A company can approve an AI assistant for employees without deciding whether the same technology belongs in a product offered to customers. A policy limited to employee use can set rules for uploads and reviewing outputs while giving a product team little direction on performance claims or what happens when the feature fails.
Companies offering AI products or features need governance that covers both the above activities and internal AI use. The safeguards should reflect what each system does and who could be affected. An internal hiring tool warrants closer scrutiny than a tool that formats documents.
This alert is the second installment in our Cybersecurity Awareness Month series and details five pillars of AI governance that companies should be considering when using the technology. Throughout October, we will be sharing insights on emerging cybersecurity, privacy, and AI governance developments affecting businesses. You can find our first alert here and be on the lookout for future alerts.
Organization and oversight
Senior leadership should designate an executive or team to oversee AI governance. We generally favor a hybrid model in which the governance lead sets company-wide policies and reviews higher-risk proposals, while department leaders manage routine uses within those policies.
Under this model, department leaders identify where AI is used and assign a business owner to each use. That owner documents its purpose, the people and data involved, and how the company will judge whether it delivers value. Uncertain or higher-risk proposals go to the governance lead for review.
This inventory gives the governance lead a basis for policies that fit the business. Employees using purchased tools require guidance on approved tools and permitted data, training on their limitations, and a process for checking outputs and reporting problems. Where the company develops, customizes, or integrates AI, the policies also assign responsibility for testing, release approval, and changes within its control. Those responsibilities follow the work the company performs, whether the system serves employees or customers. Business owners, working with legal and technical personnel, identify further safeguards and who can suspend a problematic use.
Legal and regulatory compliance
Legal counsel uses the inventory to determine which requirements apply to each activity. Under the EU AI Act, a company's obligations depend in part on whether it provides or uses the AI system; calling a use internal does not, by itself, establish an exemption. The GDPR and CCPA apply according to their respective scope and the processing involved. Covered entities and business associates also need to assess HIPAA requirements when AI involves protected health information. Internal hiring and employee-evaluation tools warrant employment-law review.
Offering an AI product can bring additional obligations based on its function, users and market, including consumer-protection, sector-specific, and AI disclosure requirements. Counsel should distinguish binding duties from voluntary practices and revisit the analysis when the law or the system changes.
A tool approved for employees may require a separate contract review before it becomes part of a customer product. If the license restricts it to internal use, the company should negotiate permission for the customer feature or choose another tool. Customer terms and performance claims should reflect the system's capabilities and any vendor or license restrictions, including those governing open-source software or models. For either type of use, contracts should address rights to inputs and outputs, vendor use of company data, confidentiality, and liability. Audit rights, service commitments, and change notices should reflect the transaction's risks.
Ethics, transparency, and interpretability
AI that screens job applicants, informs credit decisions, or evaluates housing applicants warrants scrutiny for discriminatory effects, whether used internally or offered to customers. Business owners should work with legal and technical personnel to determine what evaluation could detect those effects. A tool used to summarize public technical manuals may need accuracy or safety checks, but its use of AI alone does not require demographic bias testing.
The consequences of error determine how much human review is appropriate. When an AI output informs a consequential decision, a reviewer needs enough information, time, and authority to challenge it. The governance lead should identify who can resolve ethical concerns that a business owner cannot address.
Employees require guidance on what approved tools can reliably do and when human review is required. Customers need information appropriate to an AI product's intended use and limitations. Counsel identifies required disclosures or explanations, including for internal uses affecting employees or other individuals. Where warranted, affected people should have a route to human review.
Data, AI operations, and infrastructure
Rules on permitted data have to account for what happens after someone enters it into an AI system. Restrictions on confidential information, personal data, and contractually protected material need to align with the vendor's retention, training, and access practices. Customer products also require rules on what customer data is collected and how the company and its vendors may use it, and whether the system is hosted by the company or a vendor.
Companies training or customizing models should document their rights to use the data, its source and suitability, and its retention period. A customer feature requires testing as a complete product, including its configuration and connected data. Testing the underlying model alone does not establish how the finished feature will perform.
Approval of a purchased tool should rest on its suitability for the task and what the vendor can show about testing and monitoring. The depth of review should reflect the use's risk. Reports of errors or changes to the service can justify revisiting the assessment that supported approval. Where the company controls the system, deteriorating performance calls for investigation and a decision about correction, restriction, replacement, or appropriate retraining.
Records of these reviews and decisions support investigation and applicable obligations. The company should limit access to those records and set a retention period. Logging every prompt and output can itself create privacy and confidentiality risks.
AI security
An AI tool authorized to retrieve company records or take actions should receive only the access required for those tasks. The company remains responsible for its configuration, data protections and permissions, even when the vendor controls the underlying model. During Cybersecurity Awareness Month, security teams can check whether approved tools have access beyond their authorized tasks.
Malicious instructions in documents or websites can attempt to redirect a generative AI tool, a threat known as prompt injection. When the tool can reach company records or other systems, its permissions help determine the consequences of that attack. Screening inputs alone should not be treated as a complete defense. Companies developing or operating AI systems also have to protect the components they manage, including training data, model files, software dependencies, and deployment.
For customer products, the security review should consider foreseeable misuse and whether one user could obtain another's information. Incident procedures should give responsible personnel authority to investigate failures and suspend affected functions, with a clear path for coordinating with vendors. These procedures apply to internal uses and external offerings. They should provide for assessing notification duties under applicable law and contracts, along with other communications warranted by the incident.
For more information, please contact us or your regular Parker Poe contact. Click here to subscribe to our latest alerts and insights.