AI Automation
Turning a repetitive back-office process into an automation nobody has to think about.
- Status
- LIVE
- Date
- January 2026
- Stack
- n8n / AI / Node.js / SQLite
The problem
A business was receiving supplier invoices as PDFs and photographs, then retyping every line into a spreadsheet. Four hours a week, every week, with the kind of transcription mistakes that only surface at the end of the quarter.
The idea
The process was already well defined — it just happened to be executed by a person. Extract the structure with a model, validate it against rules that already existed, and only involve a human for the cases that genuinely need judgement.
Building it
How it came together, step by step.
Writing down the rules first
Before any model was involved, the existing implicit rules were written out: which fields matter, what a valid total looks like, which suppliers format dates backwards.
Terminal
Extraction with a confidence score
The model returns structured fields plus how sure it is. Anything below the threshold is queued for a human rather than silently guessed.
Editor
Wiring it together in n8n
Inbox trigger, extraction, validation, database write, and a notification only when something needs attention.
Interface
A review queue for the edge cases
Roughly one invoice in twelve needs a human glance. That review takes seconds instead of minutes, from a phone.
Mobile
Features
What it actually does.
Document extraction
PDFs, scans and phone photographs into structured, validated fields.
Confidence thresholds
Low-confidence extractions route to a human instead of being guessed.
Rule validation
Totals, tax and dates checked against the rules the business already had.
Review queue
The exceptions surface in one place and clear in seconds.
Audit trail
Every extraction keeps the source document alongside what was read from it.
Runs unattended
Triggered by the inbox. It only speaks up when something needs a decision.
Tech stack
What it is built with.
- n8n
- AI
- Node.js
- SQLite
The result
Where it ended up.
- Hours saved weekly
- 4
- Needs a human
- 1 in 12
- Transcription errors
- ~0
- Runs per week
- 60+
Build in public