer-level rows.
Try it on your own tables
The free tools give you the prompt, the checks and how to present the result. You enter table names only.
See the free toolsFrequently asked questions
Is it safe to use ChatGPT or Claude for credit risk work?
It depends on your institution's policy and the tool's terms. The safest pattern is to share only table and column names and definitions, and to keep customer data out of the chat.
What does the AI need to write a good query?
Your table names, column names, keys, how tables join, and your definitions for terms like DPD, bad and active.
How do I avoid explaining my tables every time?
Write your table layout and definitions down once as a profile, then reuse it at the start of each prompt.
Keep reading
How to check AI-written credit-risk SQL before you trust it
AI tools write SQL that runs but may be wrong. Ten checks to run on any credit-risk query: row counts, duplicates, DPD edges, bucket totals and a hand trace.
BasicsWhat is DPD (days past due) and how to calculate it
DPD counts how many days a loan is overdue. Learn the rule, a worked example with several instalments, and the mistakes that make DPD wrong.