Selene

The Conversational AI Implementation Gap

While chatbots brought about a breakthrough in customer care through automation of responses they have largely failed to solve persistent challenges like personalization, adapting to human language, and gauging the context of a customer’s query. Data on customer dissatisfaction with chatbots keeps surfacing, and a recent survey reported 68% consumers⁴ having a bad experience when the chatbot was unable to understand their need or answer their question.

Combining the capabilities of Conversational and Generative AI underpinned by a next-generation processing platform can help issuers overcome these challenges. With these capabilities, an advanced Conversational AI card support solution can deliver:

• A natural language interface enabling intelligent, unscripted conversations
• Intelligent automation of routine query resolution at scale
• Awareness of cardholder intents and context continuity through the conversation
• Personalization of conversation through cardholders’ historical and real-time data
• Intelligent insights and action to meet customer needs
• Contextual escalation to human agents

These capabilities form the foundation of Selene, Zeta’s advanced Conversational AI that reimagines what card support can be.

Introducing Selene: Advanced Conversational AI for Card

Zeta is pleased to introduce Selene, an advanced Conversational AI for cardholders that brings intelligent and personalized card support to life. Selene breaks through the limitations of conventional support solutions to deliver what cardholders truly need - instant and contextual engagement, relevant insights and automated actions - that scale seamlessly for issuers of all sizes and without the limitations of today’s solutions.

This solution paper presents Selene capabilities and impact as follows:

Part 1: Selene: AI-Powered Card Support at Scale

A view of the 4-layer architecture and system orchestration that enable dynamic cardholder experiences and rapid deployment of new workflows on Selene

---

Part 2: System Design: Selene's Architecture and Orchestration

A view of the 4-layer architecture and system orchestration that enable dynamic cardholder experiences and rapid deployment of new workflows on Selene

---

Part 3: Building Trust: Selene’s Banking-Grade Guardrails and Security

Measures to ensure accuracy and reliability of Selene’s responses, prevent hallucinations and deliver quality and regulatory compliance

---

Part 4: Processing Foundation: The Role of Next-Gen Processing Systems in Conversational AI

An introduction to Selene’s core capabilities, its differentiators with respect to current chatbot models, and its enterprise-grade deployment capabilities

---

Part 5: Path to Value: Framework for Conversational AI Adoption

A 4-part adoption framework covering business, operational, compliance and security, and technology considerations to help issuers maximize value realization from Conversational AI

Watch Selene Transform
Cardholder Support Experiences

Experience card support that understands context, responds intelligently, and takes action instantly. See how Selene combines human-like understanding with banking-grade security to deliver support across the card lifecycle. 

Transformed Cardholder Experience

Selene offers cardholders an intelligent support experience through natural, human-like conversation across voice and chat. While conventional apps and automated voice responses force users through static menus, forms, and workflows, Selene leverages Generative AI models that enable intelligent, natural conversations, recognition of user intent, and context continuity in conversations without the constraints of pre-defined scripts.

Cardholders can request real-time insights on their transactions and ask for automated actions through these natural interactions. When situations require human expertise, Selene ensures seamless transitions by intelligently escalating to live agents with full context preserved.

Selene’s ability to understand intent and resolve cardholder queries can be classified as follows:

1. Knowledge Driven Queries

Queries that can be resolved entirely from issuer knowledge base and product documentation.

Examples:

‘How do I apply for a credit card’

‘How do I increase my credit limit’

---

2. Data Driven Queries

Queries that require customer, account, card or transaction data along with the issuer knowledge base to resolve.

Examples:

‘Why was my transaction declined’

‘Why have I not received my card yet’

---

3. Insight Driven Queries

Queries that require customer, account, card or transaction data, analysis on this data as well as the issuer knowledge base to resolve.

Examples:

‘Show my expenses by merchant category’

‘Have I spent more money on groceries or gas last month’

---

4. Action Driven Queries

Queries that require a multi-step workflow leveraging customer, account, card or transaction data, its analysis and the issuer knowledge base to complete the desired cardholder action.

Examples:

‘I’d like to file a dispute against the last transaction on my card’

‘I need to make a payment’

Use Case: Cardholder Making a Payment Using Selene

The following example showcases Selene in action helping a cardholder make a payment against their outstanding balance.

Intelligence Beyond Scripted Chatbots

Most current implementations of chatbots fail to deliver intelligent, personalized customer experiences due to limited context awareness and poor integration with underlying core systems.

Selene’s intelligence, allowing it to overcome these limitations, stems from two powerful integrations:

• Direct access to issuer knowledge bases ensuring responses align with institution-specific policies and operating procedures
• Native integration with Zeta’s next-generation card processing platform, Tachyon to create, read, and update customer, account, card or transaction data in real-time

With these integrations firmly in place, Selene can deliver human-like customer service through capabilities across three important dimensions:

1. Seamless context navigation

Selene maintains continuous conversation context, just like a human agent. Users can modify requests mid-conversation - for example, a user can initiate a request to pay $100 against their credit card bill and change the amount to $150 in the same conversation without restarting the process. When escalation is needed, the entire conversation context is transferred to human agents.

