Data Security in Legal AI

Data security in legal AI starts with a decision an in-house lawyer can explain: which work can the team do, with which information, under which controls? Reviewing a public filing and analyzing an internal investigation may require different permissions, terms, and settings. Before signing, trace what happens to the documents and record the conditions for approved use.

GC AI is enterprise legal AI software built for in-house counsel by Cecilia Ziniti, a former general counsel at Anki, BloomTech, and Replit. With GC AI, You can compare contract language with your company’s approved terms, ask questions across company documents, and check the source passages behind an answer. As of October 2026, 2,100+ legal teams use GC AI, including Nestlé, TIME, Riot Games, Wayfair, and Columbia Sportswear.

Define the Legal Work and Permitted Data

Start with a task the business needs done. For example: commercial counsel will review draft supplier agreements against the company's approved terms. That description gives legal, procurement, and information security a specific workflow to evaluate.

On CZ and Friends, our podcast with legal and business leaders, SimplePractice Chief Legal Officer Ali Hartley described evaluating third-party AI tools through a matrix of the task, the data involved, and the features needed. Her security team also used AI as a first step in vendor review. Start by identifying what people will put into the tool.

Work: Name the task, expected output, and lawyer responsible for checking it.
Information: Identify the documents, personal data, confidential business information, and privileged material the task could involve.
People and systems: List the users, shared workspaces, connected applications, and external recipients.
Approval: Record who can authorize the use and which requirements must be satisfied.

Read the applicable agreement before deciding what to upload. A security certification does not authorize every category of data. Contract restrictions, client instructions, and your organization's policies may limit an otherwise useful workflow. The same distinction applies to ChatGPT confidentiality: privacy settings alone do not establish that an upload is permitted under your NDA.

Understanding the technology and protecting client information are core AI legal ethics duties addressed in ABA Formal Opinion 512. For privileged material, assess the proposed use, disclosures, vendor terms, and applicable law. No platform feature can decide privilege for a matter.

Five Security Questions to Ask Before You Sign

Gather the vendor's current security and contract materials. Record their dates and product scope, then use the questions below to identify findings your team can approve or send back to the vendor.

Who Receives Our Data, and What Can They Do With It?

Follow the original file, extracted text, prompts, and outputs through the application, model providers, document-processing services, and connectors. A model name alone does not establish the terms of the service using it. Identify each recipient's role, then separate these commitments:

Training: Can the recipient use inputs or outputs to train or improve models?
Provider retention: What does the model provider retain, for how long, and under which exceptions?
Application storage: What does the legal AI platform save so users can reopen chats, retrieve files, or reuse context?

A no-training commitment can coexist with stored files and chat history. Zero Data Retention (ZDR) at a model provider does not delete copies held by the application. Check the actual service, model, and feature, including any fallback used when the preferred provider is unavailable.

Who Can Access Documents and Connected Applications?

Customer isolation and permissions within your own team address different questions. Confirm how the vendor separates organizations, then check which colleagues, administrators, and support personnel can access a document or conversation.

Use a synthetic nondisclosure agreement (NDA) to demonstrate the permissions. Have an authorized user upload and share it, then confirm that a user without access cannot retrieve it. Include saved chats, folders, and any organization-wide sharing in the demonstration.

For connected applications, identify whose account supplies access, what the AI can read or change, and which actions require approval. Ask an administrator to demonstrate the configuration your team plans to use.

What Do the Audit Report and Encryption Controls Cover?

Read the SOC 2 Type II report for its legal entity, systems, audit period, exceptions, and subservice organizations. AICPA's review checklist also asks whether the covered Trust Services Criteria meet your needs and whether a bridge letter addresses the period since the audit. The five categories are security, availability, processing integrity, confidentiality, and privacy; not every engagement requires all five.

A SOC 3 report gives a general-use overview with less detail. It can support initial screening. Request the detailed SOC 2 report for procurement, observe its distribution restrictions, and identify any controls your organization must implement.

For encryption, check uploads, stored files, saved outputs, backups, and transfers to providers. Ask which methods apply at each step, who can access the data or keys, and which administrative actions are logged.

What Happens During an Incident or Provider Outage?

Commvault Chief Trust Officer Danielle Sheer urged boards to identify critical systems and establish backup plans before a crisis.

Read the incident-notification terms for the event that triggers notice, the timing, the recipient, and the vendor's cooperation obligations. Give your team an owner and a contact route for reporting a suspected incident.

