The best credit analysts of 2025 are becoming the best prompt engineers. They know how to coax the hidden insights out of the AI.
Structure Your Query: The Role-Task-Constraint Framework
A junior analyst asks: “Is this borrower risky?”
The AI gives a generic, safe answer.
A senior analyst asks: “Act as a conservative Credit Risk Officer (Role). Analyze the attached cash flow statement for the last 3 years (Task). Highlight any quarters with negative operating cash flow that deviate >20% from the trend, and cross-reference these dates with the sector’s performance (Constraint).”
The difference in output quality is night and day.
Few-Shot Prompting for Spreading
When teaching the AI to spread a financial statement, don’t just give instructions. Give examples (shots).
“Here is how we treat ‘Directors Remuneration’: It should be added back to Operating Profit. Here is an example: [Example A]. Now do the same for this new company.”
Giving the AI 3-5 examples of how your bank specifically classifies tricky line items improves accuracy from 70% to 95%.
Chain of Thought Verification
Always ask the AI to “Show your work.” If it calculates a Debt Service Coverage Ratio (DSCR) of 1.5, ask it to output the formula it used and the exact numbers it plugged in. This allows the human analyst to verify the logic, not just trust the number.