An inbox that sorts itself, and asks a person when it should
How a mid-sized online retailer let AI draft routine replies, routed urgent emails directly to Slack, and won back 106 hours a month.
01 At a glance
The short version
A mid-sized online retailer with one busy support inbox. We put an AI sorter in front of it. It rates every email, drafts replies for the routine ones and pings a person on Slack for the ones that matter.
Mid-sized online retailer
About 3,000 customer emails a month land in one shared support inbox. Order questions, returns, delivery problems and the odd angry message, all in the same pile.
Every email waits in the same line
A broken-parcel complaint sits behind a question about shipping times. Someone reads each one, works out how bad it is, then writes a reply. At 5 minutes each, that is about 250 hours a month.
AI sorts, a person decides
Each email gets an urgency score from 1 to 5. Routine ones get an AI draft, approved in Slack with one click. Urgent ones go to a person with a reply box.
43% less time on email
By month 3, time spent dropped from 250 hours a month to 144. Urgent emails receive a response within 10 minutes.
The shift
Before: one inbox, one queue, and every email read cold by whoever is next. After: the inbox is sorted before anyone opens it. Routine replies arrive drafted, urgent emails reach a person on Slack, and an urgent email never goes quiet, even when the team is busy.
02 The pain
Urgent and routine
look exactly the same
An inbox can't tell which emails matter. People can, but only after they have opened each one.
Nothing jumps the queue
A customer whose order arrived smashed writes the same kind of email as one asking about delivery times. Both sit in one list, in the order they arrived.
Someone reads every one
Each email is opened, understood, checked and answered by hand. Most take about 5 minutes. Across 3,000 emails, that is about 250 hours a month.
Routine questions get routine answers
Many emails ask something the team has answered hundreds of times. Writing the same reply again is the dull part of the job, and the part that gets rushed.
A busy team means a silent inbox
When the team is stretched, a customer in trouble hears nothing until someone reaches them. They can't even tell whether the message arrived.
03 The solution
Sort first. Draft the routine.
Ask a person for the rest.
A workflow (a chain of automatic steps) in n8n watches the support inbox. Every new email is read by an AI model, which rates how urgent it is from 1 to 5. A 4 or 5 means a person is needed. Anything lower can be answered from a standard reply.
Each email is read the moment it lands. The AI gives it a score from 1 to 5, and saves 4 and 5 for emails that need a person to fix the problem.
Scores of 4 or 5 go to the urgent route, with a Slack alert and a reply box. Scores of 1 to 3 go to the routine route, where a second AI writes the draft.
A person types the urgent reply or approves the routine draft in Slack, and the email goes out. If no one answers an urgent alert in 10 minutes, the customer gets a holding note.
A person stays in the loop
No AI-written reply goes out until someone approves it. Urgent replies are typed by a person. The one email the system sends by itself is the holding note, a fixed message that only fires when an urgent alert has sat unanswered for 10 minutes.
04 Implementation
What we set up,
and what each part does
Six parts, one workflow. It sits on top of the inbox the team already uses, so customers write in exactly as they do now.
| Part | What it gives the team |
|---|---|
| Inbox watcher | An email trigger using IMAP (the standard way programs read an inbox) fires on every new message in the support mailbox. |
| Urgency rater | An AI agent on gpt-4o reads the subject and body and returns one number from 1 to 5. Its prompt (the written instructions) reserves 4 and 5 for emails that need a person. |
| The router | A switch sends scores of 4 and 5 down the urgent route and everything below 4 down the routine route. |
| Slack reply box | For urgent emails, a Slack message shows the customer's email and a form. The person on duty types the reply there. The form waits 10 minutes. |
| Draft writer | For routine emails, a second AI agent writes a subject and a body. Its prompt sets a casual, short tone and the shop's name. |
| Approve and send | The draft goes to Slack beside the customer's email for a yes or no. A yes sends it from the support address as a reply. |
Link the support inbox, the Slack channel and the AI model. Write the two prompts: one for rating urgency, one for the reply tone and sign-off.
Send past emails (names removed) from a test address. Check every score and read every draft. Adjust the prompts until the scores match what the team would have said.
Switch it on for the live inbox. For the first fortnight, the team reads every score and every draft and fixes the prompts where one is off.
05 Under the hood
The real build,
on the n8n canvas
This is the workflow as it sits in n8n, the tool that runs it. Each box is one step. Each line shows where the email goes next. An email enters on the left, gets a score, then takes one of two routes.