For outages, ask whether the product stops, queues work, or switches providers. Any fallback needs to respect the providers and data terms your organization approved. Get the process in writing and include it in your review record.

What Happens When We Delete a Document or Leave?

Distinguish deleting a file, deleting a chat, closing an account, and requesting organization-wide erasure. Ask what each action removes from the application, indexes, logs, backups, and downstream services.

Get the deletion process, timing, retention exceptions, and export options in writing. A synthetic-file demonstration can show what users see after deletion. It cannot independently establish what remains in backend systems or backups; review the contractual commitment and supporting evidence too.

Evaluate GC AI's Controls Against Your Proposed Use

Ring-Fenced Data

We store customer data in segregated databases, with security controls including AES-256 encryption at rest and TLS 1.2+ in transit. None of our model providers train on your data. We maintain provider ZDR wherever feasible; our subprocessor list records provider-specific conditions. Review those commitments alongside the platform's storage and access controls for your proposed use.

Our Trust Center lists SOC 2 Type II and SOC 3 reports. Request the current materials and review their scope, period, and exceptions. Report access and disclosure remain subject to the applicable confidentiality terms.

Configure Providers and Connectors

On Teams and Enterprise plans, organization admins can control supported model providers. The controls cover Anthropic, OpenAI, and xAI, apply provider restrictions across product features, and record who changed a setting and when. At least one chat provider must remain available, Google/Gemini is not currently toggleable, and contractual restrictions can lock settings.

Model-specific exceptions belong in the approval decision. Our subprocessor list makes Anthropic ZDR conditional on Claude Fable 5 and newer models not being enabled on the account. Our model-control documentation says Fable 5 retains submitted data for 30 days for safety work and is off by default until an organization admin reviews the terms and enables it. Confirm the retention terms for each model before enabling it for your team's data.

Our connector controls let an organization choose Disabled, Read-only, or Full access. Each member's connected account and action permissions also constrain access. Read actions default to Always allow; write and delete actions default to Needs approval. Review those settings before connecting company systems.

Confirm Storage and Deletion Terms

Our Files feature stores documents for organization, sharing, and reuse across chats. That application storage is separate from model-provider ZDR. Decide which documents the team may keep, how they may be shared, and when they should be removed.

Our Services Agreement provides for in-app deletion and written organization-erasure requests. The DPA addresses deletion on termination and allows specified retention for backups, record-retention policies, and legal obligations. Check the agreement that applies to your purchase and confirm the process and exceptions with your procurement contact.

Build a Security Review Record With GC AI

Once your organization permits an evaluation, start with public vendor materials, a synthetic NDA, and sample company requirements. Add restricted reports or internal policies only when you are authorized to process them in the chosen workspace. The workflow below uses GC AI to organize a review; your lawyers and information security team verify the evidence and make the decision.

Consider a hypothetical review for commercial counsel who wants to use legal AI with confidential supplier agreements. This is a proposed workflow, not a customer deployment or a product test.

Collect the documents that cover the proposed use:

The applicable service agreement and order form
The DPA and security addendum
The current audit report and subprocessor list
Provider-retention and application-deletion documentation

Include amendments and incorporated policies. Record each document's date and scope, and identify anything missing.

Supply the intended users, models, connectors, and data categories separately from the vendor evidence. In this example, the company requires no model training and approved provider ZDR coverage for confidential agreements. It permits application storage under its retention policy. A provider-retention exception requires the general counsel and information security lead's approval. Until then, the permitted fallback is an evaluation using only public or synthetic documents. The hypothetical company supplies these positions for comparison with the vendor's contractual commitments.

Ask GC AI to compare the vendor's documents with those requirements:

Review the supplied documents for commercial counsel's proposed use: reviewing confidential supplier agreements. First list the documents, dates, applicable products, and missing or incorporated materials. Compare the vendor's commitments with the separately supplied company requirements.

Create a table showing each requirement, vendor commitment, supporting source passage, exception or missing evidence, proposed action, and review owner.

Separate customer isolation, model training, provider retention, application storage, access, and deletion. Identify the terms for enabled models and any fallback providers. Flag conflicts without deciding which document governs when the evidence is incomplete.

Draft a short recommendation stating what may proceed, what remains open, and whose approval is required. Do not infer upload permission from a certification or treat an unanswered question as satisfied.

That table gives counsel and information security a shared starting point. Each finding should point to a passage they can inspect, with the unanswered questions visible beside it.

