Courts are beginning to treat platform terms as evidence of whether information shared with an AI tool was confidential. Businesses should review those terms before employees upload sensitive information, not after a dispute begins.
An employee can disclose confidential information to an AI provider in seconds. The employee may be trying to work faster: summarize a contract, prepare notes from a meeting, analyze a customer complaint, or outline a response to threatened litigation. But if the tool is not approved, the organization may not know what was uploaded, where it went, how long it will be retained, or whether the provider can use it.
The practical response is not a blanket ban on AI. It is a usable governance process that tells employees which tools are approved, what information may not be entered into unapproved tools, and where to obtain a prompt answer when a new use arises. For many businesses and nonprofits, that process will not be handled entirely in-house. It may need to include outside counsel, privacy and security professionals, vendor-review support, fractional technology resources, or other trusted advisors who can help the organization make defensible decisions without building a large internal department.
The issue is the disclosure, not the label
AI is no longer confined to standalone chatbots. It appears in drafting products, meeting applications, research platforms, customer-service systems, HR tools, marketing software, and developer environments. Some of those features arrive through products the business already uses. Others are adopted by an employee or business unit without review by the people responsible for legal, privacy, security, procurement, or technology decisions—whether those roles sit inside the organization or are handled by outside professionals.
That is shadow AI: the use of AI outside the organization’s approved review process. It is usually well-intentioned. It can also move confidential business information, personal data, trade secrets, privileged communications, or regulated information to a third party under terms no one at the organization has reviewed.
Bottom line: If a platform’s terms say user content is not confidential, a court may take the provider at its word.
Courts are reading the fine print
Recent federal decisions show why the review cannot stop at a product name or an “enterprise” label. Courts are looking at the actual service used, the applicable account type, and the provider’s terms and privacy disclosures when deciding whether information was treated as confidential.
In Fry v. Fry, a court considering a request to seal documents learned that the petitioner had uploaded redacted materials to non-enterprise versions of Claude and Claude Cowork. The court cited Anthropic’s privacy and model-training disclosures and concluded that full sealing was inappropriate because the documents had already been provided to a private, for-profit AI company. No. 26-cv-3469-KSM, Memorandum at 5 n.4 (E.D. Pa. Aug. 27, 2026).
This and other decisions do not establish that every use of an AI tool waives privilege or eliminates confidentiality. The narrower business lesson is enough: courts will examine what the provider told the user about confidentiality and data use. If the terms permit retention, review, training, or disclosure, the organization may have difficulty later arguing that the upload remained confidential.
One AI use can create several legal problems
Confidentiality is only one part of the analysis. A single use can also implicate privacy, security, recording-consent laws, employment law, contractual restrictions, trade-secret protection, retention duties, and discovery obligations.
Consider an employee who invites an AI notetaker to a board, client, HR, or internal-investigation meeting. Before the first transcript is created, the organization should know:
- whether participants received appropriate notice and provided any required consent;
- what the tool records, generates, and stores;
- whether the meeting may include privileged advice, personnel information, trade secrets, or information subject to a third-party confidentiality obligation;
- whether the provider may retain or use the content to improve its services or models; and
- whether the recording or transcript must be preserved, produced in discovery, or placed on litigation hold.
The same point applies to document summarization, coding assistants, customer chatbots, applicant screening, and other AI-enabled workflows. Governance should address the business process and the information involved, not merely whether the product is marketed as AI.
A practical AI intake is the front door
An AI intake gives employees and business units a clear place to bring proposed uses. It also gives the organization a consistent way to identify proposed enterprise AI, shadow AI, assess risk, and document a decision to approve, approve with controls, restrict, or prohibit a use. The intake can be scaled to the organization. A smaller business does not need a standing committee or full internal team, but it does need an identified path for legal, privacy, security, vendor, and technology review when the use case requires it.
The intake does not need to be long. It should answer a short set of decision-focused questions:
- Identify the use. What tool or feature is involved, who will use it, and what business problem will it solve?
- Identify the data. Will users enter confidential information, personal data, source code, trade secrets, recordings, privileged communications, or regulated information?
- Read the applicable terms. Which terms apply to this account and configuration? What do they say about confidentiality, provider access, training, retention, deletion, security, subprocessors, and disclosure?
- Assess the impact. Will the tool communicate with customers, record people, or influence employment, eligibility, pricing, fraud, or other consequential decisions?
- Set the controls. What human review, access restrictions, logging, retention, notice, consent, and escalation requirements are needed?
- Plan for a dispute or incident. Can the organization preserve relevant inputs and outputs, investigate misuse, respond to privacy requests, and obtain vendor cooperation?
The result should be an operating decision, not a policy memo that disappears into a shared drive.
What an AI playbook should cover
A workable AI governance playbook usually has a small number of connected parts:
- an acceptable-use policy that distinguishes permitted, prohibited, and approval-required uses;
- an inventory of approved tools and AI-enabled features, including the business owner and permitted data types;
- risk tiers that trigger closer review for sensitive data, external-facing systems, automated decisions, employment uses, recording, biometrics, or international transfers;
- vendor-review standards focused on confidentiality, data use, model training, security, retention, deletion, incident notice, and subcontractors;
- clear controls for meeting recording and transcription;
- practical employee training with examples of what may not be entered into public or unapproved tools; and
- incident-response and legal-hold procedures that account for AI inputs, outputs, recordings, transcripts, and vendor-hosted data.
The National Institute of Standards and Technology’s AI Risk Management Framework can help organize the program around four functions: Govern, Map, Measure, and Manage. The framework is useful, but the business objective is simpler: make responsible adoption easier than going around the process.
What executives should do now
- Tell employees what information may not be entered into public or unapproved AI tools.
- Create a short intake process and a defined owner for AI review.
- Inventory AI features already in use, including meeting assistants and embedded product features.
- Review the terms that apply to the specific account and configuration, not only the vendor’s marketing materials.
- Identify the internal or outside resources responsible for Legal, Privacy, Security, Procurement, and IT review, and involve them in proportion to the risk of the use case.
- Address AI-generated and vendor-hosted content in incident response, retention, and litigation-hold procedures.
- Use outside counsel, fractional privacy or security support, and trusted technology advisors when the organization does not have those functions in-house.
Conclusion
The fastest path to unmanaged AI risk is to assume that employees will wait for a comprehensive policy before using the tools available to them. They will not. A better approach is to give them approved options, clear limits, and a fast review process.
Envisage Law can serve as part of that practical governance team for businesses and nonprofits that do not have dedicated in-house team for AI governance and risk review. We help organizations design AI intake procedures, acceptable-use policies, vendor and contract review standards, confidentiality and privilege controls, privacy and recording workflows, employee training, and incident-readiness plans that fit the organization’s size, risk profile, and existing resources.
Selected authorities and resources
