Category: blog
AI assistants are being added to email, documents, customer service, software development, meetings, finance, and internal search.
They can also access company files, cloud platforms, business applications, and APIs.
That access requires policy controls before deployment.
A useful starting point is the NIST AI Risk Management Framework and its Generative AI Profile.
Policy
Every AI assistant policy should answer five questions
- What tools are approved
- What data may be processed
- Which users may access each tool
- What actions may be performed
- Who reviews activity and incidents
The policy should apply to
- Chat-based assistants
- Microsoft and Google AI features
- Coding assistants
- Meeting transcription tools
- CRM and help desk copilots
- AI search tools
- Plugins and browser extensions
- Autonomous agents connected to files or APIs
- AI features built into business applications
Personal AI accounts should not be used for company data unless specifically approved.
Unapproved tools should be blocked or monitored where possible.
Employees should have a documented process for requesting new tools.
1. Create an AI acceptable-use policy
The acceptable-use policy should define permitted and prohibited activity.
Permitted use
Examples include
- Drafting internal documents
- Summarizing approved business content
- Creating meeting notes
- Preparing first-draft customer communications
- Generating code for review
- Searching approved knowledge bases
- Automating low-risk administrative tasks
Prohibited use
The policy should prohibit
- Uploading credentials, secrets, or private keys
- Submitting payment card data to public AI tools
- Uploading sensitive customer records without approval
- Using personal accounts for company information
- Treating AI output as final without human review
- Using AI to make unsupervised employment, legal, financial, or compliance decisions
- Connecting an assistant to systems without an approved owner
- Disabling logging, retention, or access controls
- Copying confidential information into prompts to bypass normal workflows
AI output should be treated as untrusted content.
It may contain inaccurate information, exposed data, malicious instructions, or unsafe code.
Human review should be required before output is sent externally, placed into production, or used in a high-impact decision.
2. Classify data before it reaches an AI tool
AI security depends on data classification.
A simple model can include four levels
| Data level | AI processing rule |
|---|---|
| Public | May be processed in approved tools |
| Internal | Enterprise AI environment required |
| Confidential | Approval, access controls, and retention limits required |
| Restricted | External AI processing prohibited unless specifically authorized |
Restricted data may include
- Passwords
- API keys
- Private encryption keys
- Social Security numbers
- Payment information
- Protected health information
- Legal records
- Unreleased financial data
- Security configurations
- Customer credentials
- Backup credentials
- Incident response data
Data minimization should be required.
Employees should provide only the information needed for the task.
Names, account numbers, addresses, and other identifiers should be removed or replaced where possible.
Sensitive documents should not be uploaded because an assistant appears convenient.
3. Define vendor and retention requirements
AI providers should be reviewed before use.
The review should cover
- Whether customer data is used for model training
- Retention periods
- Data deletion procedures
- Encryption at rest and in transit
- Regional data storage
- Subprocessors
- Breach notification
- Administrative access
- Audit records
- Contractual data protections
- Private model or zero-retention options
- Integration with SSO and identity systems
The provider’s default settings should not be assumed to meet business requirements.
Retention should be configured based on the data classification and business purpose.
Prompts, uploaded files, generated output, and tool activity may contain sensitive information.
They should be handled under the same retention and legal-hold requirements that apply to other business records.
4. Use identity and access management
AI assistants should inherit existing identity controls.
Required controls should include
- Single sign-on
- Multi-factor authentication
- Managed devices
- Role-based access control
- Conditional access
- Separate administrator accounts
- Timely offboarding
- Periodic access reviews
An assistant should not receive broad access simply because it needs to perform one task.
It should access only the systems and data required for that task.
An assistant connected to a document repository should respect the user’s existing permissions.
It should not operate through a shared administrator account.
For agents that call APIs or perform actions, use
- Separate workload identities
- Short-lived access tokens
- Narrow scopes
- Task-based permissions
- Automatic token revocation
- Approval for sensitive actions
- Complete activity logging
High-risk actions should require human confirmation.
