AI-Powered Client Services Assistant for Post-Trade Services

london_stock_exchange_group · Finance · 2025 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (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: Post-trade client services

LSEG's LCH client-services team needed faster, reliable answers to complex member queries that require consulting rulebooks, FAQs, CRM data, and internal documentation in a regulated financial-services environment.

Constraints:

• post-trade products are complex and highly technical
• answers must be accurate in a regulated environment
• information spans public and internal data sources
• enterprise access must be secured
• the team wanted faster retrieval without heavy custom model deployment

Architecture

LSEG LCH built a secured Amazon Q Business assistant behind a custom UI and SAML/IAM access, indexed rulebooks and internal sources, and used a Claude-on-Bedrock validation Lambda against golden answers before phased rollout.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  custom-web-ui["Custom web UI"]
  saml-iam-access["SAML IAM access"]
  api-gateway-and-lambda["API Gateway and Lambda"]
  amazon-q-business["Amazon Q Business"]
  indexed-data-sources["Indexed data sources"]
  golden-answer-benchmark["Golden-answer benchmark"]
  claude-validation-lambda["Claude validation Lambda"]
  testing-and-rollout-process["Testing and rollout process"]
  custom-web-ui --> api-gateway-and-lambda
  custom-web-ui --> saml-iam-access
  saml-iam-access --> amazon-q-business
  api-gateway-and-lambda --> amazon-q-business
  amazon-q-business --> indexed-data-sources
  golden-answer-benchmark --> claude-validation-lambda
  api-gateway-and-lambda --> claude-validation-lambda
  claude-validation-lambda --> testing-and-rollout-process
  testing-and-rollout-process --> custom-web-ui

Components

Data flow

Data connectors index LCH website content, FAQs, CRM, and internal repositories; a custom UI calls Amazon Q Business; Claude validation checks responses against golden answers.

Why this architecture

The architecture adds security, a custom UI, and validation because regulated client-service answers need accurate synthesis and more control than a default chat surface.

Models used: Amazon Q Business foundation model stack · Anthropic Claude v2

Scale

• The assistant returns answers within seconds and was rolled out in phases after factual-accuracy testing.
  “The Amazon Q Business application returned answers within a few seconds for each question.”article, stated

  “The CS AI assistant was rolled out in a phased approach to provide thorough, high-quality answers.”article, stated
• Amazon Q Business answers were returned within a few seconds for each question.
  “The Amazon Q Business application returned answers within a few seconds for each question.”article, stated
• Claude v2 returns per-question and total scores for answer validation.
  “Anthropic’s Claude v2 model returns a score for each question and answer, in addition to a total score, which is then provided to the application admin for review.”article, stated

Alternatives considered (roads not taken)

Deploying and managing a custom LLM stack

Rejected because: LSEG chose Amazon Q Business for enterprise search and ease of deployment without separate LLM deployment.

Using the built-in Amazon Q web experience

Rejected because: The team built a custom UI to maintain frontend control while accessing the Q Business API directly.

Assistant rollout without a golden-answer benchmark

Rejected because: LCH used golden answers to benchmark factual accuracy and reliability.

Unsecured enterprise assistant access

Rejected because: The team used SAML 2.0 IAM federation for the private assistant.

Patterns

rag-retrieval (retrieval)

The assistant retrieves from indexed LCH documents, FAQs, CRM records, and rulebooks so client-services agents can answer member questions quickly.

llm-as-judge (evaluation)

A validator compares Amazon Q Business answers with golden answers and scores the results using Claude v2.

safety-policy-enforcement (safety)

Enterprise authentication and authorization wrap the assistant so only approved internal client-services users can access the Q Business application.

canary-rollout (operations)

LCH rolled the assistant out in phases to preserve answer quality before expanding use.

Tradeoffs

Which LLM approach to use for client-services retrieval.

Chose Amazon Q Business managed RAG over custom model fine-tuning, pre-training, or heavier SageMaker deployment.

Rationale: The team selected Q Business for built-in enterprise search and ease of deployment without managing LLM infrastructure.

How to expose the assistant internally.

Chose custom UI with direct API control over the built-in Amazon Q Business web experience.

Rationale: A custom frontend gave LCH more control over the internal client-services experience and integration surface.

Gotchas (what bit them)

Client-services answers require accurate synthesis across complex documents.

Root cause: LCH member queries can require detailed service and policy documentation rather than simple FAQ lookup.

Resolution: Use managed RAG over multiple data sources and provide citations to the sources used for each answer.

Trust in a regulated assistant required more than a working chat UI.

Root cause: Factual errors in post-trade support could damage customer trust and regulatory posture.

Resolution: Create golden answers and validate generated answers with a separate Claude v2 scoring flow.

Built-in assistant UX was not enough for LCH's control requirements.

Root cause: The team wanted more control over the frontend and direct access to the Amazon Q Business API.

Resolution: Build a custom UI backed by ECS, API Gateway, Lambda, and IAM federation.

Evaluation

LCH tested factual accuracy, answer quality, structured and unstructured data handling, response time, and a validator that scores Q Business outputs against golden answers.

Metrics: high level of factual accuracy · answers delivered within a few seconds · Claude v2 returns a score for each question and answer plus a total score

Outcomes

• The assistant returned answers within seconds.
• The validation process helped LCH build trust in Amazon Q Business for post-trade support.
• LCH plans future integration with email and CRM interfaces and expansion to more LSEG use cases.

AWS co-authored post; outcome claims are qualitative, with seconds-level response time but no public volume or accuracy percentages.