---

2. Banking-grade hallucination prevention

Selene manages the risk of incorrect or arbitrary responses through robust guardrails. Every response is grounded in issuer-specific knowledge bases and follows pre-approved workflows, ensuring accuracy and compliance.

---

3. Human-like conversations

Interacting with Selene reduces cognitive load. Cardholders can express what they want to know or do in everyday language, without worrying about the right terminology. It enables this through four capabilities:

• Advanced Speech Recognition (ASR) for accurate voice command interpretation
• Natural Text-to-Speech (TTS) for clear vocal responses
• Precise Intent Detection to understand user needs
• Contextual Dialogue Generation for natural conversation flow

Enterprise-Grade Deployment Capabilities

Selene’s integration with a next-generation card processing core and the issuer knowledge base also enables enterprise-grade deployment capabilities like:

1. Omnichannel servicing capability

Selene provides consistent support across voice and chat platforms

---

2. Embeddability

Selene can be natively embedded into any modern mobile/web app or industry-leading contact center platform, for example, AWS Connect, Genesys, NICE or others

---

3. Monitoring and analytics

Issuers can access detailed performance data of all Selene interactions for in-depth analysis

---

4. Compliance and security

Selene has built-in guardrails to ensure it addresses only the intended issues while adhering to industry compliance and security standards

Multi-Layer Intelligence Framework

Architecturally, Selene is composed of 4 distinct layers (Image 3):

1. A central conversation orchestration engine manages all conversations, whether voice or chat
2. Advanced AI models that understand and execute requests through
   • Intent Detection
   • Retrieval Augmented Generation
   • Code Generation
   • Other models including image recognition
3. Advanced toolkit to
   • Run code
   • Run Text-to-Speech, Automatic Speech Recognition
   • Execute function calling and others
4. Integration layer that connects to
   • Zeta’s card processing core for API calls and operations
   • The issuer document repository comprising its standard operating procedures, policies and knowledgebase

This architecture enables Selene to dynamically answer questions, generate insights, and execute actions without requiring specific code to be written for each.

Orchestrating Intelligent Insights and Actions

Let’s see how it works with an example (Image 4). The conversation orchestration engine receives the cardholder request to view their recent purchases by category, visualized in a pie chart. This triggers the foundational models for Intent Detection, Retrieval Augmented Generation, Code Generation, etc., allowing Selene to understand and execute the request being made.

Intelligence Trained on Issuer Knowledge

Selene builds its knowledge foundation exclusively from issuer-approved sources:

• Detailed knowledge repositories
• Standardized operating procedures
• Approved workflow documents
• Compliance guidelines

This focused training ensures Selene operates strictly within defined boundaries, eliminating the risk of unauthorized actions or incorrect information.

Multi-Layer Hallucination Prevention

Selene employs four critical components to prevent AI hallucinations:

1. Intent-focused action generator

Selene’s Action Generator validates user intent before directing conversations through pre-approved flows, keeping the risk of hallucinations near zero.

---

2. Enterprise-grade language model

Selene’s response generation system uses an enterprise-grade large language model independently verified for its accuracy and minimal hallucination risk.

---

3. Robust prompt engineering and guardrails

Selene’s comprehensive guardrails combine precise prompt engineering with strict boundary enforcement. Every response is validated in real-time against approved topics and functions, ensuring conversations remain focused and accurate.

---

4. Rigorous input management

Each interaction undergoes systematic validation through sophisticated anti-jailbreak protections to maintain robust security against potential system manipulations.

---

Quality Assurance

Selene maintains peak performance through continuous response monitoring and systematic scenario testing. Edge cases undergo careful evaluation, with immediate corrective action capabilities ensuring consistent accuracy. This proactive approach to quality control ensures Selene consistently meets the exacting standards of card issuers while delivering the fluid, natural interactions cardholders expect.

Regulatory Compliance

Selene is engineered to support card issuers’ compliance requirements. Every implementation adheres to critical banking and payments industry standards like:

1. Payment Card Security

• Built to PCI-DSS specifications for card data handling
• End-to-end encryption of sensitive information
• Secure data transmission protocols
• Controlled access to cardholder data
• Compliant with operational security protocols

2. SOC 2 Type II certified infrastructure and processes

• Regular security audits and assessments
• Robust data privacy protocols
• Secure cloud infrastructure

Its integration with the issuer processing platform ensures that it inherits the platform’s banking compliance framework and meets the data retention and privacy requirements.

The 4 Barriers to Conversational AI Adoption on Legacy Systems

Advanced AI requires instant access to card, account and transaction data through direct card processor integration. Legacy card processors, with their batch processing and rigid workflows, can’t support the real-time operations that a Conversational AI demands.

While legacy card systems have multiple constraints, the following 4 limitations critically impact AI-powered support performance.

1. Fragmented data and systems integration

Legacy systems store data in silos across multiple subsystems. This fragmentation makes it difficult for Conversational AI to aggregate information and generate coherent, context-aware responses. While middleware can consolidate data between systems, it adds significant costs and delays, creating bottlenecks that slow down AI implementation.

