Agentic Orchestration in Action: Transforming Customer Support with Camunda & AWS

In today’s competitive landscape, customer expectations are higher than ever—fast responses, accurate information, and seamless interactions are no longer optional. Automating customer support cases using AI-driven processes can significantly boost customer satisfaction and operational efficiency. In this post, we'll demonstrate how Camunda’s agentic orchestration, combined with AWS’s scalable AI and cloud services, can modernize customer support workflows, using loan support as a detailed example.

To understand the impact, imagine a customer sends an email asking: “What's my remaining loan balance and payoff schedule?” Traditionally, addressing this request could take days, involving manual context gathering, calculations, and multiple emails. With Camunda and AWS, the scenario transforms dramatically:

1. Immediate Email Intake: The customer’s email arrives via Amazon SES, automatically triggering a Camunda workflow instance.

2. Contextual Understanding: An AI agent, powered by Amazon Bedrock, processes the email to understand the customer's intent and determines the next best steps.

3. Dynamic Decision Making: The agent autonomously retrieves past conversation context using semantic search with Amazon OpenSearch and determines additional necessary actions—like calculating loan repayment schedules via AWS Lambda.

4. Personalized Resolution: The AI agent responds via outbound email through Amazon SES, delivering a clear, detailed answer to the customer.

This entire process takes minutes—not days.

Camunda's agentic orchestration enables AI agents embedded within BPMN workflows to dynamically choose actions in response to real-time information. Leveraging an ad-hoc subprocess, the AI agent evaluates context and iteratively selects from predefined tools to fulfill its goal. Key AWS services integrate seamlessly into this workflow:

● Amazon SES (Simple Email Service) handles inbound and outbound email communication.

● Amazon Bedrock provides the Large Language Models (LLMs) for reasoning and decision-making.

● Amazon OpenSearch Service serves as the agent's long-term memory.

● AWS Lambda hosts critical business logic for precise calculations.

● Amazon S3 acts as the short-term memory repository for in-process data.

Choosing the right AI model is like picking the right brain for the job. Amazon Bedrock offers a diverse landscape of cutting-edge LLMs from leading AI companies like Anthropic, Cohere, Meta, and Amazon's own Titan family. For agentic behavior within Camunda, several factors are critical:

● Instruction Following and Tool Use: This is paramount. The model must be adept at understanding complex instructions and deciding which "tool" (an API call, a database query) to use. Models like Anthropic's Claude 3 family are particularly strong in this area.

● Accuracy and Reliability: The agent's decisions must be correct. For financial calculations, a model with a low error rate, like Claude 3 Opus, is essential.

● Latency: In customer-facing scenarios, speed matters. A faster model like Claude 3 Haiku is ideal for initial triage and simple Q&A.

● Cost: A high-volume process will benefit from a cost-effective model, balancing performance with operational cost.

● Context Window: For agents that need to analyze long conversation histories, a model with a large context window (like the 200K tokens offered by Claude) is a significant advantage.

For an AI agent to be truly effective, it needs deep, contextual understanding of your business. This is where a vector database, powered by Amazon OpenSearch Service, becomes the agent's long-term memory.

Vector databases create a "map of meaning" for your unstructured data (emails, support tickets, internal documents).

1. Embeddings: An embedding is a numerical representation of text that captures its semantic meaning. The phrases "how much do I owe?" and "what's my remaining balance?" will have very similar numerical representations.

2. Chunking Strategies: Documents are broken down into smaller, meaningful "chunks." The right chunking strategy is key to preserving the semantic integrity of the information.

3. Indexes: These vector embeddings are stored in a specialized vector index within Amazon OpenSearch, which is highly optimized for finding the most semantically similar information to a given query.

Camunda simplifies this process using its vector database connector, which integrates with libraries like Langchain4j. With just a few lines of configuration in your BPMN process, you can connect to OpenSearch, specify an embedding model, define a chunking strategy (e.g., using a recursive text splitter with configurable maxChunkSize and chunkOverlap), and execute semantic searches to retrieve the context your agent needs.

With great power comes the need for great governance. An AI agent operating autonomously must have guardrails. Camunda’s BPMN-driven orchestration ensures this is not a "black box" process, which is crucial for regulated industries like financial services. Every action, decision, and communication made by the AI agent is traceable and auditable through Camunda’s monitoring tools—Operate and Optimize.

But what happens when the agent gets stuck? This is where the human-in-the-loop capability becomes a powerful safety net.

An AI agent should escalate to a human expert in several key scenarios:

● Low Confidence: The LLM indicates it's unsure about the customer's intent or the correct next step.

● High-Stakes Decisions: The request involves a large financial transaction or requires a level of approval the agent isn't authorized for.

● Customer Sentiment: The customer is expressing extreme frustration or anger that requires empathetic human intervention.

● Ambiguity or Novelty: The request is unusual and doesn't fit known patterns, presenting a learning opportunity.

This seamless escalation is a native part of the Camunda stack.

1. Modeling the Escalation: In the BPMN model, a gateway can check the agent's confidence score or a specific flag. If it falls below a threshold, the process is routed to a User Task.

2. Assigning the Task: The User Task appears in a human expert’s queue in Camunda Tasklist.

3. Providing Full Context: The expert isn't starting from scratch. Using Camunda Forms, a custom user interface presents all the context the agent has gathered: the full email thread, retrieved documents, and the agent's suggested (but un-executed) next action.

4. Empowering the Expert: The human can then validate the agent's suggestion, override it, or perform a completely different action. Their decision is captured in the form and fed back into the process.

5. Closing the Loop: The process continues, now armed with the human's expert decision. Crucially, this interaction doesn't just solve one problem. The human's input can be fed back into the Amazon OpenSearch knowledge base, training the AI to handle similar situations correctly in the future. This creates a powerful, continuous improvement cycle where human expertise makes the automation smarter.

Integrating Camunda’s composable, agentic orchestration with AWS’s robust services delivers a powerful and measurable return on investment. Early results from our beta program highlight a significant transformation in customer support operations.

● 85% Reduction in Average Response Time: Simple inquiries that once took hours or days are now resolved in minutes, drastically improving the customer experience.

● 40% Increase in Operational Efficiency: Automating routine tasks has freed up customer service specialists, allowing them to handle 40% more complex, high-value cases that require a human touch.

● 95% First-Contact Resolution for Automated Queries: The AI agent successfully and accurately resolves the vast majority of inquiries it handles without needing human intervention.

● Continuous Improvement: The system learns and improves over time. Human expert interventions update the knowledge base, progressively enhancing the AI's capabilities and accuracy.

● Enhanced Customer Satisfaction: Customers receive timely, accurate responses, dramatically improving their experience and fostering loyalty.

● Empowered Support Teams: Automation significantly reduces manual workload, allowing specialists to focus on the most challenging and rewarding aspects of their roles.

● Flexible and Composable: This isn't a rigid, one-size-fits-all solution. Organizations can swap, scale, or add AWS components easily, ensuring future-proof, adaptable processes that evolve with your business.

By leveraging the combined strengths of Camunda and AWS, organizations can rapidly deploy intelligent, secure, and scalable customer support processes. Agentic orchestration brings adaptability and responsiveness to workflows, while robust AWS services provide the foundational AI and data capabilities needed for modern customer service operations.

Together, Camunda and AWS empower you to revolutionize customer support—ensuring superior customer experiences, operational excellence, and seamless governance.

Explore how your organization can quickly implement intelligent customer support automation by visiting Camunda’s Agentic Orchestration documentation and exploring available AI agent blueprints.

Mathieu Stennier - Senior Engineering Manager - AI at Camunda