Helping Customers in Distress: An LLM-powered Agent that Converses, Probes, and Routes

Technical Summary: Helping Customers in Distress – An LLM-Powered Agent That Converses, Probes, and Routes

Problem Statement
The UK financial sector faces a rapidly increasing volume of unauthorized fraud, scams, and disputed transactions, rising over 10% year-on-year. Traditional customer service mechanisms, such as menu-based Interactive Voice Response (IVR) systems and manual human triage, are becoming unsustainable due to their inflexibility, slow processing times, and high rates of misrouted cases. These inefficiencies lead to customer dissatisfaction and operational bottlenecks, preventing specialist teams from focusing on complex investigations. There is an urgent need for a scalable, accurate, and compliant automation solution that can effectively triage these cases within the customer journey.

Methodology
The authors propose a customer-facing AI triage agent powered by Large Language Models (LLMs) designed to conduct multi-turn conversations, ask targeted probing questions, and classify cases into specific categories (Fraud, Scam, Dispute, or Inconclusive).

• Agent Architecture: The system utilizes third-party LLMs (Claude, Gemini, GPT series) via prompt engineering rather than fine-tuning, ensuring flexibility and reduced data risks. The agent is orchestrated by a prompt manager containing specialized sub-prompts that define the agent's role, behavioral instructions, investigative workflow, and prohibitions ("Don'ts"). These prompts embed the bank's internal policies directly into the agent's reasoning.
• Safety and Hand-offs: The system employs layered safety guardrails. AWS Bedrock provides foundational filtering, supplemented by custom input controls blocking unauthorized product requests, internal process probing, and non-English messages. A dedicated "digression agent" detects vulnerability indicators or user desires to end the conversation, triggering a hand-off to human specialists via direct phone lines.
• Evaluation Framework (Digital Twins): To address the lack of labeled chat logs for training and testing, the authors developed a scalable evaluation pipeline using synthetic "digital twin" customers. These agents are generated using OpenAI Agent SDK, mimicking real customer behavior, language, and interaction styles by drawing on historical telephony transcripts and legacy case data.
• Validation Strategy: Performance is assessed through a combined approach:
  1. Synthetic Testing: Large-scale simulation of ~3,000 transcript-based cases.
  2. Human Evaluation: Subject Matter Experts (SMEs) engage in multi-turn dialogues via a web interface, rating the agent on metrics including satisfaction, empathy, compliance, factuality, and summary quality.
  3. Automated Evaluation: An "LLM-as-a-judge" (GPT-4.1) pipeline evaluates conversations against the same metrics to provide scalable, binary ratings and rationales.
  4. Red-Teaming: Adversarial testing was conducted to assess robustness against prompt injection, code manipulation, and extraction attempts.

Key Contributions

1. Deployed Triage Agent: A functional LLM-based agent capable of multi-turn dialogue for accurate case classification with integrated handoff protocols and layered safety guardrails.
2. Scalable Evaluation Pipeline: A reproducible framework that unites operational data with synthetic "digital twin" customers, enabling thorough validation without relying on pre-existing labeled chat logs.
3. Hybrid Validation Strategy: A combined human and automated (LLM) validation approach for continuous prompt tuning and quality assurance, demonstrating readiness for real-world banking deployment.

Results
The agent was evaluated against a legacy IVR baseline across synthetic scenarios and SME-led testing:

• Classification Accuracy: The system achieved significant accuracy gains. GPT-5 showed the highest improvement at +30.6% over the legacy baseline, followed by GPT-4.1 (+27.7%), Gemini-1.5-Pro (+28.4%), GPT-4.1-mini (+21.3%), and Claude Sonnet 3 (+20.0%). SME testing confirmed an average improvement of +16.0%.
• Operational Metrics: Key metrics including compliance, satisfaction, and summary quality consistently exceeded the 75–80% threshold.
• Hand-off Performance: The hand-off agent demonstrated precision and recall rates exceeding 90%, effectively detecting when human intervention is required.
• Safety and Guardrails: Input guardrails detected prompt injection and hate speech with >98% accuracy. Output guardrails identified hallucinatory content with >95% accuracy. Red-teaming exercises confirmed the system's ability to block adversarial prompts and prevent the disclosure of internal reasoning.
• Evaluation Agreement: Automated and human ratings showed strong agreement (79–92%) for objective metrics (compliance, factuality, relevance) but lower agreement (~60%) for subjective aspects like empathy and frustration, suggesting a need for human-in-the-loop validation for emotionally nuanced cases.

Significance
The paper claims that the proposed LLM-powered triage agent successfully addresses the scalability and accuracy challenges of current banking fraud and dispute handling. By leveraging synthetic digital twins for evaluation, the authors demonstrate a robust method for validating AI agents in regulated environments where labeled data is scarce. The results indicate that targeted probing via multi-turn dialogue leads to more effective triage, reducing customer frustration and operational delays. The system's high accuracy, combined with rigorous safety guardrails and compliance with internal policies, confirms its suitability for deployment in high-stakes, regulated financial domains.