AI-Assisted Decision Support for Payroll Credit Operations: Design, Implementation, and Outcomes of an Intelligent Customer Service System in a Regulated Financial Environment

This paper presents the design, implementation, and measured outcomes of an AI-powered customer support system deployed in a regulated payroll credit (crédito consignado) environment in Brazil. The system was developed to address a well-documented operational challenge in financial services contact centres: customer service agents must navigate extensive product knowledge bases — in this case, over forty product files covering rates, eligibility rules, and client-type-specific constraints — while simultaneously conducting live customer interactions. The resulting inconsistency in agent responses, dependency on experienced individuals, and absence of customer context aggregation created measurable service quality and compliance risks. We describe an intelligent decision support system built on AWS cloud services, including AWS Bedrock for LLM-based reasoning, Amazon S3 and OpenSearch for knowledge base management, Amazon RDS and DynamoDB for customer data, and AWS Lambda for real-time orchestration. The system profiles customer eligibility in real time — identifying client type (civil servant, INSS retiree, pensioner), calculating available consignable margin, and applying product-specific eligibility rules — and surfaces qualifying questions and ranked product recommendations to agents during live interactions. The system operates as an assistive tool augmenting human judgement rather than replacing it, which was determined to be the appropriate design model for a regulated environment where accountability for credit decisions is legally significant. Outcomes include approximately 85% improvement in decision accuracy and consistency across the customer service team, 73% reduction in manual knowledge lookup time, 90% operational efficiency improvement across the CredCesta payroll loan platform, 99.95% system availability, and 10% utilisation of AWS AI tooling across the organisation. A parallel AI onboarding assistant structured individual development plans for 100% of employees, transforming product knowledge from a property of individuals into a property of the organisation. We identify five design principles transferable to practitioners building AI-assisted decision support in regulated financial contexts: assistive design as the correct model for regulated environments; knowledge base governance as equally important as model quality; data residency requirements as an early architecture constraint rather than a retrofit; context aggregation as the primary value driver over model sophistication; and qualifying question generation as more operationally valuable than product recommendation generation.