Large language models are excellent when language, messy input, or judgment is the bottleneck. They are a poor fit when the job is: if status is X, call API Y.
At isomore.tech we use both. The mistake is using one for everything.
Use n8n (or scripts) when…
- Inputs are structured (IDs, statuses, dates)
- Rules are stable and auditability matters
- The path is a state machine (stages, retries, SLAs)
- Failure should be deterministic and easy to debug
Examples: routing leads, syncing CRM fields, scheduled invoice pulls, webhook fan-out.
Use Claude (or a custom skill) when…
- Input is free text or multi-format documents
- Output needs drafting (email, summary, brief)
- Classification is fuzzy but checkable by a human or rule
- The model step has a clear success criterion
Examples: summarize a support thread, draft a collections note, triage messy intake forms.
Hybrid pattern that works
- n8n moves data and enforces state
- Claude fills one narrow step
- A gate (human or rule) decides what ships
Example: invoice exception arrives → Claude drafts a reply → n8n posts draft to a queue → finance approves → send.
If you cannot describe the success criteria for the model step, do not automate it yet.
Decision table
| Signal | Prefer |
|---|---|
| Same input → same output every time | Script / n8n |
| Needs reading nuance | Claude |
| Compliance / audit trail critical | Script first |
| High cost of a wrong send | Human gate after Claude |
| High volume, simple rules | n8n |
Common failure modes
- LLM as router for hard business rules → silent wrong paths
- No state machine → model re-decides from scratch every run
- No ownership of exceptions → drafts pile up unread
How we scope it
Discovery names the workflow. Scope states which steps are deterministic vs model-assisted. Build is fixed-price; you keep the source. Brand: isomore.tech · registered entity Compass Ventures LLP.
Next step
Not sure where Claude ends and n8n begins on your stack? Book a discovery call—bring one painful workflow and we will split the jobs.