AI-Powered Customer Feedback Analysis System for Container Shipping

hapag-lloyd · Other · 2026 · source Agentic: yes · promotional bias: med · 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: Container-shipping product feedback analytics

Hapag-Lloyd needed to replace manual, reactive review of exported customer-feedback CSVs with a scalable system that identifies sentiment, themes, and actionable product insights quickly.

Constraints:

• Product managers had to analyze hundreds of comments before review ceremonies
• Feedback must be searchable by sentiment, product version, time period, feature, and individual comments
• The global stakeholder base needs secure, resilient access under EU regional constraints

Architecture

Hapag-Lloyd's feedback system is a daily AWS pipeline: Lambda ingests feedback into S3, Bedrock classifies sentiment and performs semantic detection, OpenSearch indexes records as both full-text search and vector database, dashboards support stakeholder drill-down, and a Claude Sonnet 4.6 LangGraph chatbot queries the OpenSearch knowledge base under Bedrock Guardrails. A second biweekly Lambda aggregates trend reports for product ceremonies.

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

flowchart LR
  feedback-collection-layer["feedback-collection-layer"]
  daily-lambda-ingestion["daily-lambda-ingestion"]
  s3-feedback-store["s3-feedback-store"]
  bedrock-sentiment-classifier["bedrock-sentiment-classifier"]
  opensearch-feedback-index["opensearch-feedback-index"]
  opensearch-dashboards["opensearch-dashboards"]
  langgraph-chatbot["langgraph-chatbot"]
  claude-sonnet-bedrock-model["claude-sonnet-bedrock-model"]
  bedrock-guardrails["bedrock-guardrails"]
  cross-region-inference["cross-region-inference"]
  biweekly-report-lambda["biweekly-report-lambda"]
  ses-report-delivery["ses-report-delivery"]
  monitoring-and-audit["monitoring-and-audit"]
  feedback-collection-layer --> daily-lambda-ingestion
  daily-lambda-ingestion --> s3-feedback-store
  daily-lambda-ingestion --> bedrock-sentiment-classifier
  daily-lambda-ingestion --> opensearch-feedback-index
  s3-feedback-store --> bedrock-sentiment-classifier
  bedrock-sentiment-classifier --> opensearch-feedback-index
  opensearch-feedback-index --> opensearch-dashboards
  opensearch-feedback-index --> langgraph-chatbot
  langgraph-chatbot --> claude-sonnet-bedrock-model
  langgraph-chatbot --> bedrock-guardrails
  cross-region-inference --> claude-sonnet-bedrock-model
  biweekly-report-lambda --> ses-report-delivery

Components

Data flow

Customers leave ratings and comments in web/mobile products. Once per day, Lambda fetches new feedback, Bedrock classifies sentiment and semantics, records are indexed in OpenSearch, stakeholders explore dashboards or ask a guarded Claude Sonnet chatbot, and every two weeks another Lambda creates a trend report that is emailed through SES for planning and roadmap discussions.

Why this architecture

Hapag-Lloyd chose a batch-oriented AWS architecture because feedback analysis needed consistent, scalable insight rather than millisecond latency. OpenSearch supports both dashboard filtering and chatbot retrieval, LangChain/LangGraph coordinate multi-step generative workflows, guardrails and monitoring make the internal assistant governable, and CRIS protects stakeholder exploration during bursts.

Models used: Claude Sonnet 4.6 · Amazon Bedrock foundation models

Scale

• Hapag-Lloyd operates 313 container ships with 2.5 million TEU transport capacity.
  “operating a modern fleet of 313 container ships with a total transport capacity of 2.5 million TEU”article, stated
• The company has about 14,000 employees and more than 400 offices across 140 countries.
  “With approximately 14,000 employees in the Liner Shipping Segment and more than 400 offices spread across 140 countries”article, stated
• The system processes more than 15,000 feedback items per month.
  “The solution processes over 15,000 feedback items per month with an accuracy of 95% for sentiment classification on a labeled test dataset.”article, stated
• Sentiment classification reached 95% accuracy on a labeled test dataset.
  “with an accuracy of 95% for sentiment classification on a labeled test dataset.”article, stated
• Manual feedback analysis previously took hours or days; summaries now arrive in seconds and decisions in days rather than weeks.
  “Instead of spending hours reviewing raw feedback, teams can now receive clear, structured summaries in seconds”article, stated

