Day 2 of 14 · AI in Lending & Credit Control

Your Credit AI Toolkit (and Ground Rules)

⏱ 6 min 📊 Beginner

Yesterday you saw the opportunity: hours of information work at every lifecycle stage that AI can compress into minutes. Today you get equipped — and, more importantly, you learn the ground rules that let you use these tools without ending up in a meeting with compliance.

Because here's the truth about this course: the tools take an afternoon to learn. The data discipline is what separates the professional from the person who becomes a cautionary tale in next year's staff training.

The big three assistants

You don't need special software to start. Three general-purpose AI assistants cover almost everything in this course:

ChatGPT (OpenAI) — the most widely used, strong at structured analysis and drafting. Claude (Anthropic) — excellent with long documents, careful reasoning, and nuanced tone in customer communications. Gemini (Google) — tightly integrated with Google Workspace if that's your world.

All three do the work in this course well. What matters far more than which one you pick is which tier you're using:

Consumer tiers — the free or personal accounts you sign up for yourself. Convenient, but your bank almost certainly hasn't approved them for anything involving borrower data, and depending on settings, what you type may be used to improve the model.

Enterprise tiers — business versions your employer contracts for, typically with data protection commitments, no training on your inputs, and admin controls. This is what banks and lenders roll out when they approve AI use.

Your first move is a question, not a prompt: ask your manager or IT which AI tools are approved at your institution, and under what conditions. If there's an approved enterprise tool, use it. If there isn't, you can still practice everything in this course with synthetic data on a consumer tool — which brings us to the rule that governs everything else.

Knowledge Check

What matters most when choosing an AI assistant for credit work?

A

Whether you're on an employer-approved enterprise tier, and what your institution allows on it

B

Using several assistants at once and comparing their answers

C

Picking the assistant with the most impressive model benchmarks

D

Choosing the one with the cheapest subscription

The big three assistants are all capable of the work in this course. The real differentiator is governance: enterprise tiers come with data protection commitments and your institution's approval. Always start by asking what's approved at your organization — not by picking a favorite model.

The PII rule: anonymize before you paste

Here is the one rule in this course that is genuinely non-negotiable: never paste real borrower or customer data into a consumer chatbot. No names, no account numbers, no addresses, no dates of birth, no raw bank statements, no ledgers. Not once, not "just this time," not because the deadline is tight.

Why so absolute? Because borrower data is regulated personal data, and once it leaves your institution's controlled environment, you can't get it back. Data protection law, banking confidentiality, and your own employment contract all point the same way — and regulators have little patience for "the chatbot made me do it."

The good news: anonymized data works almost as well. The AI doesn't need to know it's analyzing "Sarah Milton, account 4471…" to spot an income pattern. It needs dates, amounts, and descriptions with identity stripped out:

Instead of: real name, account number, employer name, exact address.

Use: "Applicant A," "Employer (manufacturing, ~200 staff)," "monthly salary credit," rounded figures where precision doesn't matter.

For practice — like this course — go one step further and use fully synthetic data: invented statement lines that look realistic but belong to no one. Every example in these lessons is synthetic. Yours should be too, until you're inside an approved enterprise tool with clear internal guidance. And even then, follow your institution's policy — check with your compliance team about what may be shared, anonymized or not.

Two-column do and never card contrasting anonymized and synthetic data, approved enterprise tools, and verified outputs against pasting real borrower PII, account numbers, and raw statements into consumer chatbots

Print this mentally on every prompt you write. The left column builds a career with AI; the right column ends one.

Ground Rules Check

You're on a deadline and only have a consumer chatbot available. What's the acceptable way to get AI help analyzing a borrower's bank statement?

A

Paste the statement as-is — one time won't matter if you delete the chat afterward

B

Screenshot the statement instead of pasting the text

C

Paste it but change the borrower's name to initials

D

Strip all identifying details and use anonymized or synthetic statement lines — dates, amounts, descriptions only, per your institution's policy

Real borrower data never goes into a consumer chatbot — deleting the chat doesn't undo the disclosure, and swapping a name for initials or using a screenshot still exposes regulated personal data. Anonymized or synthetic data preserves the patterns AI needs while protecting identity — and your institution's policy always has the final word.

Three prompts to try today

Enough rules — let's get your hands dirty. Three safe, high-value prompts you can run today with public or synthetic material:

1. Summarize a policy document. Take a public lending policy, regulatory guidance note, or your institution's published terms and ask: "Summarize this document in plain English for a new credit analyst. List the 5 rules most likely to affect day-to-day underwriting decisions."

2. Explain a covenant. Paste a standard covenant clause from a template (not a live agreement) and ask: "Explain what this covenant requires, what would breach it, and what a lender typically monitors to check compliance."

3. Draft a customer email. "Draft a polite, professional email to a business customer whose account has gone 10 days past due for the first time. Friendly tone, clear ask, offer to discuss. Keep it under 120 words." Then edit it into your voice — and remember, nothing AI drafts goes to a real customer without a human reviewing and sending it.

🤖

ChatGPT

online

Summarize this anonymized bank statement extract for an affordability review. Applicant A, synthetic data: 01 Jun — Salary credit — +3,150 03 Jun — Rent — -1,100 05 Jun — Loan payment (Lender X) — -290 12 Jun — Card payment — -410 18 Jun — Overdraft fee — -25 25 Jun — Gambling site — -180 28 Jun — Salary credit (secondary) — +420

Here's a structured summary of the extract: **Income:** ~3,570/month from two sources — a stable primary salary (3,150) plus a smaller secondary credit (420). Verify whether the secondary income is regular across other months. **Fixed commitments:** Rent 1,100 + existing loan 290 + card payment 410 = ~1,800/month, roughly half of income. **Points to verify:** • Overdraft fee suggests the balance ran negative mid-month — check the running balance • Gambling transaction (180) — a single instance; check whether it's a pattern across other months • Confirm the secondary income source and its reliability This is a one-month snapshot from synthetic data — a real review needs at least 3 months of statements, and every figure verified against source documents by you before it informs any decision.

↻ Replay conversation

Your toolkit, assembled

That's the whole kit for now: one approved assistant, the anonymize-before-you-paste rule, and three prompts that already save you time this week. Notice the pattern in the chat demo — the AI summarized, flagged, and suggested what to verify. It didn't score the applicant, didn't recommend approval or decline, and the human keeps every decision. That division of labor is permanent.

Tomorrow we go under the hood: what's actually inside an AI credit scoring model, and why understanding it matters even if you never build one.

🧰

Day 2 Complete

"The tool is the easy part. The discipline — no real borrower data in consumer chatbots — is what keeps you employed."

Tomorrow — Day 3

How AI Credit Scoring Actually Works

Tomorrow you'll build a plain-English mental model of how ML credit scoring differs from traditional scorecards — and why more accurate doesn't automatically mean fair.

Continue to Day 3 →