Which processes should be automated with n8n, and which should be delegated to an AI Agent?

Good examples include syncing form leads to a CRM, creating tickets from alerts, sending invoice reminders, approving leave within defined thresholds, normalizing fields, scheduling reports, invoking an API after payment, and recording backup status. These flows benefit from transparency: every step can be observed, replayed, and reconciled.

Examples include classifying multi-intent support requests, reading emails to recommend an owner, synthesizing incidents from logs and runbooks, retrieving evidence to draft an answer, comparing contract clauses against policy, or planning an investigation before calling approved tools. An Agent should not approve spending, change access, or confirm a transaction merely because it understood the request.

A practical rule: choose n8n for high determinism and low ambiguity; consider an Agent for high ambiguity with low or reversible impact; use a hybrid with approval for high ambiguity and high impact; and keep work manual while collecting a baseline if evaluation data and auditability are missing.

A reference flow: n8n receives an event → validates schema and tenant → removes unnecessary sensitive data → the Agent classifies or proposes a plan → structured output is validated → a policy engine decides whether approval is required → n8n calls APIs with narrow credentials → audit events and metrics are recorded. The Agent neither owns credentials directly nor expands its own toolset.

Sales: route leads by territory with n8n; summarize calls and suggest next steps with an Agent; update CRM through a hybrid flow with schema validation. Support: ticket creation and SLA timers use n8n; intent understanding and answer retrieval use an Agent; refunds require policy gates and approval. Finance: rule-based reconciliation uses n8n; exception explanation may use an Agent; payment requires independent controls. IT/DevOps: alert routing and proven remediation use n8n; log investigation and runbook recommendations use an Agent; production changes require approval, a change window, and rollback.

Minimum guardrails include tool allowlists, least privilege, step and budget limits, structured outputs, input/output validation, prompt-injection defenses, idempotency, timeouts, controlled retries, human approval for sensitive tools, immutable audit records, and a kill switch. RAG must filter by tenant and authorization; answers without sufficient evidence should abstain or escalate.

Measure workflows and Agents separately. For n8n: success rate, failure/retry rate, p95 latency, duplicate executions, and processing time. For Agents: task completion, groundedness/citation accuracy, tool-selection accuracy, escalation rate, cost per successful task, and policy-block rate. Compare both against the manual baseline for time, errors, and cost; token count and workflow count are not business KPIs.

Can n8n run an AI Agent?

Yes. n8n offers an AI Agent node and tool integrations; permissions, approvals, and observability still require deliberate design.

Should every workflow be replaced by an Agent?

No. Deterministic steps should remain code or workflow for testability, cost, and reliability.

Is a chatbot an AI Agent?

Not necessarily. A response-only chatbot may not plan or call tools. Agents typically pursue a goal and use bounded tools and state.

When is human-in-the-loop required?

When an action affects finances, legal commitments, data, access, customers, or is difficult to reverse.

Where should a pilot begin?

Choose a high-volume, low-risk process with historical data and explicit pass/fail criteria; run in shadow mode before enabling actions.