Agentic Orchestration in Action: How Camunda and AWS Transform Customer Support

We are in a new era of automation, and customer expectations for accurate and seamless service have never been greater. Recently, we partnered with AWS to demonstrate the value of agentic orchestration in transforming customer support agents. We recommend you read their blog post, Agentic Orchestration in Action: Transforming Customer Support with Camunda & AWS, which is an excellent summary and has particularly interesting details how AWS services such as Amazon Bedrock, OpenSearch, and Lambda combine to deliver intelligent, AI-driven customer support.

This post is designed as a companion post, which tackles the same example but explains more what happens inside Camunda’s orchestration layer—how Camunda structures, governs, and operationalizes these AWS components so that agentic behavior becomes transparent, reliable, and enterprise-ready.

Transforming customer support: A customer journey

In this example of a customer interaction, a customer sends an email asking something like: “What’s my remaining loan balance and payoff schedule?”

Traditionally, this would trigger a manual chain of work, with the request routed to an agent, who would then gather data from multiple systems and run various calculations before drafting one or more responses. The process might take days.

With Camunda and AWS, that same interaction now completes in minutes.

1. Email Intake and Triggering – The customer’s message arrives via Amazon SES, which triggers a Camunda process instance. Camunda immediately logs and contextualizes the request within a governed BPMN model, which you can observe and understand via Operate and Optimize.
2. Understanding and Intent Extraction – Using Amazon Bedrock, am AI agent is embedded via Camunda’s AI Agent connector, which helps set boundaries around which tools the agent can use and provides traceability. The embedded AI agent can interpret the customer’s message and determines what information is needed.
3. Context Retrieval and Calculation – The agent retrieves past interactions and loan data through Amazon OpenSearch, taking advantage of Camunda’s Vector database connector, and triggers an AWS Lambda function for repayment calculations.
4. Personalized Resolution – Once validated, the agent goes back to the customer and sends them a personalized answer by email via Amazon SES.

Technical deep dive: How it works

At the core of this solution is Camunda’s agentic orchestration, which embeds AI agents within BPMN workflows and coordinates every AWS interaction as a discrete, governed step.

An ad-hoc subprocess represents the agent’s “thinking loop”: the AI evaluates context, selects the appropriate AWS service from the options explicitly provided to it, executes an action, and feeds results back into the process. This enables real-time adaptation while preserving security and control.

AWS components can integrate seamlessly and powerfully within this structure:

• Amazon SES manages inbound and outbound communication events.
• Amazon Bedrock provides the LLM reasoning layer, invoked through Camunda with configurable limits on calls, latency, and retries.
• Amazon OpenSearch acts as the long-term memory layer, connected through Camunda’s vector database connector for controlled retrieval and updates.
• AWS Lambda executes the business logic for loan calculations and other transactional steps.
• Amazon S3 serves as the agent’s short-term data store, linked through Camunda’s stateful workflow engine.

Camunda ties these together through Zeebe’s event-sourced execution, ensuring every interaction is durable, replayable, and visible. Overall, the process is both agentic and accountable, combining AWS’s intelligence with the visibility and governance of Camunda’s agentic orchestration.

Governing the agent’s brain

Amazon Bedrock provides the reasoning engine behind the agent—the brain that interprets customer intent, chooses tools, and plans the next step.

Within Camunda, this intelligence is guided and governed by BPMN-based process orchestration. Camunda defines the boundaries of the agent’s cognition: which tools it may invoke, how often it may retry, and when it must escalate to a human.

Each invocation of Bedrock occurs within an explicitly modeled process step. Developers can configure conditions such as:

• Tool availability and access – Camunda determines which endpoints (whether AWS services, human tasks, or anything else) the agent can call and in what sequence.
• Retry logic and circuit breakers – Process variables control the number of permitted retries, timeout thresholds, and fallback paths when an AWS call fails or returns low confidence. This can help manage the ultimate accuracy and cost of the agent.
• Confidence-based branching – Camunda gateways can inspect metadata returned by Bedrock (e.g., confidence scores) and dynamically route execution to either continue automation or escalate to a human through a User Task.
• Governance and auditability – Every agent decision and model interaction is captured in Operate, enabling replay and traceability for compliance or optimization.

By defining these parameters at the process level, Camunda ensures that even the most capable AI model runs within governed, observable limits.

Extending the agent’s memory

Long-term memory is essential for an AI agent to have context about your business process and to learn from past experiences. For AWS, Amazon OpenSearch Service acts as the agent’s long-term memory, storing semantic embeddings that allow the system to recall relevant documents or past interactions.

Camunda extends this through its vector database connector, which integrates seamlessly with OpenSearch.

This allows process designers to configure how and when memory is updated or queried:

• Memory retrieval – Camunda orchestrates calls to OpenSearch during process execution, retrieving embeddings based on current case context.
• Memory writing – When the process completes or a new insight is generated, Camunda triggers updates to OpenSearch, appending vectorized context so that the agent “learns” from each interaction.
• Short-term context management – For transient state, Camunda coordinates with Amazon S3 or in-memory process variables within Zeebe.
• Process awareness – Because all retrieval and update operations occur within the same BPMN process, each memory action is versioned and auditable.

This integration ensures that the AWS-based memory layer is both contextually rich and fully orchestrated.

Keeping the human in the loop

Inevitably, certain cases will require escalation to a human expert. Guardrails are essential to ensure that in the right circumstances, a human can weigh in to provide guidance to the AI agent or make the final call.

In Camunda, this is an explicit part of the process model. A gateway evaluates the agent’s confidence or policy constraints; when thresholds are not met, control easily passes to a User Task for a human to evaluate.

When the human intervenes, Camunda provides full visibility:

• The entire case history, conversation thread, and the agent’s proposed next action are visible within a custom Camunda Form.
• The expert can approve, correct, or override the AI-generated response.
• Their input is then captured as part of the process data and, through Camunda’s orchestration, fed back into OpenSearch—continuously refining the agent’s knowledge base.

Not only does this provide a backstop to ensure humans can step in when necessary, but the visibility and auditability of everything the agent has done is critical for both human understanding of the agent and regulatory compliance.

Delivering powerful autonomy with reliable guardrails

Camunda and AWS together enable a governed form of autonomy. AWS provides the intelligence, memory, and compute capacity; Camunda provides the orchestration that ensures each action is purposeful, compliant, and observable. Through this combination, organizations can safely deploy intelligent agents that act within controlled, measurable parameters.

Learn more

For a closer look at how AWS services are configured within this architecture, be sure to read the AWS Builder Center post Agentic Orchestration in Action: Transforming Customer Support with Camunda and AWS.

To explore how Camunda orchestrates and governs agentic AI, visit our agentic orchestration page or check out this blog post on building an AI agent with Camunda.