Packaging Products - An AI Bot That Sorts the Inbox and Drafts the Replies

Client Overview

Packaging Products makes custom boxes, printed boxes, labels and more to order, and promotes that production capability to businesses around New Zealand. Enquiries for production work arrive in one shared inbox, and someone on the team reads and answers every one of them.

The Challenge

Reading the production inbox was taking a large part of a staff member's day, and the volume was rising. A growing share of it was junk, sales pitches or vague, mass-sent enquiries, often AI written, mixed in with genuine requests for quotes. The business expected volume to keep climbing as it pushed its production work harder.

The owners were clear on one thing: nothing should go out to a customer, and nothing should be deleted, without a person seeing it first.

The Solution

Automate AI built an inbox triage and draft-reply bot that runs on AWS and reads the mailbox through Microsoft Graph once an hour. Every new message is classified on intent, not on whether it reads like AI: a buyer who used AI to write a clear brief with sizes, quantities and a deadline is still a real enquiry. High-confidence junk moves to a separate folder. Real enquiries get a draft reply, threaded under the original, written with Claude models on Amazon Bedrock.

Each draft is built from the business's own material: its websites, product documents, standing facts like usual lead times, and similar replies the team has sent before. If a price, size or lead time is not in those sources, the draft does not guess. It leaves a clear marker for the reviewer to fill in, and a short internal note lists the category, the sources used and any open questions.

The safety rules are enforced in the system itself, not left to the AI model to follow:

How It Works Day to Day

Staff keep working in Outlook. Colour-coded categories show what the bot thought of each message, junk sits in its own folder in case anything needs rescuing, and drafts wait in the Drafts folder. Dragging a message out of junk, or tagging it as wrong, records a correction. A colleague's forwarded customer email is read as the customer's, so the draft goes to the customer, not back to the colleague.

Every morning a digest email summarises what came in, what was drafted, what was sent with or without edits, any new rules and the day's AI spend. The same codebase runs one stack per client, so a second mailbox can be added through configuration rather than new code.

Key Takeaways