The AI models
Five models that run the full behavioral credit workflow
From the first conversation to an auditable risk index — each model does one job well, and every one is decision support: evidence-backed, human-reviewed, never an automated verdict.
Interview · phone or chat
AI Credit Interviewer
Conducts the loan conversation like an experienced credit officer — one question at a time, over phone or text. It clarifies the business, finances, amount and purpose, repayment plan and collateral; explains why it asks; never pressures; and closes with a structured recommendation — approve, decline, or request completion — grounded only in what was actually said. Built from six real recorded credit calls.
Output → a complete, consistent application file + a reasoned recommendation, with a clear note that the lender makes the final call.
Structured interview
Behavioral Interview Battery
Administers a fixed 31-item behavioral & socioeconomic battery through natural conversation, in the borrower's language, with tappable response options — collecting responses only, never advising.
Output → clean, coded responses ready for scoring.
Deterministic scoring
Scoring Engine
Turns the interview into item-level risk scores, subscale totals and a composite Default Risk Index. Fully deterministic — the same input always yields the same output, so every score is auditable.
Output → Default Risk Index + subscale breakdown, imputed-data flagged.
Evidence extraction
Loan Risk Variable Extraction
Reads any transcript or application and extracts factors known to predict repayment — finances, income stability, financial literacy, behavior, repayment realism. Every finding is tied to a verbatim quote and a confidence level, with inconsistencies and red flags surfaced.
Output → a decision-support table with quotes, confidence, red flags and follow-up questions.
Personality & veracity
Personality & Veracity Analysis
Assesses Big Five traits and a Veracity score across one or more interviews — flagging linguistic patterns that signal a front-man (målvakt) or a fabricated business, and summarizing how the profile bears on the credit rating.
Output → trait scores with quoted justification + a fraud/veracity flag.
Validated against simulated applicants — a companion model plays realistic Swedish borrowers with distinct finances and personalities, so the assessment models are tested before they ever meet a real case.
