Generative AI Implementation in Banking Customer Service and Knowledge Management

various · Finance · 2023 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: banking customer service and knowledge management

Multiple banks described how they are using generative AI on top of cloud data platforms and contact-center knowledge bases while meeting financial-services expectations for risk management, explainability, human oversight, data quality, and measurable customer-service improvement.

Constraints:

• financial services requires stronger explainability, model governance, and risk management than many other domains
• customer-service agents increasingly handle only complex questions because simple banking tasks are self-service
• agents must search long procedure documents and multiple knowledge sources while staying in a customer conversation
• banks need usable PII-enabled cloud data without treating risk prevention as the only goal
• business teams must own knowledge-base and data quality for generative AI to work
• generated procedures still require traditional legal and compliance review
• scaling use cases requires risk frameworks and acceptable boundaries, not only engineering delivery

Architecture

The panel describes a banking pattern built on cloud data foundations, governed data platforms, RAG over contact-center knowledge bases, human feedback loops, and AI risk councils. Scotia put generative AI over a knowledge base for contact-center agents and improved answer quality by having the contact-center team clean acronyms, titles, old content, and incorrect content. Discover built a generative AI risk council, then used Vertex AI/Gemini and RAG to summarize relevant procedures for agents, attach links back to source documents, collect thumbs-up/down and qualitative feedback, and evaluate outputs against golden answers.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  cloud_data_foundation["Cloud data foundation"] --> knowledge_base["Knowledge base"]
  business_data_owners["Business data owners"] --> knowledge_base
  business_data_owners --> risk_and_ai_council["Risk and AI council"]
  knowledge_base --> rag_summarization_layer["RAG summarization layer"]
  risk_and_ai_council --> rag_summarization_layer
  rag_summarization_layer --> contact_center_agent_interface["Contact-center agent interface"]
  contact_center_agent_interface --> contact_center_agents["Contact-center agents"]
  contact_center_agent_interface --> agent_feedback_loop["Agent feedback loop"]
  agent_feedback_loop --> golden_answer_evaluation_tools["Golden-answer evaluation tools"]
  golden_answer_evaluation_tools --> rag_summarization_layer
  agent_feedback_loop --> knowledge_workers_and_technical_writers["Knowledge workers and technical writers"]
  knowledge_workers_and_technical_writers --> knowledge_base
  knowledge_workers_and_technical_writers --> legal_and_compliance_review["Legal and compliance review"]
  legal_and_compliance_review --> knowledge_base

Components

Data flow

Cloud data and knowledge-base content form the foundation. Business teams clean and contextualize knowledge articles, including acronyms and outdated or incorrect content. A RAG/summarization layer retrieves relevant procedures or knowledge-base content for contact-center agents, returns summaries with links back to source documents, and keeps agents in the customer conversation. Agent feedback, expert reviews, golden-answer comparisons, and AI council risk ratings drive iterative tuning and production gating.

Why this architecture

The panelists frame banking generative AI as a data-quality and risk-management program as much as an engineering project. Cloud platforms and RAG make answers faster, but banks need business-owned knowledge cleanup, source links, golden-answer evaluation, human feedback, legal/compliance gates, and cross-functional risk councils to stay inside regulated-industry boundaries.

Models used: Google Dialogflow · Vertex AI · Gemini · generative AI models

Scale

• Macquarie described being 96.4% in public cloud at the time of the panel.
  “we wanted to be a 100% public Cloud um uh and we're 96.4 this morning”transcript, stated
• Discover started building its digital assistant with Google Cloud/Dialogflow in 2017.
  “we started in 2017 when we set out to build the Discover digital assistant and uh you know that's something that runs on Google dialogue flow”transcript, stated
• Scotia described hundreds of productionized models, each traditionally taking around 18 months to build.
  “we have hundreds of models already deployed and productionize across the bank”transcript, stated

  “hundreds of models deployed it takes 18 months to build each one of those”transcript, stated
• One banking AI program targeted 500 million euros of bottom-line benefit by 2025.
  “we have objective by 2025 to produce 500 million bottom line benefit in euros”transcript, stated
• Discover reported roughly 70% search-time reduction for real agents on real customer questions.
  “we cut the search Time by about 70% and that's with real agents on real questions that they face with customers”transcript, stated

Alternatives considered (roads not taken)

Treat cloud security as pure risk prevention

Rejected because: Scotia argued banks need risk management, not risk prevention, because data must remain usable for customers and employees.

Build cloud/data platforms and assume the business will come later

Rejected because: Scotia described the classic 'if you build it they will come' mistake and moved toward business engagement and ownership.

Use generative AI before cleaning knowledge-base content

Rejected because: The panel says answer quality improved because the contact-center team cleaned acronyms, titles, outdated content, and incorrect content.

Let employees freely use open generative AI tools

Rejected because: Discover established enterprise capabilities and risk boundaries rather than allowing open tool use at leisure.

