How Hapag-Lloyd makes use of Amazon Bedrock to remodel buyer suggestions into actionable insights

Hapag-Lloyd stands as one of many world’s main liner delivery firms, working a contemporary fleet of 313 container ships with a complete transport capability of two.5 million TEU (Twenty-foot Equal Unit—a regular unit of measurement for cargo capability in container delivery). The corporate maintains a container capability of three.7 million TEU, which incorporates one of many trade’s largest and most fashionable fleets of reefer containers. With roughly 14,000 workers within the Liner Transport Phase and greater than 400 workplaces unfold throughout 140 nations, Hapag-Lloyd maintains a strong international presence. By way of 133 liner providers worldwide, we facilitate dependable connections between greater than 600 ports throughout the continents.

The corporate’s Digital Buyer Expertise and Engineering group, distributed between Hamburg and Gdańsk, drives digital innovation by growing and sustaining customer-facing internet and cell merchandise.

Over the previous years, the Digital Buyer Expertise and Engineering group has advanced from a delivery-focused channel into a real digital product driver, with sturdy buyer focus, engineering excellence, and measurable enterprise impression. We take end-to-end possession of our digital merchandise, combining customer-centric innovation with engineering craft to immediately help development and enterprise outcomes. Constructing on a contemporary, independently owned tech stack and a excessive degree of engineering maturity, we’re dedicated to staying on the forefront of know-how. Now, we’re taking the subsequent step by transferring towards turning into AI-native, investing closely in synthetic intelligence as a core functionality. This journey is about amplifying highly effective engineering with AI to construct smarter merchandise, quicker innovation, and higher buyer worth.

Understanding consumer impression.

Thus far, our buyer suggestions evaluation course of had largely been guide and reactive. Particularly forward of assessment ceremonies, manually analyzing buyer suggestions may take hours, typically days, when a whole lot of rankings and feedback wanted to be reviewed. Each two weeks, Product Managers exported buyer suggestions information as CSV information, learn by means of giant volumes of feedback, and manually categorized sentiment and themes. Though this work was priceless and deeply related to product selections, it was additionally repetitive, time-consuming, and tough to scale, limiting flexibility at any time when quicker or deeper insights have been wanted.

With our generative AI answer, we basically modified this strategy. As an alternative of manually aggregating and deciphering suggestions, we now automate all the workflow: amassing buyer feedback, extracting sentiment, figuring out themes, and surfacing actionable insights. Product Managers and groups can focus much less on operational evaluation and extra on technique, innovation, and creating distinctive consumer experiences.

On this publish, we stroll you thru our generative AI–powered suggestions evaluation answer constructed utilizing Amazon Bedrock, Elasticsearch, and open-source frameworks like LangChain and LangGraph. Amazon Bedrock is a completely managed service that provides a selection of high-performing basis fashions from main AI firms comparable to AI21 Labs, Anthropic, Cohere, DeepSeek, Luma, Meta, Mistral AI and Amazon by means of a single API, together with a broad set of capabilities it’s essential construct generative AI functions with safety, privateness, and accountable AI. With this answer, you may routinely ingest buyer feedback, generate wealthy summaries, and ship focused insights. This enables our product groups to make quicker, smarter selections and drive steady enchancment.

We stroll you thru the structure and implementation of this answer, demonstrating how utilizing generative AI foundations, comparable to orchestration, information administration, safety, and privateness, allowed us to quickly construct a scalable, production-ready suggestions processing pipeline.

Answer overview

The answer is constructed on AWS structure designed to handle these challenges by means of scalability, maintainability, and safety. It’s deployed through the use of AWS CloudFormation.

1. Steady & Quarterly Suggestions Assortment
   • Our internet and cell functions serve a whole lot of 1000’s of shoppers every month.
   • Customers can go away a ranking plus textual content feedback, serving to us perceive consumer expertise and enhance providers.
2. Day by day Suggestions Ingestion & Processing
   • A AWS Lambda perform runs as soon as per day to fetch the brand new suggestions entries.
   • We use Amazon Bedrock to classify sentiment (optimistic, unfavorable, combined, or impartial) for every open remark, streamlining downstream evaluation.
   • Processed data are listed in Amazon OpenSearch Service, serving each as our full-text search engine and vector database.
3. Interactive suggestions exploration through OpenSearch Service
   • Stakeholders can entry real-time suggestions insights by means of OpenSearch Dashboards, giving them a chicken’s-eye view of consumer sentiment, rankings, and tendencies over time.
   • Beginning with high-level visualizations, comparable to sentiment distribution, ranking scores, and suggestions quantity, customers can drill down into particular functions, options, and even particular person feedback.
   • Dashboards help filtering by time interval, remark sentiment, product model, and extra, enabling focused root trigger evaluation.
   • For instance, a Product Supervisor can visualize how sentiment round a current app replace modified week over week, or zoom into unfavorable feedback mentioning a selected function.
4. AI-Powered Inner Chatbot
   • Our stakeholder-facing chatbot queries the OpenSearch index as its data base.
   • We use Bedrock Guardrails, to implement security and reliability and ensure responses align with our model and compliance requirements.
   • Product managers and help groups can ask natural-language questions, for instance, “What ache factors do clients point out most frequently?” and obtain prompt, context-rich solutions.
5. Biweekly Insights Report
   • Each two weeks, a second Lambda perform aggregates and analyzes the most recent suggestions tendencies.
   • It generates a concise report with key metrics, highlights, and sentiment breakdowns.
   • The report is routinely delivered to our Product Managers and Product House owners, feeding immediately into dash planning and roadmap discussions.

