Home / Documentation / AI Assist

AI Assist

These guides cover the full workspace included in the Pro trial. Compare Free, Starter and Pro for ongoing access; Assist and iPad Companion require Pro.

AI Assist has two views. Ask is where you choose the work, sources and provider. Activity keeps the progress and result of each run. Dashboard is a separate workspace.

Ask

1. Write an instruction or choose a task preset, such as Summarise documents, Compare documents, Chronology narrative or Find passages on…. Edit the instruction to say what you need.
2. Choose Add documents or scope, select your sources and choose Add to request. Sources can be individual documents or Whole brief, Pleadings, Chronology or Hearing bundle.
3. Choose the provider and, where offered, a model. Read the source check above the request. If selected documents are unavailable, follow the stated remedy or explicitly choose Exclude unavailable and continue when offered.
4. Choose Preview & run for a cloud request and review the send details. A fully local Ollama request uses Run locally. Follow progress in Activity.

Sources determine how the work runs. Individually selected documents can be sent as text for tasks such as summarising or comparing them. Whole brief, Pleadings and Hearing bundle are searched through the read-only connector with Claude or ChatGPT; their documents are not pasted wholesale into the request. These scopes require AI search permission. Ollama has no connector, so choose individual documents for its supported tasks. Chronology sends the structured accepted events with the request.

Issues memo accepts up to three individually selected documents as text, or uses the connector for a document scope or an AI-search selection. It does not accept Chronology; use Chronology narrative for that work. Changing the Issues memo instruction does not remove its document limit.

Choose the provider

• Claude — runs through Claude Code installed and signed in on your computer. Some requests can instead be handed to Claude Code or Claude desktop.
• ChatGPT — runs through Codex installed and signed in on your computer.
• Ollama — can run a model on this computer, at a remote endpoint or on Ollama Cloud. Lexware labels the selected arrangement before the run.

Lexware does not receive or relay these requests, and it does not hold the provider credentials. Claude and ChatGPT use your own sign-in. A local Ollama run stays on this computer; a remote Ollama endpoint or Ollama Cloud does not.

Before anything leaves

Cloud AI is off by default for each matter. In Ask, open Details on the privacy panel, then turn on Cloud AI for this matter to use Claude, ChatGPT, a remote Ollama endpoint or Ollama Cloud. Identity masking and AI search are separate controls there. A model running locally in Ollama does not need cloud permission.

Before a cloud run starts, Lexware shows the provider, model, sources, masking posture and whether AI search will be available. Nothing is sent until you confirm.

Source text sent with the request

For an Ask request that sends selected source text, Identity masking replaces detected identifiers in the sources and instruction with neutral tags and shows the masked payload before sending it. The reply is restored locally. If masking is off, Lexware asks you to confirm that the selected material and instruction will be sent as written. Masking preparation may appear in Activity first; return to its review action when it is ready.

AI search

Turn on AI search in the privacy details to allow Claude or ChatGPT to search and read matter documents during a Lexware run. It also requires Cloud AI permission. The connector is read-only and access ends when the run finishes or is cancelled.

A search over individually selected documents, Pleadings or Hearing bundle is limited to the documents resolved for that run. Whole brief is a live grant over visible matter documents: it can also reach pending Intray uploads, email-only attachments and documents indexed while the run is in progress. Other unmasked tasks given Matter tools can have that same broader access. Read the confirmation's scope; an initial list of source documents does not always limit what the tools can retrieve.

The search instruction and retrieved passages travel as written and are not covered by Identity masking. Passages retrieved during the run cannot appear in an exact payload preview beforehand, so the confirmation shows your instruction and search scope instead. Tasks that send a masked source-text payload run without Matter tools. Check the result's coverage and citations before treating it as a review of the whole scope.

Activity

Activity records current and earlier runs, including status, provider, model and whether transmission was confirmed. Its actions let you review a prepared payload, cancel an active run or open a result. A failed Ask run with its instruction retained can be revised and run again. View kept work opens available work from an interrupted Claude run; that work is kept for 30 days.

A text result can be saved through Save to notes…. Cited passages from an AI-search result can be added to Scratchpad. Returned files can be previewed before a separate import: PDF, Word, Markdown and text files become documents, while an Excel chronology stages candidates in Review. Activity also links to the content-free transmission log.

Claude hand-offs

The Matter workspace section in Ask maintains a local hand-off folder. Update workspace prepares it and Reveal folder opens it on your computer. Open in Claude Code hands over the workspace; Claude desktop hands over only the typed prompt. Both hand-offs require matter consent and a payload review. The Claude Code workspace contains unmasked matter material, even when Identity masking is on for other Assist requests. Review returned Inbox files from Activity before importing them.