In one sentence
The canvas editor is the surface on which a draft is actually written: the AI works on the selected paragraph instead of on the whole document, its suggestions arrive as genuine tracked changes — and when you bring the client in, their role is enforced in the database rather than hidden in the interface.
Why a chat window is not a writing tool
A chat delivers text in one piece. For a first version that will do; for the second it is the wrong place. Legal writing consists of many small interventions: this sentence sharper, a condition missing there, a citation belongs here. Anyone who has the whole document regenerated every time loses precisely what matters — control over what has changed. The canvas turns that around: you select a paragraph, and the AI works on that selection.
16 actions, four groups
- Rephrase — revise, shorten, expand, simplify.
- Tone — more formal, more professional, more assertive, more diplomatic.
- Structure — the outlining interventions on the selection.
- Legal — more precise, check risks, add a condition, add citations.
The fourth group is the actual justification for the tool: “check risks” on a single clause is something other than a risk analysis across forty pages — a narrow excerpt, a concrete question, a verifiable answer.
Changes arrive as changes
The result of an AI action does not replace your text. It appears as genuine tracked changes — every insertion and every deletion with author and date. You decide on each suggestion individually; “accept all” and “reject all” exist as separate commands for when a round is useful as a whole or useless as a whole.
The difference sounds technical, but it is one of responsibility: a draft you release is yours. A tool that cannot show which five words it changed makes that impossible.
Three levels — and the server holds them, not the screen
If the client is to have access to the draft, you release it at exactly one of three levels:
- Comment only — read and annotate, no intervention in the text.
- Suggest changes — write, but exclusively as a tracked change.
- Edit freely — full editing.
What matters is where these levels take effect: laid down by a database constraint and checked on the server when saving. “Comment only” rejects a text change; “suggest changes” rejects any change not marked as tracked. A manipulated client cannot get past this, because the check does not sit in the interface. That is the difference between a rule and a greyed-out button — a greyed-out button is a request.
Internal comments cannot seep through
The second concern with any release is the internal note that goes along by accident. That, too, is not solved as a filter but structurally: the view released to clients has no read access to the internal columns at all. There is nothing to filter out because nothing arrives — and a forgotten filter can consequently expose nothing either.
Where the text comes from
Every AI answer in the canvas carries the source line that we recently introduced for all AI surfaces: what the answer was built from, and what no longer fitted into it. A suggestion from a well-developed case file therefore looks different from one drawn only from the selected paragraph.
What the canvas is not
- It does not work by itself. No action starts without your selection and your click.
- It decides nothing. “Check risks”, too, delivers a note on a paragraph, not an assessment of your mandate.
- It does not relieve you of the review. Releasing remains a decision for the lawyer — which is precisely why every change is made visible instead of being worked in silently.
→ Canvas editor · Pleadings · Client portal · Firm DMS