Generative AI Orchestration

Orchestration is a core basis of our answer, as a result of generative AI workflows usually contain a number of steps that should be coordinated. In our pipeline, information ingestion and processing steps, comparable to sentiment evaluation, embedding technology, and indexing, are orchestrated utilizing LangChain, which offers modular, reusable elements for calling fashions, reworking information, and integrating with exterior programs like Amazon OpenSearch Service. For our inner chatbot, we depend on LangGraph to implement a multi-agent structure. Every assistant is outlined declaratively in LangGraph, encapsulating its personal logic and instruments. This design makes assistants versatile and composable: as a substitute of inflexible step-by-step flows, we use an agent-based strategy the place an LLM selects the proper instruments and actions dynamically to reply consumer queries. This offers product managers and help groups a pure, interactive option to discover suggestions and associated operational insights.

Integration with Amazon Bedrock fashions is simple utilizing LangChain’s native help. For instance, our AI-powered inner chatbot makes use of the Claude Sonnet 4.6 mannequin through Amazon Bedrock. We selected Claude Sonnet 4.6 as a result of it affords frontier efficiency throughout coding and agentic workflows. The mannequin excels in multi-turn conversational exchanges and agentic workflows, making it preferrred for our inner chatbot that requires dependable efficiency throughout single and multi-turn interactions with stakeholders. With its exact workflow administration capabilities and skill to serve in each lead agent and subagent roles, Claude Sonnet 4.6 delivers the constant conversational high quality our product managers and help groups want when exploring suggestions insights at scale. Moreover, we leverage geographic Cross-Area Inference Service (CRIS) endpoint to seamlessly handle unplanned visitors bursts by distributing compute throughout a number of EU AWS Areas. This cross-region functionality ensures our suggestions processing pipeline stays resilient throughout peak utilization durations whereas sustaining constant efficiency for our international stakeholder base. The mannequin is configured with guardrails utilized immediately by means of LangChain configuration:

from langchain_aws import ChatBedrockConverse

def get_chatbot_model():
    return ChatBedrockConverse(
        consumer=session.consumer("bedrock-runtime", region_name="eu-central-1", config=config),
        mannequin="eu.anthropic.claude-sonnet-4-6",         guardrail_config={
            "guardrailIdentifier": settings.GUARDRAIL_ID,
            "guardrailVersion": "DRAFT",
            "hint": "enabled"
        }
    )

Knowledge Administration

An AWS Lambda perform runs as soon as per day to fetch the brand new suggestions entries from the suggestions repository into Amazon S3, after which the info is categorized with semantic detection by means of Amazon Bedrock. The information is then listed in Amazon OpenSearch Service, serving each as our full-text search engine and vector database.

Accountable AI

To responsibly use the answer, we implement safeguards utilizing Amazon Bedrock Guardrails. This enables us to connect Amazon Bedrock Guardrails to an AI interplay and implement content material moderation insurance policies and ensure responses align with our model and compliance requirements.

Utilizing AWS CloudFormation, we outline guardrail insurance policies as infrastructure-as-code, offering examples of configurations to assist block dangerous content material.

Guardrails as Code: CloudFormation

ChatbotGuardrail:
  Sort: AWS::Bedrock::Guardrail
  Properties:
    Identify: guardrail
    Description: Primary guardrail to dam violence and dangerous content material.
    BlockedInputMessaging: "Enter blocked by security coverage."
    BlockedOutputsMessaging: "Response blocked by security coverage."
    WordPolicyConfig:
        ManagedWordListsConfig:
          - Sort: PROFANITY
    ContentPolicyConfig:
        FiltersConfig:
            - Sort: HATE
              InputStrength: HIGH
              OutputStrength: HIGH
              OutputAction: BLOCK
            - Sort: INSULTS
              InputStrength: HIGH
              OutputStrength: HIGH
              OutputAction: BLOCK
            - Sort: SEXUAL
              InputStrength: HIGH
              OutputStrength: HIGH
              OutputAction: BLOCK
            - Sort: VIOLENCE
              InputStrength: HIGH
              OutputStrength: HIGH
              OutputAction: BLOCK
            - Sort: MISCONDUCT
              InputStrength: HIGH
              OutputStrength: HIGH
              OutputAction: BLOCK
            - Sort: PROMPT_ATTACK
              InputStrength: NONE
              OutputStrength: NONE
              OutputAction: BLOCK