---

2. Outdated messaging protocols Leading to Reduced Interoperability

Legacy systems store data in silos across multiple subsystems. This fragmentation makes it difficult for Conversational AI to aggregate information and generate coherent, context-aware responses. While middleware can consolidate data between systems, it adds significant costs and delays, creating bottlenecks that slow down AI implementation.

---

3. Lack of Real-Time Event Handling

Legacy systems have limited real-time processing capabilities. This limits the AI’s ability to respond dynamically to state changes and user interactions, often resulting in delayed responses and outdated information.

---

4. Limited Availability of Purpose-Built APIs

Without comprehensive APIs, Conversational AI cannot deliver personalized support or automated workflows. It needs real-time access to cardholder data to provide tailored responses and alerts. Lacking these APIs, it can only offer generic answers, forcing users to rely on traditional support channels for anything beyond basic queries.

The 4 Next-Generation Processing Enablers of Conversational AI

Delivering AI-powered personalized, and conversational cardholder support demands a next-generation processing platform built for real-time data access and action across systems – such as Zeta’s Tachyon processing platform.

Tachyon delivers four key enablers for advanced Conversational AI solutions like Selene to deliver on their promise:

1. A comprehensive API catalog that supports every create, read, update operation in scope for the AI solution.

For example, Selene’s access to Tachyon’s extensive API library enables it to perform actions typically handled by agents, including payment processing, account freezing/unfreezing, transaction searches, and adding payment methods.

---

2. Real-time events stream that allow the AI to learn and adapt its actions and decisions.

Selene adapts continuously through real-time events and conversation context, enhancing decision-making at scale. For example, during payment conversations, Selene can display balances due after accounting for partial payments in real-time.

---

3. Open data stack for continuous supervised learning of the AI models on comprehensive user data available in structured data lakes.

Selene analyzes varied phrases used in real user conversations to request payments, such as ‘clear my balance’ or ‘settle my dues’ to boost its recognition of payment intents in natural language.

---

4. Hyper-personalization and configurability of the AI’s conversational flows, enabling unique and context-aware responses to cardholder needs.

For example, a cardholder can ask Selene to freeze a lost card in natural conversational style, eliminating the need to locate specific app functions.

Essential considerations under each dimension

1. Business Considerations

• Objective Definition: Prioritize business objectives, e.g. reducing customer service costs, enhancing user satisfaction, or improving resolution times
• Customer Expectations: Identify customer needs and preferences for personalization, quick resolutions, and accessible support across multiple channels
• Competitive Analysis: Benchmark capabilities of Conversational AI solutions offered by leading issuers to ensure competitiveness of your solution
• Expected Outcomes: Quantify expected outcomes. e.g. reduced agent-handled queries, faster response times, and higher customer satisfaction ratings

---

2. Operational Readiness

• Scope: Ring fence the scope of service that your Conversational AI can provide. e.g.
  1. Information like balances, statements, or billing cycle
  2. Financial transactions like repayments
  3. Advanced insights like spend categorization
• Omnichannel Service Delivery Support: Evaluate what is needed for the assistant to be seamlessly accessible via the mobile app, website, and IVR, catering to different customer preferences.
• AI Agent - Human Agent Handoff: Implement a smooth handoff process to human agents when queries go beyond the scope of the AI agent, with the context preserved in transitions.
• Continuous Feedback Loop: Integrate performance analytics and customer feedback to iteratively improve the assistant’s responses and expand its capabilities over time.

---

3. Compliance and Security

• Data Privacy Compliance: Adhere to CARD Act, CCPA, and other regulations for customer data, ensuring consent and robust data encryption.
• Banking-Grade Guardrails: Prevent unauthorized actions or responses by limiting the AI’s access strictly to predefined, bank-approved workflows and knowledge sources.
• Security and Fraud Prevention: Employ strict anti-fraud measures, including transaction verification, and ensure the Conversational AI detects any signs of suspicious activity.
• Hallucination Prevention: Integrate multiple layers of control to ensure accurate responses, such as intent-focused action generators, input validation, and anti-jailbreak protections to maintain reliable, secure interactions.

---

4. Technology & Infrastructure

• Next-Gen Card Processing Platform: Leverage a cloud native, real-time card processing platform to handle API requests, track transactions, execute payments, and fetch details to ensure quick and accurate responses.
• Comprehensive API Catalog: Ensure the availability of an extensive API library to support all types of customer interactions, from simple queries to complex requests like payment processing and account management.
• Scalability and Resilience: Design the architecture to scale on demand, with cloud-native infrastructure capable of handling high volumes and concurrency without degradation in response time.

The successful implementation of an advanced Conversational AI like Selene for card support represents a significant leap forward in modern banking services. By following this comprehensive framework - encompassing business considerations, operational readiness, compliance protocols, and technological infrastructure - financial institutions can deliver a solution that not only meets current customer needs but also scales for future demands.

Bibliography

References