24. AI Client Onboarding Agent: Auto Welcome Email Generator

A single slow external call can block the entire workflow execution. At high trigger frequency, accumulated stuck executions can exhaust the n8n execution queue.

Set a timeout under the node's Additional Options. A value of 10,000–30,000 ms (10–30 seconds) is a reasonable starting point for most APIs; adjust based on the provider's documented SLA.

These expressions work reliably against test data where all fields are present. In production, fields may be absent, null, or differently typed — causing a runtime error or silent null propagation at the point of the expression.

• In `text`, change {{ $('Trigger on New Client Form Submission').item.json['Client name'] }} to {{ $('Trigger on New Client Form Submission')?.item?.json['Client name'] }}.
• In `text`, change {{ $json.Checklist }} to {{ $json?.Checklist }}.
• In `text`, change {{ $('Extract and Structure Client Data').item.json.fields }} to {{ $('Extract and Structure Client Data')?.item?.json?.fields }}.

These expressions work reliably against test data where all fields are present. In production, fields may be absent, null, or differently typed — causing a runtime error or silent null propagation at the point of the expression.

• In `sendTo`, change {{ $('Trigger on New Client Form Submission').item.json[' email '] }} to {{ $('Trigger on New Client Form Submission')?.item?.json[' email '] }}.
• In `message`, change {{ $json.text }} to {{ $json?.text }}.
• In `subject`, change {{ $('Trigger on New Client Form Submission').item.json['Client name'] }} to {{ $('Trigger on New Client Form Submission')?.item?.json['Client name'] }}.

Operators cannot tell from external signals that a run failed. Silent failures go undetected until a downstream system shows missing data or a user reports a problem.

Connect a dedicated Error Workflow in workflow settings, or add a per-node onError: "continueErrorOutput" path that sends an alert via Slack, email, or a logging service.

Without a cap, models occasionally produce very long outputs — particularly when given an ambiguous or open-ended prompt. This increases cost unpredictably and can cause downstream parsing to fail if the consumer expects a bounded response.

Set maxTokens to a value appropriate for the task. Classification tasks rarely need more than 100 tokens; structured JSON output typically needs 200–500; long-form summaries may need 1,000–2,000. Start conservative and increase if needed.

Free-form text output varies between invocations. The downstream node assumes a specific format to parse or act on. When the model produces a slightly different structure — different key names, trailing text, markdown formatting — the downstream logic fails intermittently in ways that are hard to debug.

Enable JSON mode (responseFormat: "json") on the AI node or connect an Output Parser sub-node with a defined schema. This forces the model to produce output in a predictable format and fails fast when it does not.

If this action is disputed, replayed, or causes a data inconsistency, there is no structured record of what was sent, when, and with what data. Debugging requires cross-referencing n8n execution history with the external system's own logs.

After the write node, add a step that records the key fields of the action to a log — a Google Sheet row append, a database insert to an audit table, or a Slack message to a dedicated ops channel. Include: timestamp, action type, affected ID, and result.

In production, agents can enter loops — repeatedly calling a tool that returns ambiguous results. Without an iteration cap, a single execution can generate hundreds of tool calls, causing unpredictable cost spikes and potentially triggering rate limits on connected APIs.

Set maxIterations to 10–20 for most agentic tasks. Monitor actual iteration counts in execution history and adjust up if legitimate tasks are being cut short.

Token costs scale with the size of the payload. A full $json object or a field containing HTML, transcripts, or email bodies can be orders of magnitude larger than the task actually requires.

Select only the fields the model needs for its task. If the content must be large (e.g., summarizing a full document), set a maxTokens output cap and consider chunking the input across multiple calls.

These expressions work reliably against test data where all fields are present. In production, fields may be absent, null, or differently typed — causing a runtime error or silent null propagation at the point of the expression.

• In `assignments.assignments[0].value`, change {{ $json['Client name'] }} to {{ $json['Client name'] ?? '' }}.

• Full payload in prompt — unbounded input tokens
• No output cap — unbounded output tokens
• Agent iteration risk — no max iterations set

Set maxTokens to a value appropriate for the task. Classification tasks rarely need more than 100 tokens; structured JSON output typically needs 200–500; long-form summaries may need 1,000–2,000. Start conservative and increase if needed.