From Documents to an Approval Decision

Assemble the inputs: Put the permitted source documents and company requirements in Files. Label their dates and identify the intended users, task, and data.
Analyze the commitments: Run the prompt above. Ask follow-up questions where an answer combines training with retention, overlooks an exception, or lacks a supporting passage.
Check the findings: Our Exact Quote feature opens the source document with the quoted words highlighted. Check the surrounding language, definitions, exceptions, and applicable version. Counsel checks the agreement and legal implications; information security evaluates the controls and the vendor's demonstrations.
Act on the gaps: Send the vendor focused questions, propose contract changes, or limit the approved use. Record the permitted data, users, providers, connectors, and conditions that must be met before access expands.
Keep a reusable record: Save the verified decision in your approved vendor-review system. Keep permitted source materials in Files for later work, with their dates and a named owner for reviewing changes.

Suppose the hypothetical vendor prohibits training but refers to a provider-retention schedule that was not supplied. Procurement needs to obtain that schedule for the enabled models and fallback providers.

After verifying the findings, counsel could send procurement this recommendation:

Continue the permitted evaluation with public and synthetic documents. Keep confidential supplier agreements out until procurement obtains the provider-retention schedule and legal and information security complete their review. Any provider-retention exception needs the general counsel and information security lead's approval.

Security is one part of the purchase decision. Our guide to evaluating legal AI vendors covers workload fit, Word integration, pricing, and pilot planning.

Revisit the record when a model, connector, contract, or retention policy changes. The saved evidence should let the next reviewer identify what changed and decide whether the approved use needs to change with it.

Bring a synthetic NDA, permitted vendor materials, and sample company requirements to a GC AI demo. Follow a finding from its source passage to the review table and proposed recommendation. Check whether missing evidence and approval conditions remain visible before deciding what your team can use.

See GC AI in Action

Frequently Asked Questions

What Security Documents Should We Request From GC AI?

Request the current SOC 2 Type II report, applicable service agreement, DPA, subprocessor information, and retention documentation. Our Trust Center and Security page are starting points. Confirm which documents cover your plan and proposed use, and observe any confidentiality or sharing restrictions on the reports.

Does GC AI Train on Our Data?

Our Services Agreement says we do not use Customer Data to train generative AI models, and our Security page prohibits model providers from training on customer data. Training restrictions are separate from the storage needed to keep files and chats available and from any model-specific retention exceptions.

Can We Upload Health Data to GC AI?

Our public Services Agreement, which applies to self-serve and Individual purchases, prohibits personally identifiable health data, payment-card industry data, classified information, and ITAR data. Order Form customers have separate applicable terms. Confirm the agreement and permitted data for your purchase before uploading health information; do not infer permission or a business associate agreement from a security certification.

Does Zero Data Retention Mean GC AI Saves No Files?

No. Provider ZDR governs specified model-provider retention. Our Files feature saves documents for reuse and sharing, and application storage has its own deletion terms. Check the enabled model's exceptions, the documents your team may retain, and the process for deleting files or requesting organization-wide erasure.

Does Using Legal AI Preserve Attorney-Client Privilege?

No platform feature guarantees privilege. Assess the particular use, confidentiality terms, disclosures, client instructions, and applicable law. Security controls and retention terms inform that assessment, but a lawyer must decide whether the proposed workflow is appropriate for the matter.

Does a DPA Make a Legal AI Workflow GDPR Compliant?

A DPA alone does not establish compliance. Identify the parties' roles, the basis for processing, and where personal data is stored, processed, and accessed. The European Commission’s guidance distinguishes processor terms from international-transfer safeguards. Review the ICO's separate guidance for UK personal data and determine which requirements apply to the workflow.

Can We Evaluate Legal AI Before Approving It for Confidential Work?

If your organization permits an evaluation, use public materials and synthetic documents to check the workflow, source citations, and visible permissions. Keep confidential or restricted information out until the applicable terms, settings, and proposed use are approved. A demonstration with synthetic data does not independently verify the vendor's backend controls.

Which AI Tools Are Safe for In-House Legal Teams?

Start with the workflow and data your team needs to handle, then check the applicable agreement, permitted data, enabled models, access, retention, and deletion. On GC AI Teams and Enterprise plans, administrators can also review provider controls. Record the approved workflow and any conditions before expanding use.

Data Security in Legal AI: What to Know Before You Sign