Three AI workflows I built in n8n: a phone receptionist, an SDR and an accounts-payable clerk. Each one ships as a JSON file you can import, with a diagram of how it's wired. Claude does the reading and the talking. Plain code enforces the rules around it.
Each card opens an architecture diagram, a step-by-step trace of one run and the engineering calls behind it. The JSON imports straight into n8n, with a setup note pinned to the canvas.
01● Voice agent
n8n canvas · 43 nodes● Claude
AI Voice Call Center
A Twilio phone line answered by a Claude agent. It books visits, checks job status, answers pricing questions and escalates emergencies, then scores every call once it ends.
3 Twilio webhooks · 7 agent tools · guardrails before the model
Researches, scores and answers every inbound lead in about a minute. Hot leads reach a rep with talking points, and everyone else gets a follow-up sequence that stops the moment they reply.
3 triggers · Claude in 5 roles · score caps the model can't override
Reads every emailed invoice with Claude vision and checks it against the vendor master and purchase orders. Bank-detail fraud gets stopped before money moves, approvals happen in Slack, and the bill posts to QuickBooks.
Three different jobs, one set of rules. These are the parts that decide whether a workflow survives its first bad week.
boring on purpose ↓
01
One config, one router
Every trigger lands on a single config node, then one switch sends it down its lane. Changing a threshold or a Slack channel is one edit in one place.
↳ Call Center Config → Route Twilio Event
02
Rules run as code
Emergency keywords, lead-score caps, duplicate invoices and changed bank details are checked in plain code, before or after Claude runs. The model can't override them.
A failed step takes a fallback path instead of killing the run. Extraction and QuickBooks failures land in Slack, and an AI error mid-call hands the caller to a person.
↳ AI Failure Fallback · Slack: Extraction Failed
04
Imports in one go
Every file ships a Run Once node that creates its Postgres tables and demo data, plus a setup note pinned to the canvas. Setup is mostly credentials.