Camunda 8.8 introduces agentic orchestration capabilities that fundamentally change how enterprises can combine process automation with AI. For companies already running SAP S/4HANA, this release opens up new possibilities — but only if you understand where AI agents add value and where deterministic BPMN must remain in control.
What Camunda 8.8 Brings
┌──────────────────────────────────────────────────────────────┐
│ CAMUNDA 8.8 — NEW CAPABILITIES │
│ │
│ ┌──────────────────┐ ┌──────────────────┐ │
│ │ AI Agent │ │ Ad-Hoc │ │
│ │ Connector │ │ Subprocess │ │
│ │ │ │ (Tool Belt) │ │
│ │ • Claude │ │ │ │
│ │ • AWS Bedrock │ │ Agent picks │ │
│ │ • Azure OpenAI │ │ tools at runtime │ │
│ └──────────────────┘ └──────────────────┘ │
│ │
│ ┌──────────────────┐ ┌──────────────────┐ │
│ │ MCP Connector │ │ Vector DB │ │
│ │ │ │ Connector │ │
│ │ Standardized │ │ │ │
│ │ access to any │ │ Short-term & │ │
│ │ enterprise system│ │ long-term memory │ │
│ └──────────────────┘ └──────────────────┘ │
│ │
│ ┌──────────────────┐ │
│ │ A2A Connector │ All capabilities governed by BPMN. │
│ │ │ Deterministic backbone. │
│ │ Agent-to-Agent │ Intelligent edges. │
│ │ coordination │ │
│ └──────────────────┘ │
└──────────────────────────────────────────────────────────────┘
AI Agent Connector
Camunda 8.8 includes a native AI Agent Connector that integrates directly into BPMN workflows. The connector supports Anthropic Claude, AWS Bedrock, and Azure OpenAI as backend providers.
Unlike standalone AI agent frameworks, Camunda's approach embeds agents inside a governed process. The BPMN model defines when an agent is invoked, what tools it can access, and what happens if it fails.
Ad-Hoc Subprocess as Tool Belt
The ad-hoc subprocess gives AI agents a "tool belt" — a collection of activities the agent can call dynamically based on context. The agent decides which tools to use and in what order, while BPMN defines the boundary of what's available.
This is the critical governance layer. The agent cannot access tools outside the subprocess. If the agent behaves unexpectedly, the process catches it via error boundaries and routes to human review.
MCP (Model Context Protocol) Connector
MCP gives AI agents standardized access to enterprise data. Instead of building custom integrations for every system, MCP provides a protocol that any AI model can use to read from and write to enterprise systems — including SAP.
Short-Term and Long-Term Memory
The Vector DB connector gives agents persistent memory across process instances. Short-term memory maintains context within a conversation or task. Long-term memory stores patterns and learnings that improve agent performance over time.
A2A (Agent-to-Agent) Connector
Multi-agent architectures are now natively supported. Agents can delegate tasks to other agents, coordinate work, and aggregate results — all within a BPMN-governed process.
What This Means for SAP Customers
1. Intelligent Action Determination
Today, DMN decision tables determine which downstream SAP documents need updating when a source document changes. This works well for known, static mappings.
With Camunda 8.8, an AI agent can augment or replace the DMN table by reasoning about which fields need updating based on historical patterns, business context, and cross-document dependencies. The agent handles the edge cases that a static table cannot.
Example: A freight booking change triggers updates to linked Sales Orders. The DMN table maps known fields. The AI agent catches an unusual combination of field changes that the table doesn't cover and flags it for review — or handles it autonomously if the pattern has been seen before.
2. Intelligent Exception Handling
The three-tier error handling pattern (automatic retry → team notification → human review) works well for transient errors. But many SAP errors require judgment — understanding the business context, not just the error code.
An AI agent embedded in the user review task can analyze the error, cross-reference it with SAP documentation, check historical resolutions for similar errors, and recommend whether to retry, skip, or escalate.
Example: An SAP validation error occurs because a material's base unit of measure changed since the process started. The AI agent detects this, reads the current material master, adjusts the payload, and retries — instead of waiting for a human to diagnose the issue.
3. Payload Validation Before SAP Writes
Before writing to SAP via OData, an AI agent can validate the update payload against business rules that are too complex for DMN. The agent catches data quality issues, missing mandatory fields, and business logic violations before they hit SAP's validation layer.
This reduces the number of SAP-side errors and the load on the retry mechanism.
4. Anomaly Detection
An AI agent monitoring synchronization patterns can flag unusual behavior:
- A freight booking with 50 linked Sales Orders when the average is 5
- A sudden spike in lock errors suggesting a batch job conflict
- A document that has been retried 10 times without resolution
These anomalies route to human review before they cascade into bigger problems.
The Architecture: Deterministic Backbone + Intelligent Edges
The key principle is that BPMN remains the deterministic backbone. AI agents operate at the edges — where judgment, reasoning, and adaptation add value.
