AI assistant
A conversation that knows the instance you are on, behind four permissions that all start off.
It knows where you are
The assistant reads the instance you have open: the record in front of you, the table it belongs to, the update set you are in. You do not paste context into it.
The conversation exists from the moment you send the first message rather than when a reply comes back, so nothing is lost if you navigate away mid-answer. Open the same conversation in a second tab and both stay in step.
Four permissions, all off
Nothing that touches your instance is granted by default:
| Permission | What it grants |
|---|---|
aiAutoApply | Applies a proposed change without you pressing Apply, and skips the update set check with it |
aiConsoleRead | Records uncaught exceptions and rejected promises, and lets the model read them |
aiRunScript | Runs JavaScript in the page realm: g_form, GlideAjax, the DOM |
aiBgScript | Runs a server-side script with your own rights, through Scripts - Background |
They are read strictly: a preference has to say true to count. Every other setting in the
widget defaults to on, and a permission written in that looser style would end up granted on
every machine that had never heard of it. A profile older than a permission has not refused it
either, so it is left alone rather than overwritten with a refusal nobody gave.
The check runs once, before a tool is dispatched, rather than inside each tool. A capability that decides for itself whether it is allowed becomes allowed the day somebody edits it. When a tool is refused, the model is told which switch to ask you for and not to retry, so a denied tool costs one pass instead of three.
Granting takes effect at once
Save a permission and it applies in the tab you already have open. No reload.
One honest exception: the error tap only records from the moment it starts. Turning on console reading does not recover what broke before you consented, and the report says so rather than letting silence read as a healthy page.
Proposals
Larger changes come back as a deck of proposals rather than one wall of text, and each is
applied on its own. With aiAutoApply off, which is the default, every one of them waits for
you.
Attachments
The paperclip, a drop on the panel or a paste in the box adds up to four files to a question:
- images (PNG, JPEG, WebP, GIF) and PDFs;
- text files, from a log or a CSV to a script or an XML export;
- Word, Excel and PowerPoint files (
.docx,.xlsx,.pptx, and their macro-enabled versions).
An Office file is read in your browser, not on a server. What reaches the assistant is its text: paragraphs and headings, lists, tables as rows, every sheet with its dates, slides in order. Pictures, charts and layout are left out. Because it becomes text before it leaves the machine, it goes through the same filter as any other text you send, which takes out credentials and tokens, email addresses, phone, card and bank numbers, tax codes and people's names.
The old binary formats (.doc, .xls, .ppt) cannot be read; the refusal says to save the file in
the newer format. A file is limited to 4 MB, or 25 MB for an Office file, and very long text is cut
with a note saying so.
Dictation
The microphone button in the panel records, sends once, and throws the audio away. It is not kept in storage, not kept in the conversation, and not kept in memory. What survives is the text, and it lands in the box so you can correct it before sending.
The limits are two minutes and 8 MB. The states are idle, recording, transcribing, in that order with no shortcuts.