Grella
Back to Guides & Insights

Legal AI and Client Confidentiality: Questions Before Uploading

A practical way to check data use, retention, access and deletion before approving a legal AI workflow. Includes a vendor questionnaire.

Editorial illustration of a document passing through separate, clearly defined boundaries.

Before putting a client document into a legal AI tool, establish what the firm is permitting the provider to do with it. A useful answer identifies the data, the service, the processing route and the conditions of use.

“We do not train on your data” answers one question. It does not establish how long documents are stored, who can access them, where other providers process them or what deletion removes.

This article gives you a practical set of questions to ask. It does not approve a particular tool or client upload. Apply your firm's requirements, the relevant professional rules and the terms governing the matter.

Start with the proposed use

Describe one workflow before reviewing a general security pack. For example: a lawyer uploads a document packet, asks factual questions, checks cited passages and saves a draft for review.

Then identify what would leave the firm's control. That may include the source files, extracted text, prompts, answers, document titles, user details and diagnostic information. An integration may add another route. The provider's answer should cover each step you intend to use.

Record the categories of information involved. Public documents, personal information, commercially sensitive material and privileged communications can require different decisions. Removing a name does not necessarily remove identifying details or the substance of confidential advice.

The ABA's July 2024 opinion addresses confidentiality and provider diligence under the US Model Rules. Australian OAIC guidance addresses privacy obligations when commercial AI products handle personal information. Those are different frameworks; neither is a worldwide permission to upload client material. 1 2

Ask what the training statement covers

Ask whether the statement applies to your exact plan, account type and features. Does it cover uploaded documents, prompts, outputs, feedback and support attachments? Does it cover every model or other provider involved?

If there is an opt-out, request its default setting, who can change it and how the firm can verify it. If an exception exists for abuse monitoring, product improvement or feedback, ask what information it captures and for how long.

Keep training and storage as separate questions. A service can retain information without using it to train a model. A short retention period can still involve access or processing that the firm has not approved.

Follow the information through the service

Ask for a current description of the processing route. It should identify the provider operating the product and any services used for hosting, document conversion, OCR, model inference, search, embeddings, support and analytics where they receive the relevant information.

For each route, request the purpose, categories of data, processing location and applicable terms. Distinguish the place where the main file is stored from places where a copy or extracted text is processed. Ask how the firm learns about a new provider or a material change.

The Law Society of Ireland's December 2025 guidance cautions against treating an enterprise label as a compliance guarantee. Request the terms and evidence that apply to the actual service. 3

Make retention and deletion specific

“Delete the matter” may describe a user action without describing every remaining copy. Ask separately about source files, extracted text, prompts, outputs, search indexes, backups, logs and material supplied to support.

For each category, record the retention period or trigger, the deletion method, the expected completion time and any exception. If a backup expires on a schedule rather than being edited immediately, record that limit. Ask what evidence the firm can obtain after deletion.

Also ask what happens when an account closes, a subscription ends or the provider changes a service. Export and deletion are different requirements. The firm may need an intelligible copy of its work before access ends.

Check who can access the information

Ask who inside the firm can reach a matter and its derived work. Can the firm's administrator restrict access? What happens when a lawyer leaves or changes teams? How does an integration inherit or bypass those restrictions?

Separately, ask when provider staff or other service providers may access content. Request the authorization process, the purpose, any time limit and the audit record. A general statement that access is restricted needs an explanation of the restriction.

Ask for evidence proportionate to the proposed use. That may include current contractual terms, a data-processing agreement, a subprocessor list, configuration documentation or independently assessed controls with their scope and date. A certification for one part of a business does not automatically describe every feature you plan to use.

Turn answers into a bounded decision

Do not reduce the questionnaire to a score that averages away an unanswered requirement. Mark each response as evidenced, unresolved or outside the proposed use. An unanswered required condition remains a reason to hold that use.

Record a decision with a named owner: hold, approve a limited use, or approve the described workflow under stated conditions. A limited approval should name the permitted information and features, prohibited uses, review date and changes that require a fresh decision.

Information

An illustrative decision note, using no client information

Permitted use: a demonstration using the supplied fictional records. No client documents, client-derived prompts or client-identifying details. Owner: the firm's designated reviewer. Reconsider before changing the data categories, service plan or integrations.

This is a way to document a boundary. It is not a conclusion that fictional data makes every service or environment suitable.

Once the data-use questions are settled, compare the tool against the work your firm needs to do.

The working kit

The vendor questionnaire separates the proposed use, provider answers, supporting evidence and the firm’s decision. It includes twenty questions, an evidence register, open items and a decision record. Use the editable Word document to collect responses and record the firm’s review. The PDF is a reading and print copy.

The kit is neutral and tool independent. It is not a completed security assessment of Grella or any other provider. Blank fields are deliberate: the provider and the firm's reviewer must supply and evaluate the answers.

What this means for evaluating Grella

Do not infer approval to upload confidential or privileged material from this article, a demonstration or Grella's factual-workflow features. That approval requires separate security, contractual and firm review. Ask about the current permitted data scope before arranging an evaluation.

For the broader buying decision, use Choosing Legal AI. If a trial is permitted, the one-matter pilot guide helps define the task and review method. Read the security page alongside current provider evidence; marketing copy is not a substitute for that evidence.

Sources and scope

Sources checked on 13 October 2026. The linked guidance has a jurisdiction and publication date. The questionnaire is an original practical aid, not a regulator's form or a determination of legal compliance.

ABA Formal Opinion 512, 29 July 2024

OAIC: Privacy and the use of commercially available AI products, updated 17 January 2025

Law Society of Ireland: Generative AI guidelines, December 2025

All guides & insightsBack to top