Replace legal and compliance review for AI-assisted procedure writing

Rejected because: Discover kept the same human attorney and compliance officer reviews for AI-assisted procedures as traditional procedures.

Patterns

rag-retrieval (contact-center knowledge access)

Discover uses RAG to tune summaries to its procedure documents, return direct links to source documents, and help agents find the exact relevant spot.

data-quality-loop (business-owned knowledge cleanup)

Scotia's contact-center team improved answer quality by taking ownership of knowledge-base content, including acronyms, titles, outdated content, and incorrect content.

human-in-the-loop-gate (banking review and feedback)

Banks keep humans in the loop through agent feedback, expert review, legal review, compliance review, and explicit AI program principles.

eval-driven-development (golden-answer and risk evaluation)

Discover combines qualitative agent feedback, golden-answer tools, and council risk scaling to tune output quality and decide which use cases go to production.

safety-policy-enforcement (regulated-industry governance)

The banking AI programs require principles such as explainability, fairness, human-in-the-loop review, data quality, legal/compliance review, and secure enterprise capabilities.

observability-tracing (source-linked answers and monitoring)

Discover's agent-assist summaries include direct links back to referenced documents, and Scotia discusses platform needs for monitoring, drift management, cataloging, validation, and documentation.

Tradeoffs

Enable usable PII in cloud rather than locking data away.

Chose Risk-managed PII-enabled cloud data over Security so strict that end users cannot use the data.

Rationale: The bank needed PII to serve customers and personalize experiences, while still managing risk appropriately.

Start with lower-risk enterprise use cases.

Chose Least risky or least disruptive use cases through an intake and risk framework over Open employee use of public generative AI tools.

Rationale: Discover wanted organization-wide comfort, secure enterprise capabilities, and risk-scaled experimentation before production deployment.

Use expert-agent feedback to tune summaries.

Chose Skill-group-by-skill-group tuning with knowledge workers and agent feedback over Pure platform tuning by engineers.

Rationale: The heavy work was determining which answers are high quality for each procedure set, so experts and agents had to guide the loop.

Return summaries with source links rather than only answer text.

Chose RAG summaries linked to the referenced document and spot over A generated answer without source navigation.

Rationale: Agents need to trust and act on answers during regulated customer calls, so direct references improve search accuracy and usability.

Gotchas (what bit them)

Overprotecting data can make it useless.

Root cause: Early cloud programs treated data protection as an absolute objective rather than balancing usability and risk management.

Resolution: Scotia moved toward a more rational approach to security that permits PII-enabled cloud use for customer and employee service.

Generative AI exposes knowledge-base quality problems.

Root cause: The first knowledge-base RAG attempt produced low answer quality because acronyms, titles, old content, and incorrect content were not business-owned and cleaned.

Resolution: The contact-center team cleaned and contextualized the knowledge base themselves, raising answer quality dramatically.

Digital banking makes the remaining agent calls harder.

Root cause: Simple questions move to self-service channels, leaving contact-center agents with difficult, procedure-heavy customer issues.

Resolution: Discover used generative summarization and RAG to help agents find procedure answers without putting customers on long holds.

Tuning summaries is harder than coding the platform.

Root cause: Vertex/Gemini integration made the traditional engineering work quick, but making summaries relevant to procedures required skill-group-by-skill-group experimentation.

Resolution: Discover involved expert agents and technical writers to run tuning cycles and provide qualitative feedback.

AI-assisted procedure writing still needs standard banking controls.

Root cause: Generated or rewritten procedures affect regulated operations and cannot skip legal or compliance checks.

Resolution: Discover kept human attorney and compliance officer review exactly as it would for traditional procedures.

Evaluation

The panel describes several evaluation modes: answer-quality improvement from Scotia's knowledge-base cleanup, Discover's real-agent search-time reduction, thumbs-up/down and qualitative feedback from agents, automated golden-answer comparison, risk-council scoring, and standard legal/compliance review for AI-assisted procedure writing.

Metrics: answer quality from about 33% to about 96% · search-time reduction of about 70% · minutes or forever reduced to seconds · thumbs-up/down agent feedback · qualitative answer-quality feedback · golden-answer approximation · risk framework rating · explainability, fairness, human-in-the-loop, and data quality principles

Outcomes

• Scotia's contact-center knowledge-base use case improved answer quality from about 33% to around 96% after business-owned content cleanup.
• Discover's RAG summarization reduced agent search time by about 70% on real customer questions.
• Agent search moved from minutes or indefinite searching to seconds, helping agents stay in conversation flow.
• Discover built an organizational generative AI risk framework before productionizing use cases.
• The panel emphasizes that data quality ownership, human review, and risk governance are as important as model choice.

The panel spans multiple banks and maturity levels rather than one deployment. It cites Macquarie's 96.4% public-cloud progress, Discover's 2017 Dialogflow assistant foundation, Scotia's hundreds of productionized models, an AI program targeting 500 million euros of benefit by 2025, and Discover's 70% real-agent search-time reduction.