Orchestrate people, AI, and systems in one governed model
Camunda brings deterministic workflow, AI agents, and human tasks together in a single unified agentic orchestration approach that never breaks.
Used for
End-to-end business process orchestration, not siloed task automation. Running in production, at scale, across the world’s most complex regulated environments.
Public sector · Benefits processing
The U.S. Department of Veterans Affairs automated pension, DIC, and burial benefits claims with Camunda and machine learning — raising the automation rate by 65% and cutting average processing time by over a third. Some claims now complete the same day they are filed. Time to production: under four months.
U.S. Department of Veterans Affairs →Financial services · Post-trade
1.8 million digital issuances, ~€20bn in value — more than any other securities digitization platform globally. Camunda orchestrates a microservices architecture covering 95%+ of German retail structured products.
Clearstream →Healthcare · Medication fulfillment
85% of complex medication orders handled by AI agents. Manual intervention dropped from 50% to 15%. Agents handle verification and logistics across a multi-step workflow — human reviewers focus on exceptions, with full context surfaced automatically.
Wellpointe →Insurance · Claims
200-year-old insurer replaced an obsolete legacy workflow tool. Claims processing time down 35%. Cost per claim down 30%. Real-time dashboards now guide adjuster prioritization dynamically.
Norfolk & Dedham →Technology · Financial operations
30× ROI on automation spend. Atlassian orchestrates financial close, revenue recognition, and reconciliation across ERP, CRM, billing, and data warehouse systems — with AI agents layered on for anomaly detection and variance analysis.
Atlassian →Telco · Network operations
Switzerland’s leading telecom provider orchestrated its full national network rollout — coordinating field teams, supply chain, compliance checks, and system updates in one process model.
Swisscom →Real enterprise processes don’t flow in a straight line. They branch on conditions, wait days for external responses, involve humans at critical moments, trigger exceptions that need compensating, and span multiple systems that weren’t designed to talk to each other. Simple automation handles the happy path. It breaks on everything else.
Most processes also contain both deterministic steps and agentic reasoning. A loan application moves deterministically through fraud checks, but hands off to an agent when the credit profile is unusual. A claims journey runs straight through for standard damage, but routes to an agent, and then a human, when fraud is suspected. The mistake is building two separate systems for these two types of work. Camunda runs them in a single model.
Workflow patterns Camunda handles natively
Parallel and sequential branching: paths that split, run concurrently, and rejoin
Event-driven continuation: the process resumes when an external signal arrives, days later
Human tasks modeled as first-class participants, not bolted on after the fact
Automatic compensation: if something fails downstream, completed steps are cleanly undone
AI agents for exceptions: complex cases route to an agent when no rule covers the situation
The foundation
BPMN (Business Process Model and Notation) is the open standard for modeling, visualizing, and executing business processes. What makes it different from a flowchart is that it runs. The diagram you design is the process that executes. No translation layer, no gap between what you intended and what the engine does.
BPMN was built for deterministic work: rules, sequences, human tasks, timers, error handling. It does that better than anything else, with two decades of production hardening behind it. But it had no concept of an AI agent: no way to scope what an agent can do, no inner orchestration governing the space between a decision and an action.
Camunda extended BPMN to make agents first-class participants in the process model. The result is the unified model — deterministic workflow, AI agents, and human tasks in one notation, one runtime, one audit trail.
Standard BPMN
Camunda’s agentic extension
A loan application waits for a credit bureau. An insurance claim pauses for additional documentation. An onboarding process spans weeks across multiple teams. Most agent frameworks are built for request-response patterns. Enterprise processes are long-running by nature.
This requires an orchestration engine built for it from the ground up — event-streamed, horizontally scalable, capable of running millions of concurrent process instances without degradation.
Explore ZeebeProcess state is persisted at every step. When a reply arrives days later — from a bureau, a customer, or an external system — the process resumes exactly where it left off. No timeout. No lost context. No one restarting manually.
Waiting for a signature? The process fires a reminder after a week, another after two, and expires the case automatically after a month — without custom polling logic or a cron job. BPMN handles this natively.
When a step fails — an API timeout, a gateway error, a service outage — the orchestration retries automatically. If retry fails, it compensates by undoing completed steps cleanly. Nothing is left half-done. No case is lost.
From discovery to continuous improvement — one platform, the same model throughout.
Discover
ProcessOS analyzes existing systems, operational data, and documentation to surface how processes actually work — not how they were designed to work. It identifies where automation stops, where exceptions pile up, and which processes have the highest return on AI investment before a line of BPMN is drawn.
Tools in this phase
ProcessOS
AI process intelligence
Discovers how processes actually run, identifies automation opportunities, and generates improvement proposals from production data.
Learn about ProcessOS →What we design is what we implement, is what we monitor, is what we optimize. You have that single source of truth through the entire process.
Dan Avidan
Lead Platform Architect, HSBC
A process is the coordination structure that connects humans, AI agents, and systems into a continuous end-to-end flow. It defines the sequence of decisions and actions — whether each step is deterministic (fixed in advance by rules) or agentic (chosen at runtime by AI reasoning). The process owns the handoffs, maintains state, enforces governance, and provides the audit trail for everything that happens.
BPMN (Business Process Model and Notation) is an ISO-standard visual language for modeling and executing processes. Because it is executable — not just documentary — it becomes the coordination layer for AI agents: enforcing governance gates, persisting state across long-running work, handling timers and failures, and capturing a complete audit trail for every decision. Camunda extends BPMN with native constructs for agentic behavior while keeping full backward compatibility with the standard.
The unified model combines deterministic orchestration and agentic reasoning in a single BPMN process. Standard cases flow through rule-based steps. Complex cases route to an AI agent. Human approval is injected where required. All of this runs in one runtime, with one audit trail and one place to look when something goes wrong — rather than two separate systems connected by custom glue code.
The Camunda engine persists process state at every step, so a process can wait hours, days, or months for an external reply and resume exactly where it left off. Timers fire automatically — sending reminders, escalating, or expiring cases — without custom polling logic. When steps fail, the engine retries automatically and compensates by undoing completed steps if retry fails, so nothing is left in an inconsistent state.
Yes. Running process instances continue on their current version when you deploy a new one. Instance migration moves existing cases to the new version when needed, preserving complete execution state. You can also run A/B tests across process versions, canary-release to a percentage of traffic, and roll back immediately if quality degrades.
Talk to the team about your processes — where AI can go further and how to get it into production.