Appearance
Overview
Customize AI is where you teach Overdesk how your team writes. Overdesk ships with its own rules for drafting and reporting, and the settings here add your team's guidance on top of them. They never replace the defaults.
Open Customize AI in the sidebar. Owners and admins can change these settings. Agents and viewers can read them: the cards are locked, under the line "An admin changes these settings. You can read them here."
What is on the page
The page has three cards. Each one saves on its own, so a change to one card leaves the others untouched.
| Card | What it shapes | Who reads it |
|---|---|---|
| Workspace context | Facts about your product and how your team works | Drafts, reports and analyses |
| Draft replies | Who a draft writes as, its tone, length and language, your house rules, the review link, whether drafts read pictures and linked pages, and whether mail that is not support is screened out | Every reply draft |
| Reports & analyses | How the AI commentary reads your queue, and what it should never assume | Every report and analysis |
A change applies to the next draft or report Overdesk writes. Drafts already sitting in your help desk stay as they were written.
What is already built in
Before any of your guidance is added, every draft already follows Overdesk's own rules:
- It grounds the reply in your knowledge base and links to your docs.
- It never invents features, fixes, versions, refunds or timelines.
- It answers every question the customer raised, with concrete steps.
- It keeps the customer's email address and phone number, international or North American, out of the model and puts them back in the finished draft.
- It reads what the customer wrote new in each message, leaving out the earlier emails quoted beneath it and their email signature. A forwarded message stays, because it is often the problem the customer is asking about.
- In a long conversation it keeps the customer's first message and the messages where they gave access or login details beside the newest ones, and it is told how many earlier messages it does not see.
- It follows lessons that Overdesk's team approved from your team's corrections to earlier drafts. A lesson steers wording and never licenses an invented fact, and your own instructions win when the two conflict.
Your instructions shape style and substance, but they cannot talk a draft into making something up. A draft follows your instructions unless they conflict with its grounding rules, and it never invents a fact to satisfy one.
In this section
- Tone, length and language: how a draft sounds, how long it runs and which language it uses.
- Product context and instructions: what you sell, who the draft writes as, your house rules and the review link.
- Drafting model: which Claude model writes your drafts, reports and analyses.
- Pictures and links: letting a draft read screenshots, PDFs and the pages a customer links to.
The Draft replies card also holds Only draft when Overdesk finds relevant sources, which leaves a conversation alone when your knowledge base has nothing on it. It has its own page: Only draft when sources are found. It also holds Screen out non-support mail, on by default, which gives marketing, spam, automatic notices and out-of-office replies a note instead of a draft: see Mail that is not support.
Your Anthropic API key is not on this page. It lives in Settings → AI: see Add your Anthropic API key.