Case study · Inbox automation · Customer Service
Every channel into one triage system. Every reply drafted. Nothing sent without a human.
A multi-channel consumer brand
Inbox automation
Noise filtering
Reply drafting
Human approval
7
mailboxes triaged
81
noise messages filtered in one morning
0
replies sent without human approval
48
workflow improvements
Around 50 customer emails a day in peak season, plus 10 to 15 direct messages across Shopify and Meta, all landing in the same unmanaged stream.
Real enquiries sat behind newsletters and notifications for days. Wholesale pricing questions, refund requests and delivery problems waited while a team member cleared the backlog between 8pm and 1am.
Nothing was shared. What had been answered, and what still needed a reply, lived in people's heads.
Each channel had its own app, its own notifications and its own version of unread. Nothing converged anywhere except on the team.
Noise buried real enquiries Newsletters and notifications sat in front of wholesale, refund and delivery questions for days.
Late-night backlog A team member cleared the inbox between 8pm and 1am just to keep the queue moving.
Six moving parts, from raw intake to an approved reply.
Seven mailboxes plus WhatsApp, Meta DMs and Shopify Inbox feed one automation layer: n8n workflows running on Railway.
Dashboard-editable rules drop newsletters and notifications before the AI sees anything. One 8.5-hour morning removed 81 messages.
Every remaining message is classified FLAG or DROP. A question inside a notification means FLAG; no question means DROP. The system reads the body, not the subject.
Flagged messages get a reply drafted in the brand's voice, saved into that mailbox's own Drafts folder with a clear [AI DRAFT] prefix.
Follow one enquiry from arrival to approved reply.
A wholesale enquiry lands in one of seven mailboxes.
Editable rules drop pure newsletters and notifications first.
The system reads the whole thread and classifies FLAG or DROP.
Six layers, each with one job. The review layer is not a feature of the system; it is the system.
Every place a customer can write in, all feeding one intake.
n8n workflows on Railway move every message from intake to decision, with Google Sheets feeding the dashboard.
The system prepares the work; people keep responsibility for the customer relationship. Every reply is saved as an [AI DRAFT] in the mailbox it arrived in, and a person reviews, edits, escalates or sends it.
Sensitive messages never become a draft. Refunds above an agreed threshold, legal or regulator mentions and diet-critical questions go straight to named people on the team.
Auto-send disabled by design
Every reply saved as an [AI DRAFT] in the correct mailbox
Sensitive messages escalate to named team members
Human corrections improve future rules and responses
Early results from the live system, measured on real customer messages.
Every customer channel now lands in one place, from Outlook and Gmail to WhatsApp, Meta and Shopify Inbox. The first week alone processed more than 50 messages and drafted more than 10 replies.
One busy morning, the filter removed 81 newsletters and notifications before they could bury a real enquiry, each one logged.
The system kept getting sharper as the team corrected it. Every improvement was shaped by real customer conversations.
What started as a handful of drafts a week became hundreds, every one still approved by a person before sending.
The inbox became manageable again Real enquiries surface ahead of the noise, with a drafted reply ready for review instead of a backlog to dig through.
Late-night clearing became a process The team works from two queues and a morning digest instead of raw inboxes. Nobody clears a backlog between 8pm and 1am anymore.
Every correction improves the system Edits become better rules, sharper escalation logic and clearer knowledge-base guidance for the next reply.
The system became reusable Freckleberry asked about extending it to another venture. The same workflow can support more channels and more teams.
Automation handles the repetitive preparation. Decisions that touch trust, money, safety or the customer relationship stay with people.
Every customer-facing send. The system drafts; people send.
Pricing decisions, which live in a knowledge base the client owns.
Anything legal, financial or diet-critical.
Refunds above the agreed threshold.
Weather dispatch holds for heat-sensitive orders. The system flags, a person decides.
This build fits founders and teams who recognise their mornings here.
Multi-channel brands juggling email, social and commerce messages at once.
Teams losing real enquiries to newsletter and notification noise.
Founders whose inbox knowledge lives in one person's memory.
Operators still clearing the inbox late at night.
Businesses that want AI speed without unsupervised sending.
We can map the workflow, identify what should be automated and build the review layer your team can trust. Built on your stack, owned by you, with people keeping the final word.
See how messages from seven inboxes are filtered, triaged and turned into reviewed reply drafts before anything reaches a customer.
No shared record What had been answered and what still needed a reply lived in people's heads.
Heat-sensitive dispatch risk Orders heading into hot weather needed a manual check before they left.
Refunds above a set threshold, legal or regulator mentions and diet-critical questions route to named humans instead of a draft.
The team works two dashboard queues, Alert and Human Review, edits what needs editing and presses send. A 7am digest ranks what still needs a person.
Refunds over threshold, legal mentions and diet-critical questions route to named humans.
A reply is drafted into the right mailbox, prefixed [AI DRAFT].
A person reads, edits and approves. Only then does anything send.
At 7am the digest ranks what arrived, what was drafted and what still needs a person.
Each message is classified and each reply drafted in the brand's voice.
The golden rule decides FLAG or DROP. The escalation matrix routes refunds, legal and diet-critical messages to named humans. The Weather Dispatch Hold pauses heat-sensitive orders.
The layer that keeps trust: nothing sends until a person approves it.
A person reads each draft, edits where needed and sends by hand. The dashboard and the 7am digest keep the whole flow visible.
What we learned Thread context beats subject lines. Pricing, policies and approved answers belong in a knowledge base the client controls. Draft-first automation is the trust layer: the system prepares the work so human judgement lands where it matters.
Case study
We replaced scattered sales reporting with a single live sales performance dashboard — every deal, stage and automation signal in one screen, refreshed automatically every day.