AI Inbox Triage for a Busy Founder

A consultancy founder was spending 2+ hours a day just deciding what to read, what to forward, and what to ignore. We built an AI pipeline to do all three automatically. Now his inbox takes 15 minutes a day.

The problem

James runs a 12-person digital consultancy. Client projects, sales pipeline, BD inbound, supplier threads, team updates — all of it flows through one Gmail account. At peak, 200–250 emails a day.

The real problem wasn't volume. It was decision fatigue. Every email required him to read it before he could decide whether it was worth reading. Filters and folders didn't help — a critical client message might have a vague subject line. So he read everything, just in case. 90 minutes every morning. Again in the afternoon. A last check before bed.

He'd tried inbox-zero systems, priority flags, hired a VA to help sort things. Nothing stuck. The classification problem was still on his plate.

What we built

A Gmail-connected n8n workflow that intercepts each incoming email, classifies it with GPT-4o-mini, and routes it automatically — no human in the loop. Here's the full pipeline:

The classifier receives sender, subject, and the first 400 characters of the body. It returns one of four categories — urgent, delegate, read-later, archive — plus a one-sentence summary. That summary is what James sees in Slack. He can decide in five seconds whether to act, without opening Gmail at all.

Delegatable emails get forwarded automatically to the right team member based on a routing table in a Google Sheet. James owns that table — no code required to update it.

At 5 PM each day, n8n compiles a digest of everything tagged read-later and sends it as one Slack message. One glance, nothing missed.

Running cost: roughly $1.50/month in OpenAI API usage for 200 emails/day. The workflow runs on a VPS he already had.

The edge cases

The classifier isn't perfect — maybe one in fifty emails gets miscategorised. That's fine. A misclassified urgent item that lands in read-later still appears in the 5 PM digest. The system doesn't need to be perfect; it needs to be better than doing it manually. It is, by a large margin.

We also built in a confidence threshold: if the model returns low confidence on a classification, the email gets tagged "review" and surfaced in the digest regardless. Belt and braces.

The result

What James mentioned most wasn't the time saved — it was the anxiety that disappeared. "I used to check email constantly because I was afraid I'd missed something. Now I know that if anything urgent arrives, I'll get a Slack ping. If I don't get a ping, nothing needs me. That's a completely different way to work."