What should approval cover before AI touches client documents?
Consider this hypothetical situation: your document system offers a new AI connection. A partner wants summaries ready before client meetings, and your administrator needs to know whether enabling it also permits changes or sharing. As of November 2026, the operating question is specific: what must your firm approve before an AI tool connects to client documents?
We recommend approving a defined job, data boundary and set of permitted actions together. Record who can use the connection, which documents it can reach, what it can do with them, and who checks the result. Approval should describe the workflow staff will actually use, including where the work stops for human review.
This article uses dated public reporting, vendor guidance and vendor research. It includes a clearly labeled constructed example. Those sources may help frame an approval review, but they do not establish how any particular firm's system behaves, whether a product satisfies applicable professional obligations, or whether a proposed connection is safe.
Why a familiar product still needs a specific review
In a November 3, 2026, LawNext interview about NetDocuments' redesigned platform, Chief Product Officer Dan Hauck discussed document-system connections to other AI platforms and a beta connector with Legora. He also described a proprietary legal context graph: infrastructure that maps relationships among matters, documents, communications and people.
Those are reported product descriptions and executive claims, not independent findings about access controls or performance. The supplied excerpt does not establish which permissions a particular connection receives or how those permissions are enforced.
A connection inside software the firm already uses may introduce a different route to documents or a new ability to act on them. Familiarity with the original product therefore may not answer the approval questions for the added workflow.
We recommend reviewing the connection separately from the existing software approval. Ask the vendor to identify the systems involved, the information exchanged, the accounts used and the actions available. Keep what works. Change what doesn't. A review can support either retaining the current setup or considering a scoped replacement when the firm's requirements warrant it.
Ask for written answers that match the proposed job
Counsel Cyber's August 28, 2026, vendor due diligence guidance recommends defining the use case before evaluating products, getting answers in writing and comparing them with the contract. It also recommends a limited trial using non-sensitive or synthetic material, with defined participants and success measures.
This is vendor guidance, not an independent assessment or a statement of the rules governing your firm. Its reference to ABA Formal Opinion 512 does not substitute for reviewing that authority or relevant state guidance. No legal obligation is inferred here from that secondary summary.
We recommend writing a one-paragraph job description before the vendor meeting. For example: prepare a draft meeting summary from documents selected by an authorized employee, place it in a review location and stop. Specify whether sending messages, updating matter records or retrieving additional documents falls outside approval.
Ask for a written account of the data path. Identify where submitted material goes, which providers receive it, how long it remains available and how deletion works. Compare the response with the agreement and the configuration you intend to use. Record unresolved differences rather than treating a sales answer as a contractual commitment.
Check whose access the connection uses. Ask whether it operates as the individual employee, through a shared account or through a separate service account. Request an explanation of how matter restrictions apply to searches, retrieved documents and generated output.
Separate permission to read from permission to act
Consider a hypothetical document workflow in which an assistant retrieves files, creates drafts or changes records. For that example, we recommend distinguishing preparation for human review from actions that affect a client or the firm's records.
We recommend listing each permitted action separately. Distinguish searching, opening documents, creating drafts, changing originals, sending communications and deleting records. Give each action an owner and an approval condition. Our related explanation of what to verify before an AI agent handles client work provides a companion framework for that permission review.
Consider beginning with retrieval and draft creation only. Keep external sending and changes to original records disabled during the initial trial. This narrower scope could make mistakes easier to inspect before the firm considers additional permissions; it does not establish that the remaining workflow is secure.
Check the stopping point explicitly. Document what happens when the tool finds conflicting records, cannot open a document, encounters an unfamiliar instruction or lacks enough information. Name the person who receives that exception and decides whether work can resume.
Two technical developments explain why configuration matters
In the supplied release note for LLM 0.37.2, Simon Willison describes fixes for two configurations that could accidentally send a configured OpenAI API key to another provider's compatible endpoint.
The note concerns that software and those configurations. It does not establish that a law firm's tools exposed credentials, and it is not evidence of a general failure across legal AI products.
We recommend asking the technical reviewer to verify credential routing. Confirm which credential belongs to each provider and what happens if a required credential is missing. Request evidence that a missing credential stops the request rather than silently selecting another one. Record the software version and configuration tested.
Microsoft's November 10, 2026, Lazarus research account describes a different concern. Its researchers claim that modifying a small number of model components could cause an otherwise normal agent to execute malicious commands after a hidden trigger. The account says the attack requires altering model weights.
This is Microsoft vendor research, not a reproduced result here or evidence that a firm's model is compromised. Microsoft's descriptions of its own defenses are vendor claims. The excerpt also states that no scan can prove beyond doubt that a model is safe.
Ask who supplies the model and how its origin and integrity are checked. If the workflow permits tool or terminal access, have the technical reviewer document those capabilities and their limits. A successful demonstration may show that a task works under the demonstrated conditions; it cannot establish the absence of a hidden trigger.
Practical example: trial a probate meeting summary
Constructed operational example: A small probate firm wants an assistant to prepare meeting summaries from selected documents. This scenario is a recommended test design, not a client account, measured outcome or description of an available AlphaForge configuration.
- Create a synthetic matter. Use invented names and facts. Include two similarly named matters, a superseded draft and a document that the test user cannot access. Keep real client information out of the trial.
- Define the expected result. Specify the documents the assistant may use, the facts the summary should contain and the location where the draft belongs. Require source references that the reviewer can inspect.
- Test boundaries alongside accuracy. Request information from the restricted document and ask the assistant to email the summary. Check whether the underlying system enforces the approved limits, and record attempted actions separately from completed ones.
- Review the full work. Compare the draft against the permitted source documents. Record missing facts, unsupported statements, wrong-matter references and the time needed to correct the output.
- Record the decision. Name the reviewer, accepted limitations, unresolved issues and approved configuration. Keep approval pending where a required boundary cannot be demonstrated.
What should the firm measure next?
We recommend tracking review effort alongside output accuracy. Count drafts reviewed, material corrections, unsupported statements, boundary-test failures and unresolved exceptions. Measure the time spent checking and correcting work against the firm's existing process. A polished draft alone does not show that the workflow saves effort.
Review approval after material changes. Revisit the record when a connection gains new actions, the provider changes, access settings change or the vendor alters relevant terms. Counsel Cyber's guidance also recommends reviewing approved tools at least annually and when terms or architecture change.
Keep the final approval record short enough to use. It should identify the job, permitted data, allowed actions, human reviewer, tested version and reasons to pause. That record could give the managing partner a concrete basis for a decision while leaving technical checks and professional-responsibility review with the people qualified to perform them.