  “This helps them move from information to action much faster, making decisions within days rather than weeks.”article, stated
• Biweekly reports are generated every two weeks.
  “Every two weeks, a second Lambda function aggregates and analyzes the latest feedback trends.”article, stated

Alternatives considered (roads not taken)

Manual CSV review before product ceremonies

Rejected because: Manual review took hours or days and could not scale when faster or deeper insight was needed.

Dashboards only

Rejected because: The system adds a natural-language chatbot so product managers and support teams can ask context-rich questions beyond fixed visualizations.

Rigid step-by-step chatbot workflows

Rejected because: LangGraph lets an LLM dynamically select tools and actions for feedback exploration.

Uncoordinated team-by-team AI spaces

Rejected because: The AI-Native Umbrella Program standardizes infrastructure, security, and guardrails while still allowing independent team exploration.

Patterns

batch-offline-processing (feedback-ingestion)

A daily Lambda job ingests new feedback, classifies sentiment, embeds records, and indexes them before dashboards and reports consume them.

rag-retrieval (feedback-chatbot)

The internal chatbot queries the OpenSearch index as a knowledge base to answer stakeholder questions about feedback.

multi-agent-orchestration (stakeholder-assistant)

LangGraph defines assistants with their own logic and tools, letting an LLM select actions dynamically for user feedback queries.

guardrails-io-filtering (responsible-ai)

The chatbot applies Bedrock guardrails through LangChain configuration to keep feedback exploration responsible and secure.

observability-tracing (ai-operations)

CloudWatch, Bedrock invocation logging, and CloudTrail record metrics, model inputs/outputs, metadata, and API events for optimization and audit.

data-quality-loop (product-feedback-action)

AI reports identify negative feedback, guide product changes, and then track whether sentiment improved after release.

Tradeoffs

How to expose feedback insights to teams.

Chose dashboards plus a natural-language chatbot and biweekly reports over raw CSV exports and manual review ceremonies.

Rationale: Different stakeholders need both high-level trends and drill-down access to features, comments, sentiment, and root causes.

How to scale AI adoption internally.

Chose standardized infrastructure, security, and guardrails under an AI-Native Umbrella Program over uncoordinated team-by-team AI spaces.

Rationale: The company wants lower experimentation barriers while maintaining consistency, responsibility, and scalability.

Gotchas (what bit them)

Manual review was too reactive and slow for product decisions.

Root cause: Product managers exported CSVs every two weeks and manually categorized hundreds of ratings/comments before review ceremonies.

Resolution: The automated pipeline collects, classifies, indexes, summarizes, and reports feedback continuously.

Peak usage could disrupt stakeholder feedback exploration.

Root cause: Unplanned traffic bursts can strain regional model capacity for a global stakeholder base.

Resolution: Hapag-Lloyd uses geographic Cross-Region Inference Service endpoints across EU AWS Regions.

AI recommendations need traceability into operational behavior.

Root cause: Teams need to optimize latency/cost and understand request/response metadata, not just read final summaries.

Resolution: The solution logs invocations, input/output data, metadata, and CloudTrail API events.

Evaluation

Hapag-Lloyd evaluates sentiment classification on a labeled test set, monitors app/model operations, tracks latency/cost signals, and compares user feedback before and after product decisions.

Metrics: monthly feedback volume · sentiment classification accuracy · time from insight to action · positive/negative comment trends · latency · cost

Outcomes

• The system processes more than 15,000 feedback items per month.
• Sentiment classification reaches 95% accuracy on a labeled test dataset.
• Teams receive structured summaries in seconds and move from insight to action within days rather than weeks.
• The Shipping Instructions Preview feature addressed a frequent negative-feedback request.
• AI recommendations surfaced the need for Excel cargo-data upload.

Hapag-Lloyd operates 313 container ships, 2.5 million TEU transport capacity, 400-plus offices, and connections across more than 600 ports.