Introduction to AI assistant access

​

What is AI assistant access?

AI assistant access lets a tool such as Claude query Nobly Insight directly, using the Model Context Protocol (MCP) — an open standard for connecting assistants to the systems they need to answer with. Instead of pasting documents into a chat, the assistant asks Insight and gets the answer from your own content. The assistant connects as you. It signs in with your account, and every request it makes carries your identity — so it sees exactly what you would see, and nothing more.

​

What has to be true before an assistant sees anything

Three things stack, and all three must allow a request:

​

What each capability exposes

Grant these deliberately: they differ sharply in how much they reveal. Decide with the data in mind, not the feature name. Catalog is the mildest and is usually what makes the rest useful — an assistant cannot build a sensible search without knowing which document types and keyword types exist.

​

What each capability can change

Reading and changing are separate permissions. A group can be given every capability above and still change nothing; these three are what an assistant needs before it can alter anything at all. Each is a permission of its own, with one exception: MCP Administration grants all three at once, together with every read permission and every configuration permission — see Granting access. A reindex is the change to understand before you grant it. By default it replaces all of a document’s keyword values with the ones the assistant supplies. When the assistant carries the existing values over instead, the ones the new type has no place for are still removed, and a keyword group the assistant supplies replaces that whole group. The document history records that the document was reindexed, but not the values the reindex removed. The assistant’s answer also shows the document as it stands afterwards — its new type, name and dates — provided the person may view it, even in a group without MCP: View Documents. Grant these more carefully than the read ones. An assistant works at machine speed and does not pause to wonder whether it understood you, so a mistaken instruction becomes a mistaken change quickly — and a keyword value it overwrites is gone the same way it would be if a person had overwritten it. Its changes also set off the same automation a person’s change does: scripts and workflows that react to a keyword change or a reindex run for an assistant’s change too.

​

Changing configuration

An assistant can also work with your configuration — the same operations as the admin assistant, in the same areas and with the same approvals. For each admin area the assistant gets three tools: one that lists the area’s operations and what the environment holds, one that previews a change without making it, and one that makes changes. A further set of tools reads configuration — looking an item up, finding where it is used, the configuration audit log, and what became of the changes it proposed. These tools need more than an MCP permission:

• AI features are available in your environment.
• The admin assistant is switched on for the environment. When it is off, the tools are gone.
• The person holds Use AI Admin Assistant, as well as the configuration permissions for the areas — see Granting access.
• Each change needs the permission its admin page requires, exactly as in the product.
• An operation the environment has switched Off for the assistant does not exist for an outside assistant either.

​

How a change is approved

A change waits for approval unless the environment allows auto-apply for it and the person has switched that on for themselves — the same rule as in the admin assistant. One request from the assistant is one batch of up to 25 changes, approved together. The changes apply in order, and anything that deletes, removes or takes something away applies after the rest. Each waiting change waits on the one before it, whether or not the two are related, so rejecting one change of an outside assistant’s batch cancels the waiting changes after it. If one change in the request is refused, the request’s waiting changes are withdrawn. A change your own auto-apply already made stays made, and the assistant is told which changes those are, so it can tell you. How you approve depends on the assistant you use and on what the batch holds:

• An assistant that can show a form opens its own approval dialog only for a batch the dialog can show in full — every value whole, no multi-line content such as a script, at most three rows per change, and no “…more” — and only when the batch holds no change of a kind that starts at Always ask, such as a delete, and nothing that changes code or access: scripts, workflow definitions, templates, or users and access. The dialog lists the changes, with a link to the batch on the Approvals page for the full details. Accept approves the batch, Decline rejects it, and closing the dialog decides nothing — the changes keep waiting.
• Every other batch, and any other assistant, gets the link to the batch on the Approvals page, where you decide it next to the full details.

Either way, the batch waits on the Approvals page, marked MCP, until it is decided or expires after 7 days. The assistant reads what became of a change from Nobly Insight; it does not need the dialog to find out.

​

What is recorded

Assistant activity is not a side channel. It lands in the same places you already review:

• Document retrieval is written to the document history, under the person whose account the assistant is using — the same record a download through the web client produces.
• Searches are recorded in the search audit trail, marked as having come from an assistant, so you can tell them apart from searches run in the product. Searches of case data are recorded there too, under a record of their own, so you can review what an assistant looked for in Caseflow without reading through its document searches — and the other way round.
• Configuration changes are recorded in the configuration audit log as AI on behalf of the person, like changes made with the admin assistant.
• Changes to documents and case objects are recorded exactly as the same change made in the product is: a keyword edit and a reindex in the document’s history, a case object’s edit in its own history, each under the person whose account the assistant is using. Unlike searches and configuration changes, these are not marked as coming from an assistant — the history reads as if the person had made the change themselves.

Search auditing for assistants is always on. It does not follow the setting that controls auditing for the search page, so a tenant that has not switched that on still gets a record of what its assistants searched for.

​

What assistants cannot do

An assistant cannot delete a document, a revision or a case object, and no permission grants it. It cannot add a document to the archive, and it cannot change the file a document holds: what it can change is how a document is indexed and classified, not what it is a copy of. Clearing keyword values, and the values a reindex removes, are as close as it comes to deleting. It also cannot exceed the person it acts as. Every change it makes is one they could have made themselves, and it needs both the assistant permission and their own rights on the document type or Caseflow class to make it.

​

Where to read next