EU logistics group (anonymized)

Cut Gmail triage response time from 6 hours to 6 minutes with a 17-node n8n + GPT-4o pipeline.

A French mining-logistics operator was drowning in a shared Gmail inbox handling maintenance, dispatch and accent-heavy supplier threads. Two dispatchers spent roughly 4 hours each per day classifying, summarizing and drafting replies — half of which were boilerplate.

Previous attempts at Zapier-based routing kept misreading the CC field, replying to the wrong stakeholder, or mangling French diacritics. Trust in automation had eroded to the point where every draft was being rewritten by hand anyway.

The brief was specific: route routine maintenance threads automatically, summarize multi-party chains, draft accent-safe replies in the operator's voice, and never auto-send to a CC'd executive.

The win wasn't the model — it was the gate. Every Zapier attempt before us had treated this as a one-step classify-then-reply problem. It isn't. Mining-logistics email is a five-variable problem: intent, sender role, thread depth, attachment posture and CC topology. Once we modeled that explicitly in n8n, GPT-4o became a competent junior dispatcher instead of a confident hallucinator.

The diacritics fix was almost embarrassingly small: a pre-processing node that normalizes incoming subject lines to NFC before any LLM call, plus a post-processing node that re-injects accents from a lookup table built off the client's contact list. Three lines of code apiece. Without it, the model kept reading 'Maintenance générale' as a person's name.

Six weeks in, the dispatchers were not replaced — they were promoted. They now spend their day on the 30% of threads the gate refuses to auto-draft, which is exactly the work they signed up for in the first place. Response time on routine maintenance threads collapsed from 6 hours to 6 minutes. The four SaaS subscriptions the previous setup had bolted together (parser, classifier, reply generator, CRM logger) were retired the same month.