Skip to main content
QA review · Order entry

Orders from emails into the ERP without retyping

We took an existing Make scenario from a building-materials dealer apart in a QA review and rebuilt it as a hardened n8n workflow. The classification now steers the flow, uncertain matches get a review flag, nothing is lost silently.

Detected ambiguities are surfacedinstead of silently waved through
MailboxPDF✦Classifyorder?Line itemsMaster datacustomer · itemOrder fileERPconfident?Change / queryREVIEWHuman checksunclear

Pain

Orders arrive as PDFs by email. Someone types the line items into the ERP, mixes up item numbers or files a change as a new order. In the reviewed scenario the AI classification steered nothing, special characters broke the order file and items without a match vanished silently. Errors surfaced at the customer.

Solution

  • Classify every email as order, change, query or spam and let only orders into processing.
  • Extract header and line data from PDF and email body into a fixed structure, with a confidence per field.
  • Match the customer and every line item against master data; without a confident hit the line gets a review flag instead of disappearing.
  • Build the order file structured and escaped so special characters cannot break the ERP import.
  • Run an error path that reports every abort instead of swallowing it.

How it works

  1. 01
    Email intake
    Mailbox trigger collects the attachments, with a short wait for PDFs that arrive late.
  2. 02
    Classify
    An LLM sorts the mail: order, change, query, spam. Only orders continue, everything else goes to the review path.
  3. 03
    Extract
    Header and line data from PDF plus email body are turned into a fixed structure, with a confidence per field.
  4. 04
    Match
    Customer and items against master data; uncertain hits carry the REVIEW flag.
  5. 05
    Order file
    Generated structured and escaped, then stored for ERP takeover.
  6. 06
    Error path
    Every abort raises a notification to the owner. No order is lost silently.

Outcome

The rebuild closes the three break points of the original scenario: the classification steers the flow, special characters no longer break the file, no item is dropped silently. What remains is a review path for the cases a human has to decide.

Why this stack

n8n
Error paths, retries and logging in one place, operable in-house.
LLM extraction
Reads order PDFs of varying formats and returns a confidence per field.
Master-data matching
The match against customer and item master decides what may enter the ERP.
ERP import
A structured order file instead of freely assembled text.
More about our AI & Automation services

Related use cases

WebhookPOST /lead{ }Validate✦ClaudeclassifyCRMappend: rowGGeminireplyGmailsend: replysuccess{ }Error: data{ }Error: CRMTelegram alerterror
Own toolchain · 01

n8n workflows with LLM classification

Webhook → validate data → Claude classifies (heat score, intent) → save to CRM → Gemini drafts a personalized reply → Gmail. Plus an error branch with a Telegram alert.

one workflowper integration instead of a separate service
n8nClaudeGeminiWebhooks
Learn more→
Telegram voice message on a smartphone being transcribed, classified and stored as a task list on EU server infrastructure
Own toolchain · 03

Task dashboard with voice input

Telegram voice note → Whisper transcript → automatically classified tasks in the backlog. Cron jobs surface reminders.

voice instead of typingclassified tasks at no extra effort
Whisper (self-hosted)Telegram Botn8nLlama / Mistral
Learn more→