Key Takeaways:
- When rolling out an AI tool across your organization, consider how company data might be exposed across internal teams.
- Internal data access isn’t just a governance concern—contracts professionals are key in preventing internal data leakage through vendor terms.
- Different AI use cases may call for different contractual protections, such as data segregation clauses, limitations on training data, and adherence to document-level permissions.
Contract professionals often focus on external data exposure when evaluating generative AI tools, such as whether prompts and data inputs might be accessed and used by the model provider or appear in outputs delivered to third parties.
A less explored yet still significant concern is internal data leakage, where sensitive information is improperly shared within the organization itself.
While this might seem like a data governance issue, it’s also a contractual risk. And one that attorneys and contract professionals are uniquely positioned to manage. This article outlines three potential internal leakage scenarios and identifies contractual clauses that can help mitigate them.
Scenario 1: Organization-Wide Use of an LLM
Context:
Your organization rolls out an enterprise LLM available to all teams, including HR, Finance, Legal, Engineering, and Marketing. The HR team uses the tool to analyze sensitive employee data, including salaries, performance reviews, and disciplinary history.
Risk:
If model memory or workspace configurations permit cross-user access, employees in other departments might input prompts that inadvertently—or intentionally—retrieve or synthesize sensitive HR data. Even without access to the original documents, the LLM could generate summaries or insights that reveal confidential information.
Contractual Safeguards:
- Data Segregation Clauses: Require vendors to isolate data inputs by default. Contracts should prohibit cross-user retrieval or inference unless explicitly configured and authorized.
- Configuration Enforcement Provisions: Ensure contracts prevent end-users (i.e., any employee using the LLM) from altering default sharing settings. If the LLM allows data pooling or shared memory, require centralized control over those configurations.
- Use Limitation Warranties: Vendors should warrant that user data will not be retained for cross-user inference, internal indexing or model training—unless expressly permitted.
Scenario 2: Custom AI Model Trained on Internal Data
Context:
Your Marketing Team partners with a vendor to develop a custom LLM that generates campaign content based on customer feedback and CRM data. To improve relevance, the model is trained on emails, surveys, and past campaign performance—some of which include high-value client information.
Risk:
Without adequate controls, model outputs may inadvertently reveal deal-specific details. For example, a sales associate asking for “messaging for VIP clients” might receive suggestions that reflect real engagement history with specific accounts.
Contractual Safeguards:
- Training Data Restrictions: Include contract language specifying what types of data may and may not be used for training. Exclude client identifiers, financial information or strategic plans unless expressly approved.
- Anonymization & Minimization Warranties: Vendors should warrant that training data is anonymized or scrubbed of sensitive fields unless the organization grants written permission.
- Role-Based Output Access: Ensure that access to the trained model and its outputs is limited based on user roles, in alignment with the sensitivity of the underlying training data.
- Audit Rights: Include provisions allowing periodic audits of both training inputs and generated outputs to detect and address potential leakage.
Scenario 3: AI Indexing of Shared Workspaces
Context:
Your organization deploys an AI assistant within a productivity suite, like Gemini in Google Workspace or Copilot in Microsoft 365. The AI assistant indexes shared drives, documents, and communications to answer questions and provide intelligent summaries.
Risk:
If access permissions are misconfigured or indexing settings are overly broad, the AI assistant may surface information from files that a user would not normally be able to access. For instance, a general prompt like “Summarize Q3 budget concerns” could pull sensitive content from a finance document the user wasn’t meant to see.
Contractual Safeguards:
- Indexing Control Clauses: Secure admin-level control over what the AI assistant is allowed to index. Contracts should provide for the exclusion of folders, content types or departments as needed.
- Permission Adherence Warranties: Ensure that the AI assistant fully respects existing document-level permissions. If a user cannot manually access a file, they should not be able to retrieve or synthesize its contents via the AI assistant.
- Prompt Monitoring & Filtering: Work with vendors to implement filters that flag overly broad or high-risk prompts.
Internal Data Leaks Are a Contract Issue
As AI tools become more widely used across organizations, the risk of accidentally exposing sensitive information inside the company grows. This kind of internal data leakage often happens not because someone breaks the rules but because the AI was trained or configured without enough safeguards.
That’s where contract professionals come in. By building protections into vendor agreements—like limiting data sharing, controlling how AI tools are set up, and making sure sensitive data isn’t used for training—you can help prevent these risks.
————————————————————————
For more on AI and Contracts, check out my full column here.
