Real-time AI Agent Assistance in Contact Center Operations

us_bank · Finance · 2025 · source Agentic: no · promotional bias: med · 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: Financial-services contact center agent assistance

US Bank needed to help contact-center agents answer customer calls in real time while reducing manual knowledge-base search, post-call documentation burden, transfers, latency, and compliance risk.

Constraints:

• Voice interactions require answers in real time, not after a delayed retrieval or generation step.
• Agents work from different skills, job aids, and knowledge bases, so recommendations must be scoped to the right domain.
• The system operates in banking, requiring careful guardrails, prompt-injection defenses, and business-line/SME input.
• The rollout was still a limited production pilot, so enterprise-scale claims needed ongoing validation.

Architecture

US Bank built a production-pilot agent-assist system on Amazon Contact Lens, Amazon Q in Connect, Amazon Bedrock, and Anthropic Claude. Contact Lens transcribes calls in real time, Q in Connect detects actionable customer intents, agents choose which intents need recommendations, and Bedrock/Claude generate answers from tagged US Bank knowledge bases stored through S3-managed Q in Connect RAG/vector infrastructure.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  A["Customer voice call"] --> B["Amazon Contact Lens"]
  B --> C["Amazon Q in Connect"]
  C --> D["Agent desktop"]
  D --> E["Contact center agent"]
  C --> F["Tagged knowledge bases"]
  G["Knowledge data pipeline"] --> F
  G --> H["S3 buckets"]
  H --> C
  C --> I["Amazon Bedrock"]
  I --> J["Anthropic Claude"]
  C --> K["LLM evaluation and observability"]
  I --> K

Components

Data flow

A customer call is transcribed by Contact Lens in real time. Q in Connect reads the transcript, distinguishes real intents from ordinary conversation, and presents intents to the agent. When the agent chooses an intent, Bedrock with Anthropic Claude generates a recommendation against the relevant tagged knowledge base. Knowledge-base articles are cleansed through a pipeline before upload to S3, where Q in Connect builds the internal vector/RAG stores.

Why this architecture

US Bank avoided a simpler standalone Bedrock knowledge-base setup because voice-channel assistance has a strict real-time latency requirement. Q in Connect plus Contact Lens provided real-time transcription, intent detection, agent-controlled recommendation requests, and tagged knowledge access that matched contact-center role boundaries.

Models used: Anthropic Claude · Amazon Bedrock models

Scale

• US Bank describes itself as the fifth-largest financial bank in the United States.
  “US bank is a fifth largest uh financial bank in in United States”transcript, stated
• The bank has presence in 13 countries and around 70,000 employees.
  “We we have presence in 13 countries uh globally uh including u Canada, US, Europe and we have around 70,000 employees”transcript, stated
• The retail branch footprint is about 2,000 US branches.
  “Our retail branch uh uh presence within USA is 2,000 branches across uh east, west and south.”transcript, stated
• The contact-center estate spans around 100-plus global contact centers.
  “we do have around 100 plus global contact centers. So it's a it's a massive contact center”transcript, stated
• The current deployment is a limited production pilot for specific business lines and agent groups.
  “it's limited. It's only for specific business lines and for specific group of agents.”transcript, stated

Alternatives considered (roads not taken)

Use Amazon Bedrock knowledge bases alone

Rejected because: The voice channel needed lower-latency real-time agent assistance than a standalone knowledge-base approach provided.

Generate recommendations for every detected intent

Rejected because: The live conversation may contain multiple intents and experienced agents may not need help for all of them.

Let the AI search across all company knowledge bases

Rejected because: Agents are scoped to specific job aids and knowledge domains, so AI recommendations need the same access boundaries.

Patterns

rag-retrieval (retrieval)

The system retrieves against tagged knowledge bases and Q in Connect-managed vector/RAG stores to generate agent recommendations.

human-in-the-loop-gate (governance)

Agents choose which detected customer intents should receive AI recommendations rather than letting the system answer every possible intent.

safety-policy-enforcement (governance)

Knowledge-base tags, Bedrock guardrails, Q in Connect guardrails, temperature controls, and prompt-injection defenses constrain recommendations for banking use.

eval-driven-development (quality)

The team is adding LLM-as-judge/evaluator tooling and using pilot feedback to decide scaling, segmentation, and performance modeling.

observability-tracing (operations)

US Bank treats monitoring, issue detection, alert configuration, and change management as part of the generative AI rollout.

Tradeoffs

Run as a limited production pilot before enterprise rollout.

Chose Production pilot scoped to specific business lines and agents over Immediate full contact-center deployment.

Rationale: The team wanted real operational feedback while limiting risk and identifying the right segments for AI assistance.

Keep customer data out of the current recommendation flow.

Chose Knowledge-base-only recommendations over Customer-data-personalized recommendations.

Rationale: The current pilot narrows privacy and implementation risk, though it limits personalization.

Let agents request recommendations after intent detection.

Chose Agent-controlled recommendation requests over Automatically generating every recommendation.

Rationale: Experienced agents may already know answers, and real-time conversations may surface more intents than need AI help.

Gotchas (what bit them)

Voice agent assist is latency-sensitive in a way that generic knowledge retrieval is not.

Root cause: Agents need answers while the customer is still on the call, so delayed recommendations lose value.

Resolution: US Bank combined Contact Lens real-time transcription with Q in Connect orchestration rather than relying only on Bedrock knowledge bases.

AI recommendations can cross knowledge-domain boundaries if retrieval is not scoped.

Root cause: Different agents and business lines use different knowledge bases and job aids.

Resolution: The system uses Q in Connect tagging to restrict searches to the appropriate knowledge base for the call/agent.

Bad data becomes bad agent assistance.

Root cause: RAG answers are only as reliable as the underlying knowledge-base articles.

Resolution: US Bank built an automated cleansing pipeline before uploading knowledge articles into S3/Q in Connect.

Evaluation

US Bank evaluates through a production pilot with latency, intent detection, answer quality, user feedback, and planned automated LLM evaluation/observability tooling.

Metrics: Intent detection quality · Recommendation answer quality · Real-time latency · Average duration reduction · First-call-resolution reduction/improvement target · Agent/business/product feedback · LLM-as-judge or Bedrock evaluator outputs

Outcomes

• The system is running in production pilot for selected business lines and agent groups.
• Early pilot results report real-time intent detection and answer generation without latency issues.
• US Bank identified data quality, domain-specific prompts, guardrails, and automated observability as key operating lessons.
• Future plans include enterprise contact-center scaling and possible customer-facing self-service.

The pilot sits inside an organization with roughly 100-plus global contact centers, 70,000 employees, 2,000 US branches, and presence in 13 countries, but the current deployment is intentionally limited.