Workflow review

ai_email_triage

3 high3 medium

Risk before

69 / 100

Risk after fixes

23 / 100

Findings

Fixed automatically

4 / 6

Needs a human decision (2)

FlowGuard could not apply these fixes safely, so the fixed workflow leaves them unchanged.

Classify Email · Extract only the subject line and a truncated plain-text preview (e.g. first 200 characters) before passing to the classifier.
Build Prompt · Strip HTML to plain text and truncate to 500 characters before building the prompt.

Severity	Category	Node	Issue and fix	Status
high	cost	`Classify Email`	Uses premium model gpt-4 with maxTokens 4096 for a single-label classification task; this is 2–4× more expensive than necessary. Fix: Switch to gpt-4o-mini (or gpt-3.5-turbo) and cap maxTokens at 20 (only one label is returned).	Fixed
high	reliability	`(workflow)`	No errorWorkflow is configured; unhandled runtime errors are silently dropped with no alerting or logging. Fix: Set settings.errorWorkflow to the shared error-handler workflow ID.	Fixed
high	security	`Classify Email`	Full email body (including HTML content, sender address, and thread messages — which may contain PII) is forwarded to an external LLM (gpt-4) without any field selection or redaction. Fix: Extract only the subject line and a truncated plain-text preview (e.g. first 200 characters) before passing to the classifier.	Manual
medium	cost	`Gmail Trigger`	Polling trigger runs every minute, consuming Gmail API quota and incurring n8n execution charges even when no new emails arrive. Fix: Increase the polling interval to at least 5 minutes, or switch to Gmail push notifications if near-real-time routing is required.	Fixed
medium	cost	`Build Prompt`	The prompt concatenates the full HTML email body plus all thread message HTML with no truncation; oversized emails will send thousands of input tokens to gpt-4 per classification. Fix: Strip HTML to plain text and truncate to 500 characters before building the prompt.	Manual
medium	reliability	`Route by Category`	Switch node has no fallbackOutput; if the LLM returns any label other than support/sales/billing/spam (e.g. 'Support' with capitals, or an explanation), the item is silently dropped. Fix: Set parameters.options.fallbackOutput to an extra output connected to a logging or quarantine node.	Fixed

Fixed workflow

criticalhighmediumlow

flowchart LR
    n0["Gmail Trigger"]
    n1["Build Prompt"]
    n2["Classify Email"]
    n3["Route by Category"]
    n4["Create Zendesk Ticket"]
    n5["Notify Sales"]
    n6["Label Billing"]
    n7["Archive Spam"]
    n8["Unmatched Category"]
    n0 --> n1
    n1 --> n2
    n2 --> n3
    n3 --> n4
    n3 --> n5
    n3 --> n6
    n3 --> n7
    n3 --> n8
    classDef red fill:#fdecea,color:#8a1c12,stroke:#d0402b,stroke-width:1.5px
    classDef orange fill:#fdf0e3,color:#7a3a06,stroke:#d9731a,stroke-width:1.5px
    classDef yellow fill:#fbf5dc,color:#5e4b05,stroke:#b8960f,stroke-width:1.5px
    classDef blue fill:#e8f0fb,color:#153e75,stroke:#3d74c4,stroke-width:1.5px
    classDef clean fill:#f6f7f9,color:#2b3139,stroke:#c3c8cf
    class n0 yellow
    class n1 yellow
    class n2 orange
    class n3 yellow
    class n4 clean
    class n5 clean
    class n6 clean
    class n7 clean
    class n8 clean