Credit risk analysis in corporate banking has always been a document-heavy, judgment-intensive discipline. A single mid-market borrower can generate a 60-page annual report, three years of financial statements, covenant schedules, an industry outlook, and months of internal relationship notes. Before a credit officer can form an opinion, they must reconcile all of it: normalize EBITDA adjustments, trace contingent liabilities, cross-check management commentary against the numbers, and map everything to the bank’s internal rating scale. The work is repetitive in structure but unforgiving in detail. A missed footnote on a related-party guarantee or a misread working-capital swing can distort the risk grade that drives pricing, limits, and provisioning.
The friction compounds at the portfolio level. Quarterly reviews, watchlist migrations, and limit renewals arrive in waves, and analysts spend most of their time extracting and reformatting information rather than interpreting it. Senior credit committee members then receive memos of uneven quality, forcing them to re-derive the analysis in the meeting itself. Meanwhile, regulators expect a defensible, consistent rationale behind every rating decision — and “the analyst felt comfortable” is not a rationale.
ChatGPT changes the economics of this work when it is used as a structured analytical assistant rather than a search engine. Given the right inputs and a disciplined prompt, it can extract and normalize financial data, flag inconsistencies between narrative and numbers, draft a rating rationale aligned to your internal scale, and stress-test a borrower against downside scenarios — all in minutes, with the analyst retaining full ownership of the final judgment. The key word is structured: the quality of the output tracks the quality of the brief you give it.
Why generic prompts fail in credit work
Asking ChatGPT to “analyze this company’s credit risk” produces fluent but shallow commentary — the kind of text that sounds right in a summary and collapses under committee questioning. What works instead is a prompt that defines the task, supplies the source documents, specifies the output format, and constrains the tone to your institution’s credit culture. The two templates below follow that pattern: one for a single-borrower credit assessment, one for a portfolio-level early-warning review.
First, read these files completely before responding:
[borrower_annual_report.md] — audited financial statements, management commentary, and notes for the last three fiscal years
[borrower_financial_spreads.md] — my normalized spreads: revenue, EBITDA, margins, leverage, interest coverage, liquidity, working capital
[internal_rating_scale.md] — our rating grades, scorecard factors, weights, and default definitions
[industry_outlook.md] — sector trends, demand drivers, and key risks for [industry]
Here is a reference for what I want to achieve:
[Upload a previously approved credit memo as markdown that represents the quality bar.]
Here’s what makes this reference work:
It opens with a one-paragraph recommendation and rating, then supports it with a structured financial trend table, then addresses business risk and financial risk separately, then closes with covenants, conditions, and a clear list of open items. Tone is factual, hedged where evidence is thin, and every claim is traceable to a source document.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: Credit assessment memo, 1,200–1,600 words plus a financial trend table.
Recipient’s reaction: The credit committee should think “the analysis is complete and the rating is defensible” and approve, decline, or approve with conditions in one meeting.
Does NOT sound like: A sell-side equity report, a marketing pitch, or generic AI commentary with vague adjectives.
Success means: Every rating driver is cited to a source document, all three years are compared, and no material footnote or contingent liability is omitted.
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Notice what the template does that a casual prompt does not. It names the source files and requires them to be read first. It anchors quality to a real reference memo instead of an abstract instruction like “write professionally.” It defines the recipient’s reaction as the success test, which pushes the model toward decision-ready output rather than exhaustive description. And it forces a planning step, so you can correct the approach before 1,500 words are generated in the wrong direction.
From single borrowers to the whole book
Single-name analysis is where most teams start, but the larger time sink is often portfolio surveillance. Early-warning reviews require scanning dozens of borrowers for deteriorating trends — falling coverage ratios, covenant headroom erosion, negative working-capital swings — and ranking them by urgency. The second template handles that at portfolio level, producing a triage output your team can act on the same day.
First, read these files completely before responding:
[portfolio_spreads.md] — borrower-level financial spreads for [number] obligors, four quarters of history
[covenant_tracker.md] — covenant packages, headroom, and test dates for each facility
[watchlist_policy.md] — our triggers, escalation rules, and watchlist definitions
[relationship_notes.md] — recent RM commentary, payment behavior, and any covenant waiver requests
Here is a reference for what I want to achieve:
[Upload last quarter’s early-warning summary as markdown.]
Here’s what makes this reference work:
It ranks borrowers by risk migration, not alphabetically. Each entry states the trigger observed, the supporting metric with quarter-over-quarter change, the recommended action, and the owner. It separates “act now” from “monitor” with explicit thresholds, and it never flags a borrower without citing a numeric trigger.
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: Early-warning triage report, 900–1,200 words plus a ranked borrower table.
Recipient’s reaction: The head of credit should be able to assign actions immediately and defend every escalation to internal audit.
Does NOT sound like: A data dump, a list of every borrower, or alarmist commentary without numeric support.
Success means: Every flagged borrower maps to a written policy trigger, false positives are minimized, and no borrower with two consecutive quarters of deterioration is missed.
My context file contains my standards, constraints, audience. Read it fully before starting.
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
A practical tip before you deploy either template: run it first on a borrower whose outcome you already know. Compare the model’s memo against your approved version, note where it overstates or underweights risk, and encode those corrections into your context file. Two or three iterations of this calibration will do more for output quality than any amount of prompt wording tweaks.
From there, the natural next step is standardization — turning your calibrated templates into a shared prompt library with fixed context files, so that every analyst in the team produces memos at the same quality bar. That consistency, more than speed, is what credit committees and regulators ultimately reward.
Published on 12 October 2026 on growwithgpt.com