Customer Service Transformation with AI-Based Email Automation and Chatbot Implementation

sixt · Automotive · 2025 · source Agentic: yes · 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: Global customer service automation

Sixt needed to scale customer service across many products, countries, languages, and channels without simply growing routine support workload, while keeping customer-facing AI quality high enough for production.

Constraints:

Customer inquiries span sales, documents, billing, reservations, and multiple products.
Customers ask questions in many languages.
Automation needs backend integrations such as reservation systems.
Customer-facing replies require evaluation and monitoring before exposure.
Classification accuracy below the required level creates manual reclassification and longer waits.

Architecture

Sixt built Project AIR, an AI-based replies platform using Amazon Bedrock and Anthropic Claude models. Incoming emails and chat messages are classified by process and intent, routed into automation or reply-generation flows, enriched with knowledge-base retrieval and backend-system integration, then either sent automatically when quality is sufficient or handed to human customer-service agents with summaries, extracted fields, and reply proposals.

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

flowchart LR
  customer-channels["Customer channels"] --> classification-service["Classification service"]
  classification-service --> automation-flows["Automation flows"]
  classification-service --> agent-assist-workspace["Agent assist workspace"]
  automation-flows --> knowledge-base["Knowledge base"]
  knowledge-base --> reply-generator["Reply generator"]
  reply-generator --> agent-assist-workspace
  automation-flows --> backend-systems["Backend systems"]
  automation-flows --> customer-channels
  backend-systems --> chatbot["Chatbot"]
  chatbot --> customer-channels
  agent-assist-workspace --> human-support-agents["Human support agents"]

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Customer channels`	Email, messaging, chatbot, and future voice entry points for customer requests.	`Classification service`
`Classification service`	Uses Bedrock and Claude to classify inquiry stage and subcategory.	`Customer channels`, `Automation flows`, `Agent assist workspace`
`Automation flows`	Triggers response generation, self-service routing, or process-specific automation.	`Classification service`, `Knowledge base`, `Backend systems`, `Customer channels`
`Knowledge base`	Provides process and policy information for generated answers.	`Automation flows`, `Reply generator`
`Backend systems`	Reservation and other operational systems needed for customer actions.	`Automation flows`, `Chatbot`
`Reply generator`	Creates agent-facing or customer-facing response proposals from classification and retrieved knowledge.	`Knowledge base`, `Agent assist workspace`
`Agent assist workspace`	Salesforce Service Cloud workspace showing classification, extracted information, summaries, and generated replies.	`Classification service`, `Reply generator`, `Human support agents`
`Human support agents`	Review proposed replies and handle cases not safe or possible to automate.	`Agent assist workspace`
`Chatbot`	Customer-facing conversational channel that answers questions or routes customers to self-service.	`Backend systems`, `Customer channels`

Data flow

A customer email or chatbot question enters the system. Bedrock classification identifies the process stage and subcategory. The platform extracts information such as reservation number and sentiment, retrieves policy or knowledge-base content, calls backend systems when a reservation action is needed, generates a response or self-service link, and either returns the answer automatically or hands the case to a human agent.

“it goes into the first classification process where again the bedrock models are utilized and depending on that different automation flows or reply generation flows can be triggered”transcript, stated

“then it goes um to our systems the answer is generated and a Google maps link is um shared with the customer so that he or she can see how to find it”transcript, stated

Why this architecture

Sixt split the system into classification, agent assistance, and automation because routing intent correctly determines which backend systems, knowledge, and human handoffs are needed. The design supports both internal productivity and customer-facing automation while keeping evaluation and monitoring central.

“for automation it's important because then you understand what are the kind of systems what is the kind of flow you would like to go or what's the kind of AI agent that can help best here”transcript, stated

“especially when it's customerf facing genai, you need a proper evaluation. You need to understand are the answers at a level are the replies at a level is the quality at a level”transcript, stated

Models used: Amazon Bedrock models · Anthropic Claude on Amazon Bedrock

Scale

Sixt described more than 100 countries where customers can pick up vehicles or come from.
“with um more than 100 countries where customers can pick up vehicles or where they are also coming from different countries.”transcript, stated
Sixt had more than 100 and actually several hundred AI/ML models running in production.
“we have there more than 100 um models that are running in production actually several hundreds that are running in production.”transcript, stated
Classification accuracy improved from about 70% to more than 90%.
“with a out ofthe-box M solution the accuracy was not at the level that we needed was only around 70% and that means a huge amount of manual reclassifications”transcript, stated
Classification task cost was reduced to 30% of the prior cost, a 70% reduction.
“we were also able to reduce the costs for that classification task um to 30% so meaning a 70% reduction of the costs related to this”transcript, stated
The initial AI reply generation use case moved from prototype to production in roughly five months.
“Six managed to transform their customer service um brought generative AI from ideiation into tangible business value within five months of time.”transcript, stated

Alternatives considered (roads not taken)

Use out-of-the-box classification

Rejected because: It reached only around 70% accuracy and caused manual reclassification burden.

