AI Security Checklist for Small Businesses

Data

• Define sensitive information in language employees understand.
• Classify customer, financial, contract, HR, proprietary, credential, regulated, and internal data.
• Document which information is prohibited, approved, or requires review.
• Minimize the data provided to AI systems.
• Confirm the employee is already authorized to use the source information.
• Define retention and deletion expectations based on current vendor capabilities and business requirements.

Applications and vendors

• Create an approved AI tool list.
• Discover Shadow AI, including personal accounts, browser extensions, meeting tools, and AI embedded in SaaS.
• Review current official vendor documentation and applicable terms.
• Evaluate data use, retention, deletion, account controls, integrations, admin features, and logging.
• Record the business owner and approved purpose for each application.
• Remove abandoned or duplicate tools.

Integrations

• Inventory APIs, OAuth grants, connectors, webhooks, service accounts, and credentials.
• Use secure secret storage and rotation where appropriate.
• Grant the minimum read and write permissions required.
• Separate test and production environments.
• Validate inputs, outputs, and record identifiers before writes.
• Log errors and make failures visible.
• Review the integration risk model whenever a new system is connected.

AI agents

• Give every agent a named business and technical owner.
• Define its goal, permitted tools, data sources, actions, stopping conditions, and escalation.
• Apply least privilege to records, folders, mailboxes, calendars, and actions.
• Require human approval for high-impact steps.
• Test prompt injection and untrusted-content handling without using live sensitive data.
• Log meaningful actions and approvals.
• Provide a clear disable and recovery procedure.

Employees

• Train staff on approved tools and data boundaries.
• Explain that plausible AI output can still be wrong.
• Teach employees to recognize AI-enabled phishing and impersonation.
• Explain direct and indirect prompt injection at a practical level.
• Define when human review is mandatory.
• Provide a simple path to report mistakes or questionable use.
• Publish an AI acceptable-use policy with real examples.

Monitoring and governance

• Review accounts, access, integrations, and vendors on a defined schedule.
• Remove unused tools, OAuth grants, API keys, and service accounts.
• Review new AI features added to existing SaaS products.
• Track incidents, exceptions, employee questions, and policy updates.
• Evaluate whether workflows still produce useful results.
• Review security after material vendor, data, or process changes.

Human approval checklist

Consider explicit approval before sending external communications, changing customer records, deleting data, making purchases, moving money, changing permissions, handling legal-sensitive or HR-sensitive decisions, executing security actions, or completing high-value transactions.

What to do with the findings

Turn checklist gaps into an ordered plan: immediate credential or access problems, policy and training needs, vendor decisions, integration cleanup, and longer-term workflow improvements. The AI Security and Governance service can help establish the operating model. Start with the free AI and Cyber Risk Snapshot or schedule an AI Opportunity and Security Assessment.