Skip to content
All articles
Industry and business

Private AI for law firms: the matter wall has to survive the assistant

A law firm's retrieval index is the fastest way to break confidentiality between matters, so the permission model has to say no before the model sees a document.

BastionUpdated 3 min read

Summarising, comparing, extracting clauses and searching internal knowledge are the repetitive work that consumes junior time in a law firm. The material is client confidential and separated by matter.

Ask where client information goes before you ask about quality

  • Where does the client's information go while the model is processing it?
  • Who else can reach it?
  • How long is it retained, and by whom?

Those answers determine which tools are usable at all, and they narrow the shortlist faster than any benchmark.

Retrieval is where confidentiality breaks

Retrieval over the firm's documents is the feature that makes an assistant useful. It is also a search engine over everything the firm holds, and a search engine does not know about the matter wall.

A lawyer should not receive documents from another matter because the assistant was able to retrieve them. That has to be enforced at retrieval, by the user's own permissions, not by an instruction telling the model to be discreet: a model can be persuaded, and an access control cannot.

Documents from the other side are untrusted input

A firm's index fills up with material written by people outside it: draft contracts from opposing counsel, bundles, correspondence, expert reports. Those documents become model input, and text in a document can behave like an instruction.

CVE-2025-32711, disclosed in June 2025, was described as “a zero-click prompt injection vulnerability in Microsoft 365 Copilot that enabled remote, unauthenticated data exfiltration via a single crafted email”. It scored 9.3 and Microsoft fixed it; The Hacker News reported that “There is no evidence that the shortcoming was exploited maliciously in the wild.”

An assistant that reads inbound documents and can reach the matter store therefore needs its outbound paths and tool permissions constrained deliberately.

Workloads worth running locally

  • Summarising long documents within a single matter
  • Comparing versions of a contract and listing what changed
  • Extracting clauses and obligations into a checkable table
  • Searching the firm's own precedents and know-how
  • Drafting first passes of routine correspondence
  • Answering internal questions about the firm's own procedures

Each of those is verifiable against a source document, which is the property that makes a workload safe to start with.

What a local model does not replace

A local model does not exercise legal judgement: generated analysis still requires review by a qualified professional before it reaches a client or a court.

Where the work is not confidential and the strongest available reasoning matters, a cloud service may be the better tool.

With Bastion

What Bastion changes

Bastion is a private AI system delivered as one sealed box that runs inside your building. One monthly fee covers the hardware, the model, the hardened operating system and support, and nothing your team types leaves the building.

Client material is not sent to an outside AI service: the appliance, at the firm's office or in a data centre, answers on the firm's private network and needs no outbound connection.

The matter wall is still the firm's to build: retrieval permissions, logging and review remain deployment decisions.

Questions this article answers

Is it safe to use AI on confidential client documents?
It depends on where the client's information goes while the model processes it, who else can reach it, and how long it is retained and by whom. Those answers decide which tools a firm can use at all, before model quality is considered.
Can prompt injection expose a law firm's documents?
Text in a document can behave like an instruction. CVE-2025-32711, disclosed in June 2025 and scored 9.3, was a zero-click prompt injection in Microsoft 365 Copilot through a single crafted email; Microsoft fixed it, and The Hacker News reported no evidence of malicious exploitation.
Which legal work suits a local AI model?
Work that can be checked against a source document: summarising within a single matter, comparing contract versions, extracting clauses, searching the firm's precedents and drafting routine correspondence. Generated analysis still needs review by a qualified professional.

On the record

  1. 1
  2. 2
  3. 3