Kyndryl LoanIQ - Agentic Loan Application Review on AWS

Kyndryl LoanIQ uses AWS serverless, Amazon Bedrock, and AgentCore to automate first-pass loan reviews—structuring documents, validating consistency, and generating scorecards—reducing manual work while keeping final decisions human-led.

How it works on AWS Applicants and loan officers access the LoanIQ web experience through a React frontend hosted on AWS Amplify, with authentication supported by Amazon Cognito. The applicant workflow captures the required loan document set and creates a loan case, while the officer workflow presents case details, supporting documents, validation status, AgentCore scorecard, risk flags, and approve/reject controls.

Uploaded documents are stored in Amazon S3 and processed through AWS Step Functions. AWS Lambda tasks classify documents, run Amazon Bedrock Claude Vision extraction, persist extracted fields, validate the case, and invoke the loan review path. Amazon DynamoDB stores the case state, extracted evidence, validation rows, AgentCore scorecard, and final pre-review output used by the frontend.

Agentic loan review with Amazon Bedrock AgentCore Amazon Bedrock Claude Vision extracts structured fields from NRIC, payslip, EPF statement, bank statement, and employment letter submissions. Amazon Bedrock AgentCore then runs a supervisor with specialist review agents for document quality, applicant consistency, income and banking evidence, and risk exceptions.

The AgentCore review produces a 100-point scorecard, recommendation, confidence, rationale, risk flags, document findings, and next-best officer action. Results are saved back to Amazon DynamoDB as agent scorecard and agent_pre-review rows so the review dashboard can display evidence-backed recommendations.

Key AWS services

• Frontend, compute, and orchestration: AWS Amplify, AWS Lambda, AWS Step Functions.
• Generative AI and agentic review: Amazon Bedrock Claude models, Amazon Bedrock AgentCore.
• Storage and case state: Amazon S3, Amazon DynamoDB.
• Security and observability: Amazon Cognito, AWS IAM, AWS KMS, AWS WAF, Amazon CloudWatch.

Key solution capabilities

• Document extraction: Claude Vision extracts normalized evidence from the five required loan document types.
• Validation and scorecard: Rules check document completeness, identity consistency, income signals, banking evidence, and exception flags.
• Officer review: The dashboard surfaces the recommendation, score breakdown, cited findings, risk flags, and next-best action for human decisioning.

What buyers get

• Faster first-pass review: Routes submitted loan applications through automated extraction, validation, specialist scoring, and officer-ready recommendations.
• Structured evidence capture: Extracts key fields from NRIC, payslip, EPF statement, bank statement, and employment letter documents.
• Consistent scorecard: Applies a repeatable 100-point review model across document quality, applicant consistency, income and banking, and risk exceptions.
• Earlier exception detection: Highlights name, NRIC, employer, salary, EPF, bank ownership, and document-quality mismatches before approval.
• Officer-ready workflow: Presents recommendation, confidence, rationale, risk flags, specialist findings, document view controls, and approve/reject actions.
• Audit-ready case state: Persists source documents, extraction rows, validation outputs, AgentCore scorecards, final pre-review results, and decision records.

Typical flow

• Applicant or loan officer signs in through Amazon Cognito;
• Applicant uploads NRIC, payslip, EPF statement, bank statement, and employment letter through the LoanIQ portal;
• Documents are stored in Amazon S3 and linked to a loan case;
• AWS Step Functions orchestrates classification, extraction, persistence, validation, and loan review;
• Amazon Bedrock Claude Vision extracts structured document evidence;
• Validation checks completeness, identity consistency, employment, income, EPF, and banking signals;
• Amazon Bedrock AgentCore runs the supervisor and specialist review agents;
• AgentCore writes scorecard and pre-review outputs into Amazon DynamoDB;
• The officer dashboard displays findings, risk flags, recommendation, and next-best action;
• The loan officer records the final approve/reject decision for the case.

Deployment & fit

Designed for banks, SME lenders, loan operations teams, credit review teams, and digital lending units that need to accelerate document-heavy loan application review while maintaining human oversight, auditability, and consistent exception handling.

Expected outcomes:

Reduced manual first-pass review effort, faster identification of document and identity mismatches, more consistent scorecard-based review, clearer officer decision support, and audit-ready records across extraction, validation, AgentCore review, and final decisioning.