Gains:
- Ability to understand what data credit scoring is based on and that artificial intelligence does not provide the score, but a clear summary of the score and accelerates the missing document check
- Use artificial intelligence to produce a consistency check and draft decision justification for the loan file and be able to make the final allocation decision as an authorized expert
- Ability to understand that model output must be documented in an explainable, justifiable and non-discriminatory manner and maintain red lines
When a loan application arrives, there is a simple question in the background: "Can this customer repay this loan and what risk does he put the bank in paying?" The answer to this question is dispersed in a series of tasks that take up most of the credit specialist's day: reading the income statement, calculating the debt/income ratio, interpreting the credit history (KKB - Credit Registry Bureau note), evaluating the collateral, comparing it with internal policy and finally writing a reasoned decision. Artificial intelligence saves serious time in the reading, summarizing and consistency checking parts of these tasks. But be careful: the AI does not generate the score, it helps you understand what the score means and where the file is inconsistent. The credit allocation decision, together with its justification, is signed by the authorized expert.
In this unit, we will see step by step what data credit scoring is based on, at what point you can safely use artificial intelligence, how to verify its output, and which red lines you should never cross.
What is credit scoring based on?
A credit score is an estimate that reduces a customer's probability of repayment to a numerical value. Typical entries are:
- Income and employment: Monthly regular income, working hours, continuity of income source.
- Current indebtedness: Existing loans, credit card debt, debt-to-income ratio (ratio of monthly debt payment to income).
- Credit history: Past payment pattern, delays, legal follow-up records (KKB score summarizes this).
- Collateral: The asset (house, vehicle, deposit) and its value provided against the loan.
- Demand characteristics: Loan amount, maturity, product type.
Tip: The score is a "probability estimate", not a "fact". Two customers may have the same score, but their stories may be completely different. Making a decision without seeing the reasoning behind the score is judging the book by its cover.
Where does artificial intelligence stand in this process?
It is clear where artificial intelligence can be used safely and where it cannot:
Quest
AI role
Who has the decision/approval?
Summarize amounts in an income document
Reading, summary
Confirmation from expert source
Debt/income ratio calculation draft
account draft
Expert recalculates
File consistency check (finding contradictions)
scan
Expert confirms
Draft rejection/approval justification
text draft
Authorized expert author/approvals
Explain the meaning of the score
Description text
Expert reviews
Credit allocation decision
(draft only)
The authority/committee decides
As can be seen, artificial intelligence produces drafts and checks on every line; does not "decide" in any line. Maintaining this distinction is the essence of this unit.
Step by step: preparing a file with artificial intelligence support
- Anonymize. Remove name, TR ID, account number; Describe the customer as "42 years old, public employee, X segment".
- Request a draft. Have AI do file summaries, consistency checks, and draft justifications — not decisions.
- Connect it to the source. Mark which document each number in the printout comes from.
- Recalculate. Verify critical numbers yourself, such as debt-to-income ratio, collateral ratio, etc.
- Policy filter. Check compliance with your internal credit policy and non-discrimination policy.
- Decision and signature. You make the decision, write the rationale, leave the audit trail.
Four copyable templates
1) File consistency check:
Your role: credit file consistency checker. Decision making. Client (anonymous): 42 years old, public employee, monthly net income 55,000 TL, current monthly loan payment 12,000 TL, demand 250,000 TL / 48 months. Task: Are there any mathematical or logical contradictions, missing documents, or inconsistent dates in the information provided? List each finding with the label "[confirmation required]" and specifying what information it is based on. The decision is mine.
2) Debt/income ratio calculation control:
Calculate the debt/income ratio step by step with the following data and show the formula.Just do the calculation, don't make an eligibility decision.Monthly net income: 55,000 TLCurrent monthly debt payment: 12,000 TLEstimated monthly installment of the new loan: 6,500 TLDesired output: current rate, rate including the new loan, formula and intermediate steps.
3) Draft rejection/approval justification (explainable):
Your role: assistant who DRAFT the loan decision rationale. BASED on the verified findings I give you as input, adding new information. Findings: [debt/income ratio 34%, collateral ratio adequate, 1 record of delinquency]Task: Write a draft rationale based on these findings that is explainable and does not contain discriminatory language. No reference to protected characteristics (age, gender, origin, neighbourhood). I will add the final decision and signature.
4) Simple explanation of the score:
Explain to a customer what a credit score means in 4 sentences, using non-accusatory language and without using technical jargon. Do not use personal data, do not make definitive promises. I will check this text before forwarding it to the customer.