Email Trigger fires on each new email. Classify Emails Urgency scores it. The Switch then asks one question: is the score 4 or above?
Get User Input sends the Slack alert and reply box. The next step checks the 10-minute timer. A typed reply goes out by email. No reply sends the holding note.
AI Agent writes the draft. Get Approval puts it in Slack. An If step reads the answer. A yes sends the reply. A no stops there, so a person writes that one.
The small boxes underneath
Each AI step has two small boxes below it. The round one is the AI model, gpt-4o. The curly-bracket one is an output parser (a rule that fixes the shape of the answer): one number for the rating, a subject and a body for the draft. No tools or memory are attached, so each email is judged on its own.
06 The rating
How each email
gets its score
| Customer email | Score | Action taken |
|---|---|---|
| Do you ship to New Zealand? | 1 | The AI drafts a reply. Someone approves it. |
| Where is my order? It was due Tuesday. | 3 | The AI drafts a reply. A person approves it or rewrites it with the tracking details. |
| My parcel arrived smashed and I need a replacement by Friday. | 4 | Urgent route. A Slack alert opens a reply box for a person. |
| I was charged twice and my bank is chasing me. | 5 | Urgent route. A person writes the reply, or a holding note goes out at 10 minutes. |
4 and 5 are for people
The prompt tells the AI to give a 4 or 5 only when the email needs a person to fix the problem. Anything a standard guide could answer scores lower.
It keeps the rule simple
The AI returns a single number and nothing else. The router then asks one question: is it 4 or above?
The team can move the line
Too many urgent alerts in Slack? Move the cut-off from 4 to 5. Too few? Drop it to 3. It is one number in one place.
Skip what needs no reply
Promotional emails and automated notices can be filtered out before the AI sees them, so nobody approves a reply to a newsletter.
07 The two routes
Who writes the reply,
and when
Both routes end the same way: an email from the support address. What differs is who writes it.
Urgent route: a person writes it
A Slack message headed “Urgent attention needed” shows the sender, the subject and the full email, and asks what the reply should be.
A form opens with one box. A person types the reply and submits. The workflow emails that text to the customer as written.
If no one submits in 10 minutes, the customer gets a short holding note. The timer is a setting, so it can be 5 minutes or 30.
Routine route: the AI drafts it
A second AI reads the email and writes a subject and a body. The prompt asks for a casual, crisp reply that gets to the point.
The draft lands in Slack beside the customer's email. A person reads both and clicks to approve.
On approval, the reply goes out to the customer as “Re:” plus the original subject, so it stays in the same thread.
08 The outcome
Delivered operational results
Measured results tracked across the 90-day deployment.
| Area | Before | Delivered at 90 days |
|---|---|---|
| Time on email | About 250 hours a month, all by hand | About 144 hours a month by month 3, 43% less |
| Urgent emails | Wait in the queue with everything else | A Slack alert at once, and a reply or holding note within 10 minutes |
| Routine emails | Written from scratch, about 5 minutes each | Drafted by the AI. 6 in 10 approved as written by month 3 |
| A busy team | A customer in trouble hears nothing | Every urgent email gets at least a holding note |
The 90-day rollout
The workflow ran on the live inbox. The team reviewed every score and draft, noting required changes. Four in ten drafts went out as written.
Refined prompts based on team feedback. Added store returns and delivery policies to the prompt context. Balanced the urgency threshold for Slack alerts.
Verified operational time savings against baseline figures. Connected automated order lookup to answer tracking queries directly.
Verified operational return
These numbers reflect verified client metrics across live support queues. With triage automated and routine replies pre-drafted, the team maintained consistent sub-10-minute response times for urgent cases while cutting 106 hours a month from support operations.
09 The numbers
Delivered results
at 43% less time
Across 3,000 monthly emails, staff time dropped as the share of AI drafts approved as written climbed from 40% to 60%.
Baseline against month 3.
Month 1 against month 3 as prompts refined.
Worth about $4,250 at $40 an hour.
Month by month
| Measure | Before | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|
| Drafts approved as written (of 2,550) | n/a | 40% | 50% | 60% |
| Emails handled by hand | 3,000 | 1,980 | 1,725 | 1,470 |
| Hours on email | 250 | 186 | 165 | 144 |
| Hours saved | none | 64 | 85 | 106 |
| Value of time saved | none | $2,550 | $3,400 | $4,250 |
handed back to the team over 90 days, worth about $10,200 in staff time. More than six working weeks for one person.