AI nodes add probabilistic generation to a deterministic workflow. Connect them only after defining the task, permitted data, provider terms, model/version, schema, evaluation, cost limits, and human approval. Node availability and model names change; verify current n8n and provider docs.
Choose the Smallest AI Task
Good first tasks: classify a synthetic enquiry into an allowed taxonomy, summarize approved public text, or draft a response for review. Avoid using a model as payment verifier, legal/tax authority, identity decision, or database authorization layer.
Create separate provider credentials with minimum access and budget alerts. Never put API keys in prompts or Code nodes. Record provider, model identifier/alias, temperature/settings where exposed, prompt version, and data controls.
Maintain a provider-change checklist covering credentials, allowed data, model availability, pricing date, context/output limits, rate quotas, evaluation results, and fallback. Switching a node is a production change, not a cosmetic replacement.
Enforce Structured Output
Request a schema:
{
"category": "sales|support|other",
"summary": "string <= 200 chars",
"needs_human": true,
"evidence_refs": ["input:message"]
}
Validate after the model. Reject unknown enum, excess length, missing fields, or invented evidence. Model output cannot choose credentials or tool permissions.
Worked Example
A synthetic Lahore service enquiry is classified. OpenAI and Claude configurations are evaluated against the same 30 labelled fixtures. The workflow measures exact category match, unsafe claim rate, invalid JSON, latency, and estimated cost.
The selected configuration writes only a DRAFT classification. Low-confidence, unsupported, and protected topics route to human review. The business can swap provider only after rerunning the evaluation; no model is declared permanently best.
Failure Cases to Diagnose
- Entire CRM record sent: minimize input.
- Model name assumed current forever: record and re-evaluate changes.
- Free-form output drives branch: schema-validate.
- AI marks payment paid: use provider evidence.
- No provider outage path: queue or human fallback.
- Cost cap absent: limit input, calls, retries, and tenant spend.
🇵🇰 Pakistan Angle
Do not send CNIC, bank, health, student, employee, or private customer data to an AI provider without necessity, authority, and appropriate terms. Use synthetic or redacted fixtures during development.
Test English, Roman Urdu, and code-switched messages from the intended audience. Do not infer education, religion, ethnicity, or ability to pay from language.
Hands-On Exercise
- Define one bounded AI task.
- connect two provider nodes using protected credentials.
- enforce one output schema.
- run the same labelled fixtures.
- set cost, retry, fallback, and human gates.
Completion Rubric
- Task needs probabilistic language capability.
- Input data is minimized/authorized.
- Credentials and budgets are controlled.
- Output schema and evaluator gate branching.
- Models are compared on same fixtures.
- Human/outage fallback exists.
Sources
- n8n Docs — Advanced AI
- n8n Docs — OpenAI node
- n8n Docs — Anthropic Chat Model
- OpenAI — Data controls
Key takeaway: connect AI as a bounded, evaluated draft/classification component with minimal data and hard schemas—not as the authority for business truth.