“with a out ofthe-box M solution the accuracy was not at the level that we needed was only around 70%”transcript, stated

Fully automate every customer interaction

Rejected because: Cases that cannot be resolved automatically need handover to human agents.

“If it's not possible to resolve it completely automatically then we have there a handover process um in the possible to human agents.”transcript, stated

Build without market alignment

Rejected because: The team had to avoid spending capacity on capabilities AWS or partners would soon provide.

“where should we um focus our development capacity so that we don't develop something that is coming one month later by AWS or some other supplier”transcript, stated

Patterns

`tool-use-orchestration` (workflow)

Customer requests trigger different backend-system and automation flows based on classification.

“when you want, for instance, to change a reservation, then you have to integrate with different systems, you have to integrate, for instance, with a reservation system.”transcript, stated

`rag-retrieval` (retrieval)

The chatbot and reply flows retrieve relevant knowledge before generating answers.

“behind um the knowledge is retrieved and a link also um the answer is um sent what kind of um documents are needed it and it's sent um and back including the link.”transcript, stated

`human-in-the-loop-gate` (workflow)

Generated email replies can be reviewed by human agents unless quality is sufficient for automation.

“already a reply is um generated a reply proposal um which the human agent in that case can review.”transcript, stated

`eval-driven-development` (evaluation)

The team treats evaluation and monitoring as prerequisites for customer-facing GenAI.

“you need a proper evaluation. You need to understand are the answers at a level are the replies at a level is the quality at a level that can be exposed to a customer and you also want to monitor it.”transcript, stated

Tradeoffs

Automate confident replies while preserving human review

Chose automatic sends only for sufficiently confident/quality cases over manual review of all generated replies or full automation of all cases.

Rationale: The workflow balances speed gains with quality control for customer-facing responses.

“For cases where we are already confident and MC sufficient quality even that process is then automated and um the email is sent back um completely automatically.”transcript, stated

Prioritize use cases by business-process volume

Chose start with the most impactful customer-request categories over building every possible support flow at once.

Rationale: Knowing the share of requests per use case guided where automation effort paid off first.

“It was key to understand how many percentage um of customer requests are coming to a certain use case. So for instance changing the reservation that allowed us to prioritize the cases accordingly”transcript, stated

Gotchas (what bit them)

Classification quality can block automation progress.

Root cause: The team found classification accuracy was not high enough after migrating service tooling.

Resolution: Step back, improve classification with Bedrock/Claude, then proceed with automation use cases.

“we found out the classification accuracy was not at the point as I have mentioned before. Therefore, we had to go step one um we had to go one step back”transcript, stated

Customer-facing GenAI needs backend integration, not just text generation.

Root cause: Real support actions such as reservation changes depend on operational systems.

Resolution: Integrate GenAI services with backend systems and self-service pages.

“integration integration with different systems is um very important and um for this we also very thankful to the very good collaboration with the platform team”transcript, stated

A single ROI number does not capture GenAI value.

Root cause: Customer service AI creates financial, employee, and future-business value.

Resolution: Track value across reduced waits, consistency, operational efficiency, and employee enablement.

“the problem is not all value created by AI can be can be described with a concrete euro figure.”transcript, stated

Evaluation

Sixt evaluated classification accuracy, reclassification burden, cost per classification task, answer quality for customer exposure, waiting-time/customer-experience effects, and ongoing monitoring for customer-facing GenAI.

Metrics: Classification accuracy · Manual reclassification volume · Classification cost · Reply quality · Waiting time · Automation coverage · Customer consistency across countries

“You have to think about the evaluation strategy to ensure that the quality is at a sufficient level.”transcript, stated

Outcomes

Project AIR moved the first AI reply-generation use case into production in September 2023.
Classification accuracy improved from roughly 70% to more than 90%.
Classification cost dropped by 70%.
Messaging and chatbot services rolled out to all corporate countries by Q2 2025.
The system supports faster responses, consistency, operational efficiency, and reduced routine work for agents.

Sixt operates customer service across more than 100 countries and already had several hundred AI/ML models in production.

“in Q1 2025, we launched the messaging, we launched the chatbot, rolled it out um in Q2 to all corporate countries”transcript, stated

“Clearly, an enhanced customer experience, especially by reduced waiting times. When you have workloads automated, then the response is given much faster”transcript, stated