The default assumption is usually backwards
Most teams that come to us assume every automation problem is an AI problem. It usually isn’t. A surprising amount of what looks like judgment is actually a rules problem wearing a judgment costume — and rules problems are faster, cheaper, and more predictable to solve with a plain n8n workflow than with an LLM call.
The question we ask first isn’t “can AI do this” — it almost always can. It’s “does this task actually require reading intent, tone, or unstructured context, or is it a lookup with extra steps.” That question is the whole playbook.
When it’s a script, not a skill
If the input is structured — a form field, a status value, a number against a threshold — and the decision can be written as an if/then, it’s a script. Routing a support ticket by the category the customer already selected, flagging an invoice that’s 30 days overdue, moving a deal to a new pipeline stage when a field changes: none of that needs a model. It needs a workflow, and a workflow is faster to run, cheaper to operate, and doesn’t have a failure mode where it confidently gets the wrong answer.
We’ve replaced more than one client’s expensive “AI automation” with a five-node n8n workflow that does the same job in milliseconds instead of seconds, for a fraction of the running cost — because the task never needed a model in the first place.
When it’s genuinely a Claude problem
The other side of the line: anything where the input is unstructured language and the output requires understanding intent, tone, or context that isn’t captured in a field. Drafting a first-pass reply to an ambiguous support email. Summarizing a long thread before a handoff. Scoring a resume against a role’s actual requirements instead of keyword matching. Triaging an inbound message that could mean five different things depending on how it’s phrased.
In those cases, a rules engine either can’t do the job or does it so badly it creates more cleanup work than it saves. That’s where a scoped Claude skill — not a generic chatbot, a skill built against your specific playbook and edge cases — is the right tool.
The build pattern we use most
In practice, most of the systems we ship are both: n8n handles intake, routing, retries, and anything with a clear rule, and hands off to a Claude skill only for the one step that genuinely needs judgment — then n8n takes back over for whatever happens after. That split keeps the expensive, slower, judgment-based step small and auditable, instead of routing everything through an LLM because it’s available.
If you’re not sure which side of the line your task falls on, that’s usually the first thing we sort out on a discovery call — often before we’ve talked pricing, because it changes what the build even looks like.