Building Autonomous AI Agents on AWS with Celonis Process Intelligence

In complex automotive manufacturing environments, aligning production schedules between multiple companies operating on different systems is a persistent challenge. Celonis identified a critical bottleneck in the order-to-delivery process: coordinating appointments between internal resources and external partners required significant manual effort, was prone to delays, and relied on human intermediaries to bridge disconnected systems.

The core challenges were threefold:

• Fragmented data sources spread across internal and partner systems
• Multiple external partner systems requiring real-time coordination
• Manual intervention by human operators to manage scheduling and communication between systems

Celonis sought to build an AI agent capable of handling all coordination autonomously — retrieving order data, checking partner availability, and scheduling appointments — without human involvement in the loop.

To address these challenges, Celonis and AWS co-developed an agentic solution built on Amazon Bedrock AgentCore, a serverless environment designed for deploying and scaling AI agents and MCP servers. The agent was built using the Strands SDK, AWS's open-source framework for building and deploying agents, and connected to external systems via the open MCP (Model Context Protocol) standard.

At the heart of the solution is the Celonis MCP Server, which exposes a set of tools enabling the agent to:

• Load all order data and identify orders ready for partner scheduling
• Retrieve partner data and available resources
• Apply customizations from partners to orders
• Trigger action flows within the Celonis environment to write results back into the system

The agent also connects to third-party systems — such as partner calendar APIs — to check availability and avoid scheduling conflicts. By cross-referencing data from both the Celonis environment and external systems, the agent autonomously identifies the optimal time slot and schedules the appointment, then writes the outcome back into Celonis via an action flow.

The architecture leverages four key components of Amazon Bedrock AgentCore:

• AgentCore Runtime: A serverless environment for deploying and scaling agents and MCP servers, supporting any open-source framework and protocol. It automatically versions runtime deployments and exposes them through a secure endpoint.
• AgentCore Gateway: A centralized, secure hub for managing and exposing MCP servers and REST APIs to agents through a single unified interface — simplifying tool discovery, authentication, and orchestration at scale.
• AgentCore Identity: A centralized identity and credentials management system that assigns each agent its own scoped identity, controlling both inbound access (which users can trigger the agent) and outbound access (which tools the agent can use).
• AgentCore Observability: Full tracing and transparency of agent activity, logging every tool call, request, and response between the agent and downstream systems — enabling continuous process improvement by feeding logs back into the Celonis environment.

The joint solution between Celonis and AWS demonstrates a scalable, trustworthy blueprint for deploying autonomous AI agents in industrial environments. By grounding agents in process intelligence — using process mining to identify bottlenecks, define guardrails, and feed agents the right context — organizations can ensure that AI acts purposefully and reliably within their operations.

The observability layer provided by AgentCore enables a continuous improvement loop: agent logs are written back into Celonis, allowing teams to monitor agent behavior, validate outcomes, and refine processes over time. This closes the loop between AI action and process intelligence, ensuring that deployed agents genuinely improve operational performance.

Looking ahead, Celonis and AWS see this architecture as a foundation for expanding autonomous orchestration across additional use cases in automotive and manufacturing — wherever fragmented systems, manual coordination, and complex partner networks create friction in critical business processes.