┌──────────────────────────────────────────────────────────────────┐
│ CAMUNDA 8.8 PROCESS │
│ │
│ ┌─────────┐ ┌──────────────┐ ┌───────────────────────┐ │
│ │ Event │───►│ Deterministic │───►│ AI Agent Task │ │
│ │ Trigger │ │ BPMN Steps │ │ │ │
│ └─────────┘ │ │ │ ┌─────────────────┐ │ │
│ │ • Read SAP │ │ │ Ad-Hoc Subprocess │ │ │
│ │ • DMN Rules │ │ │ (Tool Belt) │ │ │
│ │ • Validate │ │ │ │ │ │
│ └──────────────┘ │ │ • Analyze Error │ │ │
│ │ │ • Check History │ │ │
│ │ │ • Read SAP Data │ │ │
│ │ │ • Suggest Fix │ │ │
│ │ └─────────────────┘ │ │
│ └───────────┬───────────┘ │
│ │ │
│ ┌──────┴──────┐ │
│ │ │ │
│ ▼ ▼ │
│ ┌───────────┐ ┌───────────┐ │
│ │ Auto-Fix │ │ Human │ │
│ │ & Continue│ │ Review │ │
│ └─────┬─────┘ └─────┬─────┘ │
│ │ │ │
│ └──────┬──────┘ │
│ ▼ │
│ ┌──────────────────┐ │
│ │ Deterministic │ │
│ │ BPMN Steps │ │
│ │ • Write to SAP │ │
│ │ • Notify Team │ │
│ │ • Audit Log │ │
│ └──────────────────┘ │
│ │
│ BPMN = Source of Truth AI = Governed by Process Boundaries │
└──────────────────────────────────────────────────────────────────┘
The BPMN model is always the source of truth. You can see exactly where AI is used, what decisions it made, and what happened as a result. This satisfies audit requirements and gives operations teams confidence that AI is governed, not running loose.
What NOT to Use AI Agents For
Not every part of the SAP orchestration process benefits from AI:
- OData API calls: These are deterministic. Call the API, handle the response. No judgment needed.
- Lock retry with backoff: This is a mechanical pattern. Wait, retry, increment. AI adds no value here.
- Field mapping for known patterns: If the mapping is well-defined and stable, DMN is simpler, faster, and more transparent than an AI agent.
- Regulatory compliance steps: Where the process must follow an exact sequence for compliance reasons, deterministic BPMN is the only acceptable approach.
AI agents should be used where the process requires reasoning about context, adapting to unexpected situations, or handling edge cases that cannot be enumerated in advance.
Comparison: Camunda Agentic vs. Standalone AI Frameworks
| Capability | Camunda 8.8 | LangChain / CrewAI / AutoGen | |---|---|---| | Process governance | BPMN-defined boundaries | None — agents run unconstrained | | Audit trail | Every step logged in Operate | Custom logging required | | Human-in-the-loop | Native user tasks with forms | Custom UI development needed | | Error handling | Boundary events, retry subprocess | Try/catch, custom retry logic | | SAP integration | OData connectors, event triggers | Custom API code required | | Multi-agent coordination | A2A connector, BPMN orchestration | Framework-specific protocols | | Production readiness | Enterprise-grade, Kubernetes-native | Requires significant hardening |
The critical difference: with Camunda, the process is always visible, always auditable, and always changeable without rewriting application code. Standalone AI frameworks give you power but no governance.
Getting Started with Agentic SAP Orchestration
┌─────────────────────────────────────────────────────────────┐
│ INCREMENTAL ADOPTION PATH │
│ │
│ Phase 1 Phase 2 Phase 3 Phase 4 │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌────────┐ │
│ │ AI-Assist│ │ AI-Auto │ │ AI-Route │ │AI-Full │ │
│ │ Human │───►│ Known │───►│ Complex │──►│Orchestr│ │
│ │ Reviews │ │ Patterns │ │ Decisions│ │-ation │ │
│ └──────────┘ └──────────┘ └──────────┘ └────────┘ │
│ │
│ Agent Agent handles Agent replaces Agent │
│ recommends, repeat errors static DMN for coordinates │
│ human autonomously context-based multi-agent │
│ approves routing workflows │
│ │
│ Risk: LOW Risk: LOW Risk: MEDIUM Risk: MED │
│ Value: HIGH Value: HIGH Value: HIGH Value: MAX │
└─────────────────────────────────────────────────────────────┘
If you already have Camunda orchestrating your SAP processes, adding agentic capabilities is an incremental step:
- Identify the judgment points — Where do humans currently make decisions in your SAP processes? These are your AI agent insertion points.
- Start with exception handling — Embed an AI agent in your existing user review tasks. Let it analyze errors and recommend actions. Humans still approve.
- Graduate to autonomous decisions — Once the agent proves reliable on a specific error pattern, let it handle that pattern autonomously while routing new patterns to human review.
- Expand to intelligent routing — Replace static DMN tables with AI agents for complex, context-dependent routing decisions.
The key is incremental adoption. You do not need to redesign your processes. You add intelligence at specific points where it delivers measurable value.
Interested in adding agentic AI to your existing Camunda + SAP processes? Book a free audit. We'll identify exactly where AI agents add value — and where they don't.