Weak prompt / Strong prompt
Weak prompt:
Should we give credit to this customer? Decide based on your score and write your reason.
This request leaves the decision to the model, does not link to the source, and does not prevent the justification from appearing discriminatory.
Powerful prompt:
Your role: decision support assistant, not decision maker. Write a DRAFT rationale based on the verified findings I gave you. Show the finding on which each sentence is based in parentheses. Do not reference protected properties. Mark any unclear points as "[confirmation required]". I will take the final credit decision and responsibility.
The strong will limits the role, connects every claim to the source, prevents discrimination and keeps the decision in the person.
three mini cases
Case 1 — Time savings. An expert performs a consistency check on a portfolio of 30 files. The AI flags 5 discrepancies such as “Income declared in number 8 is 60,000 but the payroll shows 42,000.” The expert confirms it all at the source; 3 of them are real, 2 of them are false readings of the artificial intelligence. A manual 3-hour job is reduced to 40 minutes and no warnings are used without verification.
Case 2 — Recalculation catches the error. The AI summarizes the debt-to-income ratio as “28%.” The expert runs the formula again: existing debt is not included, the real rate is 41%, which exceeds the threshold of domestic policy. Without verification, a false confirmation could occur.
Case 3 — Discriminatory ground denial. Artificial intelligence constructs the phrase "the area where the applicant lives is risky" in a rejection draft. The expert denies this; This statement carries the risk of indirect discrimination (proxy variable) based on district. The justification is rewritten with only legitimate criteria such as debt-to-income ratio and delinquency record.
Beyond the score: context and "borderline" files
The most difficult part of credit decisions is not the open acceptance or open rejection, but the "borderline" files. If the score is too close to the threshold of domestic politics, the decision cannot be left at the mercy of a single number; context comes into play. This is where AI can set a trap: it evaluates every boundary file with the same mechanical logic and misses the nuance. It is the expert who adds the nuance.
Examples of context in which an expert looks at a border file:
- The story of the delay: Was a single delay due to a temporary health problem or a persistent inability to pay? The same “1 delay record” can tell two different stories.
- Nature of income: Two clients may have equal income, but one may be a regular salary earner and the other a variable freelance income; The risk profile is different.
- Relationship with the bank: The file of a customer who has been working smoothly for a long time is read differently than that of a new customer.
AI can help you summarize these contexts, but it's up to the expert to weigh up and decide — especially in borderline cases.
Tip: In cases where the score is close to the threshold, focus on the question "what does this customer's story tell" rather than "what did the model say?" Boundary files are where expert judgment, not automation, shines.
Common mistakes
- Thinking the score is the decision. The score is a prediction; It is not the decision itself. It is wrong to judge based on the score without justification and context.
- Not recalculating the number. Using the rates and totals given by artificial intelligence without confirming them; This is where hallucination or misreading comes in.
- Not noticing the discriminatory justification. Approve justifications that include protected characteristics or surrogate variables such as neighborhood, age, gender; serious legal risk.
- Bypassing the audit trail. Failure to document on which verified finding the decision was based; cannot be defended in subsequent objection or audit.
- Sharing credentials. Entering the tool without anonymizing the file.
Attention: When a credit decision is rejected, the customer has the right to learn the reason and to object. Therefore, the justification must be both correct, explainable and free from discrimination. "The model scored it that way" is not a justification.
In summary
Credit scoring is an estimate of probability based on data such as income, indebtedness, credit history and collateral. In this process, AI produces document reading, consistency checking, account draft and justification draft; But it doesn't change the score or the decision for you. Recalculate each number, attribute each justification to the source, and pass it through the filter of discrimination. The credit allocation decision is signed by the authorized expert, together with the explainable justification. In one sentence: Artificial intelligence prepares the file; A competent person makes the credit decision.
Application task
Define a sample credit file (anonymous: age, occupation, income, current debt, claim, maturity). With templates 1 and 2 above, first do a consistency check and then calculate the debt/income ratio. Manually recalculate and compare the rate given by artificial intelligence. Then generate a justification draft with the 3rd template and check if there is a protected property or surrogate variable in it; Rewrite it with legitimate criteria, if any.
checklist
- [ ] I anonymized the file; I did not share any identification information.
- [ ] I gave the artificial intelligence a draft/control task, not a decision.
- [ ] I recalculated the debt/income ratio and critical numbers by hand.
- [ ] I attributed the rationale to verified findings and cited the source.
- [ ] I checked that there is no protected property/proxy variable in the argument.
- [ ] I have recorded the decision and its justification; I took ultimate responsibility.