The problem
A retail-services business in South East Queensland — around twenty staff, roughly A$5M turnover — ran its entire customer relationship through one shared inbox. Orders, delivery questions, warranty claims, returns, account queries, the occasional complaint: all of it landed in the same place, and three people took turns clearing it around their other work.
The volume wasn't the whole problem. Most of what arrived was repetitive — "where's my order", "do you have this in stock", "how do I return this" — and each one still had to be read, understood, checked against an order or account, and answered by hand. First-response time had crept out to most of a working day. Customers chased. Some enquiries were answered twice by different people; some slipped through entirely.
The owner didn't want a chatbot bolted to the website, and didn't want a call centre. The question was narrower and more practical: could the repetitive drafting be taken off the team's plate without a customer ever receiving something the business hadn't chosen to send?
What we built
We started with a short readiness audit — two weeks, reviewing a few months of inbox history to see what actually came in and how it was answered. The pattern was clear: a handful of enquiry types made up most of the volume, and the good replies already followed recognisable shapes.
The pilot ran for six weeks and put a single flow behind the shared inbox:
- Classification. Every inbound message is sorted into an enquiry type — order status, stock, returns, warranty, billing, general, or "unclear".
- A drafted reply, in the business's voice. For the common types, the flow drafts a reply against the business's own tone and standard wording, pulling the relevant order or account detail so the draft is specific rather than generic.
- A review step, always. Every draft lands in a queue for a team member to read, adjust, and send. Nothing goes to a customer automatically.
- Escalation. Anything sensitive — a complaint, a refund dispute, legal or health-adjacent wording, an angry tone, or a low-confidence classification — is flagged and routed straight to a named person with no draft attached, so it gets a human from the first word.
How it works
A new email arrives in the shared inbox. The flow reads it, classifies it, and decides whether it's routine enough to draft or sensitive enough to escalate.
For routine enquiries, it retrieves the specific facts the reply needs — the order record, the returns policy, the stock position — and drafts a response in the business's established tone. The draft appears in the review queue with the original message, the classification, and a confidence score beside it. A team member skims it, edits if needed, and sends. Most take a few seconds; the hard ones still get the time they deserve.
For sensitive enquiries, there's no draft. The message is tagged and pushed to the person who should handle it, so the business's judgement — not a model — writes anything that carries risk or emotion.
Brand tone is controlled deliberately: the drafting is anchored to the business's own past replies and a short style guide, and reviewers correct any drift, which keeps later drafts closer to the mark.
The results
Measured across the pilot and the first weeks in production:
| Before | After (pilot) | |
|---|---|---|
| First-response time | ~7 hours | ~9 minutes (median) |
| Enquiries auto-drafted (human-approved) | 0% | ~65% |
| Team time on the inbox | ~30 hrs/week | ~12 hrs/week |
| Duplicated or missed replies | Regular | Rare |
| Cost per enquiry (model + infra) | n/a | ~A$0.04 |
The ~18 hours a week the team stopped spending on repetitive drafting is worth an estimated A$50k–65k a year in recovered time at a loaded rate — hours that went back into orders, fulfilment, and the enquiries that genuinely needed a person. The speed change was the one customers noticed: a reply within minutes instead of the next day, in the same tone as always.
How the team owns it
The pilot was built to be run by the business, not by us. The three people who already worked the inbox are the reviewers — the flow drafts, they decide. That review step is the point, not a temporary safeguard: a person reads and sends every reply, and every escalated message reaches a human with no draft in the way.
We handed over a short runbook covering how to adjust the standard wording, how to add a new enquiry type, and how to change what counts as "sensitive" — because a business's sense of what needs a human hand shifts over time, and the team can move that line themselves without calling us. We left the escalation rules conservative on purpose; it's easier to loosen them once the team trusts the drafts than to recover from one reply that should have had a person on it.
Nothing about the customer relationship changed hands. The business still answers its own customers, in its own voice. It just stopped writing the same reply for the fortieth time to get there.
If a shared inbox is quietly eating your team's week, we're happy to take a look.