1. Access and permissions
AI tools connected to organizational content can make existing permission problems more visible. An employee may already have access to a sensitive record that an assistant makes easier to discover. Review which information the tool can reach and whose identity or connector permissions it uses.
Microsoft documents that Copilot respects authorized access, and that overshared content can increase risk. That behavior should not be generalized to every AI product or integration without checking its design and configuration.
Starting work: review broad sharing, stale guests, privileged identities and sensitive repositories. Test the relevant access and discovery settings with the system owners. Record exceptions and assign remediation responsibilities.
2. Data classification
People need a usable way to identify the information they can share with an approved tool. Define categories appropriate to the organization, such as public, internal, confidential and restricted, with concrete examples for each.
Starting work: identify where sensitive information lives, assign owners and define permitted uses. If classification labels or data-loss prevention tools support the policy, confirm their coverage and test their behavior. A label alone does not ensure that data will be handled correctly.
3. AI tool inventory and use cases
Inventory approved tools, integrations, browser extensions and the business activities they support. Ask employees how they use AI, and compare those answers with available procurement, identity and endpoint records. Technical visibility can have gaps.
Starting work: document the tool, owner, use case, information involved and approval state. Offer a practical route to request new tools and explain which uses need additional review. Keep the inventory current as products and integrations change.
4. Vendor terms and configuration
Evaluate the actual product, account type, contract and settings. “Enterprise” is not a universal privacy guarantee, and consumer-account terms vary by provider.
Starting work: review use of inputs for model training, retention, deletion, tenant isolation, access, subprocessors and contractual commitments. Confirm the relevant configuration and ask legal and privacy owners to review the proposed use when needed.
A vendor report or marketing claim does not answer every question about your specific use case. Record assumptions and the evidence behind approval.
Connect the work to ongoing risk management
These four areas are preparation topics, not a complete AI governance framework or a fixed compliance sequence. NIST’s AI Risk Management Framework organizes work around four connected functions: Govern, Map, Measure and Manage.
Use your organization’s goals and operating context to identify further work, including reliability, human oversight and evaluation of the intended use. Applicable standards, contractual obligations and legal requirements need their own scoped review.
Write a policy people can follow
Define approved and prohibited uses, permitted information, tool approval, human review and escalation. Explain who owns the policy and how employees can ask questions or report a concern.
Policy and implementation can develop together. Prioritize higher-risk uses, identify interim restrictions and review whether the controls operate as intended. Publish a document that reflects the current state rather than implying every safeguard is already complete.
Related reading
- Clarifying security oversight with your IT provider
- What a vCISO does
- Controls, owners and evidence for readiness
