Take Control of Your AI

Establish authority and accountability

Someone must own the program and someone must have authority to stop a use that exceeds the organization’s tolerance.

• Name an executive sponsor and accountable program owner.
• Define who participates in reviews and who may approve, reject, pause, or retire AI.
• Document risk tolerance, prohibited uses, exception handling, reporting cadence, and required records.

Evidence: charter, responsibility matrix, approval authority, meeting records, exception register.

Discover and inventory AI

Inventory purchased, developed, embedded, and employee-selected AI, including browser tools, personal accounts, extensions, meeting assistants, copilots, APIs, and agents.

• Record owner, purpose, users, vendor, plan, data access, integrations, permissions, and lifecycle state.
• Identify declared and suspected Shadow AI.
• Reconcile the register with identity, browser, expense, procurement, SaaS, network, and security records where practical.

Evidence: AI system register, discovery results, SaaS inventory, connector register.

Define the use case

Review the actual business process, not just the product name. The same AI tool can be low risk for drafting and high risk when it ranks applicants or takes action in another system.

• Document the problem, benefit, users, affected people, intended use, and foreseeable misuse.
• Identify the authoritative source of truth and define success, failure, and unacceptable outcomes.
• Decide whether a non-AI approach could meet the need with lower risk or cost.

Classify data and data rights

Identify everything the AI can receive, retrieve, infer, generate, or expose. Record source, owner, permission, retention, deletion, and whether the vendor may use data for training or product improvement.

• Apply the organization’s data classification scheme.
• Minimize fields and remove identifiers when they are not necessary.
• Set rules for prompts, uploads, outputs, history, memory, feedback, logs, and retrieved data.

Identify legal, contractual, and insurance considerations

Screen for privacy, security, employment, biometric, consumer, civil-rights, intellectual-property, accessibility, records, and sector-specific considerations. Review customer contracts, supplier terms, and insurance conditions.

Review the vendor and platform

• Identify the contracting entity and exact product plan.
• Review security, privacy, availability, retention, deletion, training use, subprocessors, data location, and incident terms.
• Review identity controls, logs, offboarding, export, and connector permissions.
• Set recurring review dates and change triggers.

Classify and treat risk

Assign an internal tier based on impact, scale, reversibility, data sensitivity, autonomy, affected people, and uncertainty. Identify plausible failure scenarios, current safeguards, and remaining exposure.

• Choose a response: avoid, reduce, transfer, accept, or monitor.
• Require higher authority for higher-risk acceptance.
• Define hard stops that prevent deployment or require immediate suspension.

Design meaningful human oversight

A person who clicks “approve” without enough information, time, competence, or authority is not meaningful oversight.

• Name the person accountable for the result.
• Define what the reviewer checks and when escalation is mandatory.
• Provide correction, appeal, override, and service-recovery paths.
• Log material overrides and analyze recurring patterns.

Evaluate before approval

• Test task success, factual accuracy, unsupported output, and source grounding.
• Test privacy leakage, access control, prompt injection, unsafe instructions, and foreseeable misuse.
• Test materially different outcomes for relevant groups when people may be affected.
• Verify logging, rollback, shutdown, and manual continuity.

Evidence: evaluation plan, representative test set, results, limitations, remediation, acceptance decision.

Approve and deploy

• Record a formal go, conditional-go, pilot, or no-go decision.
• Confirm approved users, purposes, data, integrations, settings, and locations.
• Communicate limitations, prohibited behavior, and reporting channels.
• Activate monitoring and set review dates before operational use.

Monitor and respond

Monitor new tools, sensitive-data interactions, access, connectors, errors, complaints, corrections, overrides, vendor changes, policy exceptions, and overdue reviews.

• Integrate AI incidents with privacy, security, legal, HR, and continuity procedures.
• Preserve logs and evidence.
• Define containment, notification evaluation, correction, recovery, and root-cause steps.
• Retest and reapprove after material changes or incidents.

Educate and continually improve

• Provide baseline AI literacy and role-based training.
• Teach which data is prohibited, restricted, or approved, and why.
• Refresh training after material system, policy, threat, or regulatory changes.
• Retire unused systems, revoke access, address retained data, and archive the decision.