OpenAI launched Astra for Law on September 17, along with a collection of plugins that connect ChatGPT to legal systems. One of those plugins lets an attorney search iManage, review documents, organize content, and save new work back to the correct matter without leaving ChatGPT Enterprise.

iManage says its existing permissions, ethical walls, audit trails, and version history remain in place. That is exactly what a law firm should want from an AI connector.

It also creates a less comfortable question: are those permissions right today?

The connector does not repair a stale security group. It does not know that an attorney changed practice groups, a paralegal left a matter, or a former client required access restrictions that someone never configured. If a user can reach information they should not see in iManage, the AI may be able to retrieve it too.

The AI inherits the firm’s permission decisions, including the old and incorrect ones.

What “preserves permissions” means

A well-designed connector should not give every AI user access to every document. It should ask the source system what the signed-in person is allowed to see and enforce the same restrictions through the AI interface.

For example, an attorney who cannot open a restricted matter in iManage should not be able to find or summarize it through ChatGPT. A user who can read a document but cannot edit it should not gain write access through a plugin.

That model is far safer than copying an entire document library into a separate AI database with weaker controls. It also means the source system remains responsible for deciding who gets access.

This is where old permission problems become AI problems.

An employee may have broad access without realizing it. Before AI, that person would have needed to know where to look, browse into the wrong workspace, and open the file. An AI assistant can search across authorized content in seconds. A normal question such as “find similar settlement agreements” could surface a document from a matter the user was never meant to see.

The AI did not defeat the ethical wall. It followed the permissions it received.

Permission drift is normal

Law firm permissions rarely become messy because one person makes an obviously reckless decision. They drift a little at a time.

Someone joins a matter and gets added to a group. The matter closes, but nobody removes the access. An attorney moves between teams and keeps old permissions. A lateral hire receives access based on a template that does not account for a conflict. A temporary exception becomes permanent.

Shared folders and Microsoft 365 groups create another layer. Matter documents may live in iManage, while related email, meeting notes, exports, or drafts sit in Teams, SharePoint, or OneDrive. The ethical wall may be correct in one system and absent in another.

This issue predates AI. We covered the Microsoft 365 version in our Copilot data-readiness guide. Legal AI plugins make the same problem more urgent inside document management and research workflows because they can bring several systems into one conversational interface.

OpenAI announced 26 vendor plugins with Astra for Law, including connections involving iManage, Thomson Reuters, and Intapp. Each connector has its own permissions, actions, and audit capabilities. Testing one does not prove the others are safe.

Read access is only part of the review

The new iManage plugin can do more than search. According to iManage, a user can organize documents and save new or revised work back into a matter while preserving version history.

That can remove a lot of copying, downloading, and refiling. It also means the firm must test what the connector can change.

Can it save a document to the wrong matter? Can it create a new version of a restricted document? Does the audit log clearly identify the user and the action taken through ChatGPT? What happens when a user’s access changes during an active conversation? Can administrators disable write-back while leaving search available?

The answers may differ by product and configuration. Marketing language about governed access is a starting point, not a completed risk review.

Test the wall instead of trusting the diagram

Before enabling a legal AI connector, choose test users with deliberately different access. Include someone assigned to a restricted matter and someone who is blocked from it.

Then test ordinary prompts beyond direct requests for the restricted client’s name. Ask for related agreements, prior work from the same industry, recent correspondence, and documents that match a legal issue. Those searches may find content that a simple matter-name test misses.

The blocked user should not receive the document, a summary, a revealing title, or enough metadata to infer that the matter exists. The authorized user should receive the expected content without gaining access to unrelated matters.

If the connector can write back to the document system, test that separately. Confirm where drafts land, which version is preserved, and what appears in the audit history. Revoke a test user’s access and verify that the change reaches the AI connector promptly.

Save the results. A dated test record is more useful than a screenshot of a configuration page if a client, insurer, or managing partner later asks how the firm confirmed its ethical walls worked through AI.

Questions to answer before connecting

The firm should be able to answer these questions before turning on a plugin:

Which repositories and matter workspaces can the connector reach?
Does it use each person’s existing access, or a shared service account?
Which actions are read-only, and which can create, change, move, or delete content?
How quickly do permission changes and terminations take effect?
What activity appears in the source system’s audit log?
Who reviews access exceptions and stale group memberships?
How will the firm retest after a plugin, model, or permission structure changes?

This does not require perfect permissions across every system before the firm can experiment with AI. It does require knowing where the sensitive boundaries are, fixing obvious drift, and proving that the connector honors those boundaries.

Legal AI vendors are putting serious effort into permissions and ethical walls. That is good news. Their controls can only enforce the access model the firm gives them.

Before asking what an AI assistant can do with your documents, confirm which documents each person can reach without it. The connector will inherit that answer.

If your firm is evaluating legal AI connectors, Artech can help audit the underlying permissions and test whether ethical walls still hold through the AI interface.