AI Email Support Agent

Agentic email-support orchestration on Camunda 8 Self-Managed. An inbound customer email triggers a process that first checks domain scope, then fetches prior conversation context, queries a RAG knowledge base, lets an agent decide, and either resolves the case autonomously or escalates to a human — with the resolution saved back to long-term memory. An "Agent as a Judge" step evaluates each interaction and routes the outcome.

Blueprint: https://marketplace.camunda.com/en-US/apps/522492/ai-email-support-agent

Camunda version support

The upstream blueprint targets Camunda 8.8. This repository runs on both, and is currently developed against 8.10-SNAPSHOT.

Version	Status
8.10-SNAPSHOT	Full flow runs end to end, including the agent-instance history writes. Current target.
8.8.x (8.8.33 / connectors 8.8.15)	Full flow runs end to end.

An earlier 8.10-SNAPSHOT image pair failed at the AI Agent connector's history write with 400: Request property [loopIteration] cannot be parsed. That was a version skew, not a defect in this repository: AgentInstanceHistoryItem renamed iteration to loopIteration, and the connectors-bundle and camunda SNAPSHOT images are published on separate build pipelines, so a pair straddling that rename disagrees on the field name. Current images are aligned and the error no longer reproduces. Because SNAPSHOT tags roll, pin both images by digest rather than relying on the tag.

How it works

1. Inbound email — an IMAP inbound connector polls for unseen mail. Messages with no In-Reply-To header start a new case; replies correlate back into a waiting instance by message id.
2. Domain-scope guardrail — a DMN decision classifies the email by keyword before any LLM call. OUT_OF_SCOPE routes straight to an escalation end event, so off-domain requests never reach the agent.
3. Long-term memory — fetch-past-conversations loads the sender's prior interactions from their own Elasticsearch index, keyed by email address, and injects them into the agent's context.
4. Agent loop — an AI Agent ad-hoc sub-process reasons over the request and selects its own tools: query the RAG knowledge base, ask the customer a follow-up by email, or hand control to a loan specialist. The loop runs until the agent answers or hits its call limit.
5. Retrieval — query-knowledge-base embeds the agent's query and runs kNN against a dense-vector index, returning hits above a cosine floor. A miss routes the agent down the "knowledge base empty" path instead of inventing an answer.
6. Write-back — the resolved interaction is embedded and stored in the customer's conversation index, so the next email from that address arrives with history.
7. Agent as a Judge — a second agent evaluates the exchange and returns a verdict: resolve autonomously and email the customer, send for manual review, or escalate to human control.

Components

Path	What it is
process/bpmn/	The BPMN process definition and its embedded agent/judge prompts
process/dmn/	DMN decision table (domain-scope guardrail)
workers/	Spring Boot job workers (Java 21): fetch-past-conversations, query-knowledge-base, save-customer-interaction, save-to-knowledge-base; RAG + Elasticsearch dense-vector logic; PII redaction guardrail
services/llm-stub/	Local LLM stub the flow calls in place of a paid LLM. Drives the agent's tool call and returns the judge's verdict JSON
services/embeddings/	all-mpnet-base-v2 sentence-transformers embedding service
images/	Diagrams used by this README
LIFECYCLE.md	Versioning and rollback procedure

The Docker Compose file and all secrets are intentionally not committed. The repo versions code and process definitions; the running stack is assembled locally from the official Camunda images.

Guardrails (governance)

• Out-of-scope rejection (deterministic) — process/dmn/domain-scope-guardrail.dmn classifies each inbound email by keyword before any LLM call. A DMN business rule task runs first; a gateway routes OUT_OF_SCOPE messages to an escalation end event, so off-domain requests never reach the agent. Deterministic and provable, independent of the LLM.
• Agent as a Judge — every interaction is evaluated, and the verdict routes the case to autonomous resolution, manual review, or human control.

Observability

• Operate — process and task state, DMN decision evaluations, and the AI Agent panel showing the agent instance's conversation and status
• Optimize — heat maps and analysis (zeebe-record-* via the exporter)

Attribution & License

Built on Camunda 8 and the AI Email Support Agent blueprint from the Camunda Marketplace. The orchestration patterns and the BPMN foundation are Camunda's; the guardrails (PII redaction, DMN domain-scope classifier), workers, supporting services, and documentation are original additions authored for learning and demonstration.

The original contributions in this repository are released under the MIT License (see LICENSE) — in the same open spirit that Camunda shares its blueprints. Camunda's platform, connectors, and original blueprint BPMN remain Camunda's and are governed by Camunda's own license terms.