These are 40 customer messages that arrived in one day at a fictional AC and small-repair shop - website form, WhatsApp, and Instagram. Open any one: that is the raw inbox. Cheap rules clear the obvious noise. A model reads only what survives and routes it. A person only sees what still needs a human.
Read top to bottom: the number a person never opens → where survivors went → who the rules closed (and why) → the raw pack.
n8n
WhatsApp
Instagram
CRM
Bayline Comfort Co., Port Meridian - fictional stand-in for any shop that answers its own inbox. AC install and service, small plumbing and electrical.
A person never opens
- / 40
Opens only - - manager queue and disputed ones.
Recorded run on synthetic requests. Bayline and Port Meridian are fictional. Changing thresholds re-runs the rules; the model answers stay the recorded ones.
Try a change:
Thresholds for the next run
Change the area, the minimum, the shift or the auto-route bar, then run again. Rules move with the new numbers. Model opinions stay the recorded ones.
Where the rest went
Survivors of the rules. The model sorted them. Disputed ones still need a person.
Closed by a rule
A rule you can argue with and correct. Open a rule to see who it closed, then open a request for the sentence it used.
The pack itself
Everything that came in, before any rule touched it. Open it and read the raw text - the decisions above are made from exactly this.
Show all inbound requests
The automation
Scroll sideways on a narrow screen to read the whole path.
Export of a working n8n workflow, not a drawing of one.