Programmatic Enter Validation

We additionally use Amazon Bedrock Guardrails independently to validate uncooked consumer enter earlier than passing it to the LLM, serving to forestall immediate injection and different misuse:

def validate_question_with_guardrail(query: str, user_data: UserData) -> bool:
    consumer = boto3.consumer('bedrock-runtime')
    response = consumer.apply_guardrail(
        guardrailIdentifier=settings.GUARDRAIL_ID,
        guardrailVersion='DRAFT',
        supply="INPUT",
        content material=[{'text': {'text': question}}]
    )

    if response.get("motion") == "GUARDRAIL_INTERVENED":
        print(json.dumps(response, indent=4))
        print(
            f"Immediate was blocked. user_id=[{user_data.user_id}] query=[{question}]"
        )
        return False
    return True

With this setup, we now have created a safer, scalable, and explainable pipeline that places Generative AI to work, responsibly and successfully, throughout our product suggestions lifecycle.

Monitoring

We monitor the components of the appliance utilizing Amazon CloudWatch, which collects uncooked information and processes it into readable, close to real-time metrics. We enabled mannequin invocation logging to gather invocation logs, mannequin enter information, and mannequin output information for the invocations, enabling assortment of full request information, response information, and metadata related to the calls. Amazon Bedrock additionally integrates with AWS CloudTrail, which captures API requires Amazon Bedrock as occasions. This generates insights that you need to use to optimize the functions additional like bettering response latency or lowering prices.

Subsequent Steps

The answer processes over 15,000 suggestions objects per 30 days with an accuracy of 95% for sentiment classification on a labeled check dataset. As an alternative of spending hours reviewing uncooked suggestions, groups can now obtain clear, structured summaries in seconds that spotlight a very powerful matters and recurring indicators. This helps them transfer from info to motion a lot quicker, making selections inside days quite than weeks. Over time, the experiences helped us perceive not solely when sentiment improved, but in addition why it didn’t. By constantly monitoring buyer suggestions, we are able to react rapidly to early indicators, modify selections, and proper course when wanted. In a number of areas, actions taken based mostly on these insights have already resulted in additional optimistic feedback and a noticeable discount in unfavorable suggestions. A robust instance is the “Preview” performance in Transport Directions. This function was prioritized immediately in response to a excessive quantity of unfavorable consumer suggestions highlighting the shortage of a preview functionality. After its launch, AI-driven experiences allowed us to trace consumer reactions intimately. Suggestions associated particularly to this function confirmed that the beforehand frequent requests for preview functionality have been successfully resolved, demonstrating that the core consumer want had been efficiently addressed. On the similar time, broader suggestions continued to floor different areas for enchancment throughout the appliance. AI insights additionally guided future function planning and prioritization. Based mostly on consumer feedback, we created new OpenSearch Service-based dashboards that assist groups rapidly confirm and analyze points reported by customers. One other instance is the power to add cargo information through Excel information, a repeatedly requested function highlighted by AI suggestions. This performance is now absolutely out there and is predicted to considerably cut back guide effort, notably for big shipments. Throughout assessment periods, the stakeholders can now see prime optimistic and unfavorable feedback in actual time, alongside AI-generated suggestions, creating a much more knowledgeable and productive dialogue.

This suggestions evaluation answer is one instance of how we’re making use of generative AI throughout our processes, and it marks the start, not the top, of our AI-native journey. Underneath our AI-Native Umbrella Program, which serves as a single supply of reality for AI adoption, our subsequent focus is to determine a shared, strong AI basis with Amazon Bedrock. By offering standardized infrastructure, safety, and guardrails, we purpose to allow each function within the division, engineering, product and supply (PM, PO, SM), UX/design, and operations/help, to create their very own AI “areas” safely and independently whereas accessing the very best in-class basis fashions. This setup is designed to decrease the barrier to experimentation, streamline discovery, and encourage hands-on exploration of generative AI use circumstances in day-to-day work. In doing so, we assist groups transfer quicker from concepts to impression, whereas sustaining consistency, duty, and scalability throughout the AI initiatives.

If you wish to scale your generative AI functions, you will get began by studying this Architect a mature generative AI basis on AWS that dives deeper on the varied foundational elements that assist speed up the end-to-end generative AI software lifecycle.

Concerning the authors