AI inbox that drafts the reply and the quote for you to approve

This is a shared inbox that reads an incoming request, looks up the customer and your price list, and prepares a draft reply with a quote attached. A person reads it next to the original message, edits if needed, and approves. Nothing leaves the building without that click.

The longer write-up of the design and its guardrails is in the related post on drafting replies and quotes with human approval. This page is about what the finished tool looks like for a small business.

What does it look like day to day?

You open one screen with a list of pending items. Each row shows the sender, a one-line summary, the category the system guessed (quote request, order status, complaint, other) and how long it has waited. Clicking a row opens three panels side by side:

• the original message, shown as untrusted text
• the draft reply, editable
• the evidence: which customer record, which past orders and which price list lines the draft used

Under the draft sits a line-item table with quantities, unit prices and totals. Those numbers come from your price list through plain code, not from the model. If the draft text mentions an amount that differs from the table, the item is flagged and the approve button asks you to acknowledge it.

How is it built?

The pipeline has a few boring, separate steps. Mail arrives and is stored untouched. A classification step sorts it, and anything unclear goes straight to a human without a draft. A retrieval step, written as normal code, fetches only the records for that sender. The model receives those records and returns structured output: reply text, proposed products and quantities, and the IDs of the records it relied on. Validation runs, the item lands in the queue, and after approval the sending code takes the approved text and checks that it has not changed since.

The model has no ability to send mail, write to your records or search across all customers. That narrow scope is the main protection against a hostile email that says “ignore your instructions”.

Where does it go wrong?

Honest list. Drafts for unusual requests will be weak, so the queue shows a confidence note and a “needs a person” route. Tone drifts unless it is anchored with examples of your own past replies. Customers who write in two languages in one message confuse classification. And a reviewer who approves everything without reading defeats the point, so I log edit rate and time spent per item so you can see whether review is real.

When is off-the-shelf better?

If your emails are mostly simple support questions with no pricing, a helpdesk with a built-in AI drafting feature may be enough, and you should try it first. A custom build earns its place when quotes depend on your own price rules, stock or customer terms that a generic tool cannot see.

What I need from you

• Access to a sample of past requests and the replies you sent, so I can test on real material.
• Your price list in a spreadsheet or export, and the rules that change prices (volume, customer type).
• A decision on who reviews, and what they should never approve unchecked.

Go-live starts with the queue running in shadow mode, drafting next to your normal work, so you compare before relying on it.