CASE STUDY 04 / AI-ASSISTED SUPPORT
Customer supportAgent.
Generate email drafts from order, product, and conversation context, with controlled delivery rules and human review for sensitive cases.
System overview
Inspect the workflowsThe agent gathers relevant context and drafts a response, then follows defined delivery rules or asks a person to review the case.
Email received
Resolve incoming Gmail messages into cases and use deduplication to keep repeated events coordinated.
Context assembled
Retrieve relevant order, product, FAQ, and conversation information before generating a structured draft.
Draft evaluated
Evaluate the draft against the configured policy. Sensitive or uncertain cases require a reviewer.
Send or review
Allow eligible low-risk delivery or route to Telegram for approval, feedback, rejection, or case linking.
Risk-based routing · capped redrafts
- Bounded autonomy
- Automatic delivery is limited to supported low-risk cases under the configured send policy.
- Reviewer control
- Staff can approve, reject, or provide feedback, with a maximum of three redrafts.
- Stalled-state recovery
- Recovery checks surface prolonged processing and uncertain sends for reviewer attention.
THE CHALLENGE
The operational challenge.
A helpful support draft depends on the right order, product, and conversation context. Automatically sending every AI-generated answer would give sensitive and uncertain cases too little oversight.
THE SOLUTION
The system design.
The agent resolves email threads into cases, retrieves relevant context, and creates structured drafts. Supported low-risk cases can follow a narrow automatic-delivery policy; other cases go to Telegram for review.
DESIGNING THE SYSTEM
Key engineering decisions.
Retrieve only relevant context.
Conditional retrieval brings together order, product, FAQ, and memory information. Deduplication and thread-to-case resolution keep repeat messages attached to the right case.
Put boundaries around autonomy.
Sensitive or uncertain drafts require human review. Reviewers can approve, reject, provide feedback, or link a case, with a maximum of three redrafts.
Make stalled work visible.
Recovery checks cover processing older than ten minutes and uncertain sends older than five minutes. A separate error handler catches execution failures.
SYSTEM ARCHITECTURE
Integrations and recovery paths.
The overview above shows the main path. Explore the architecture for the integrations, decision points, and recovery paths behind it.
INSIDE THE SYSTEM
Explore the actual workflows.
Inspect the intake paths, integrations, review steps, and recovery routines behind this system. Open any screenshot to read the original n8n canvas.
Context & drafting
Support agent & contextual drafting
Combines email intake, order and product context, structured reply drafting, evaluation, and delivery routing.
Human review
Telegram human review
Coordinates staff approval, feedback, redrafting, and case actions through Telegram.
Monitoring & recovery
Ticket recovery monitor
Finds stalled processing and uncertain automatic sends, updates recovery state, and alerts the reviewer.
Automation error handler
Connects execution errors to a configured Telegram notification for the reviewer.
EVIDENCE & OUTCOMES
What the evidence shows.
Four workflows and 147 nodes are described in the source, with configuration and selected screenshots supporting the case-handling and review narrative.
A portfolio demonstration of context retrieval, bounded drafting, and human oversight. No production response-time or savings claim is made.
Limitations and next steps.
Production outcomes have not been measured. The next step is validating the send policy and recovery behavior against a representative test set before live use.
Technology stack.
- n8n
- Gmail
- Google Sheets
- OpenAI
- Telegram
Source: owner-supplied case study and diagrams. Scope and test results are reported in the document, not independently audited.