Examples
- Sending external email
- Changing financial records
- Creating user accounts
- Modifying firewall rules
- Deleting files
- Changing backup settings
- Publishing customer-facing content
- Approving payments
5. Address prompt injection and unsafe instructions
Prompt injection is a major AI security issue.
An attacker may place instructions in
- Email messages
- Websites
- Shared documents
- PDFs
- Support tickets
- Calendar invitations
- Code repositories
- CRM records
The assistant may process the content and treat embedded instructions as trusted commands.
Policies should require separation between instructions and data.
Technical controls should include
- Input filtering
- Protected system instructions
- Connector restrictions
- Sandboxed tool use
- Output filtering
- URL and file scanning
- Sensitive data detection
- Human approval for external actions
- Logging of prompts and tool calls
The OWASP GenAI Security Project provides guidance for large language model and agentic application risks.
Employees should be trained to treat external content as untrusted.
They should not instruct an assistant to follow directions found inside an email or document without review.
6. Maintain an AI and agent registry
Every approved AI tool should have an owner.
The registry should record
- Tool or assistant name
- Business purpose
- Department
- Data sources
- Connected applications
- User groups
- Authentication method
- Risk classification
- Vendor
- Retention settings
- Approval date
- Review date
- Incident contact
The registry should include plugins, browser extensions, automation tools, and embedded application features.
A tool may be missed if only standalone AI applications are reviewed.
Shadow AI should be identified through
- Cloud access logs
- DNS monitoring
- Endpoint telemetry
- Identity provider records
- Firewall data
- CASB controls
- Procurement reviews
- Employee surveys
Blocking every AI tool may cause employees to move to less secure alternatives.
Approved tools should be available with clear usage rules.
7. Establish monitoring and incident response
AI activity should be monitored like other business system activity.
Logs should capture
- User identity
- Device
- Time
- Prompt activity
- Uploaded files
- Data sources
- API calls
- Tool actions
- Generated output where appropriate
- Administrative changes
- Policy violations
Logs should be protected from unauthorized modification.
Relevant activity should be connected to existing security monitoring and SIEM workflows.
Incident procedures should address
- Accidental data submission
- Unauthorized tool access
- Compromised AI accounts
- Prompt injection
- Malicious generated code
- Incorrect automated actions
- Vendor breaches
- Inappropriate output
- Unauthorized connector changes
Response actions may include
- Revoking tokens
- Disabling connectors
- Suspending accounts
- Preserving logs
- Isolating affected devices
- Restoring data from backups
- Notifying affected parties
- Reviewing policy gaps
- Reauthorizing the assistant after testing
8. Require backup and recovery controls
AI tools create new business records.
Prompts, configurations, knowledge bases, connector settings, generated files, and workflow rules may require backup.
The backup policy should identify
- Which AI data is business-critical
- Where configurations are stored
- How data is exported
- Recovery time requirements
- Recovery point requirements
- Administrative recovery access
- Vendor termination procedures
- Testing frequency
AI tools should not become the only location for business knowledge.
Critical files should remain in approved storage with backup and access controls.
Backup and security monitoring should cover connected systems, not only the AI application.
9. Train employees before access is granted
Training should be practical.
Employees should know
- Which tools are approved
- Which data is restricted
- How to report accidental disclosure
- How to identify suspicious AI instructions
- Why output requires review
- How to verify generated code
- When human approval is required
- How to request a new AI tool
Training should be repeated when tools, vendors, or policies change.
Access should be removed when employment ends or job responsibilities change.
X-Tek managed AI services
AI security requires configuration, monitoring, access management, and ongoing review.
X-Tek can support a managed AI program through business IT support, cloud services, network infrastructure, security controls, backup monitoring, and user support.
A managed AI engagement may include
- AI tool inventory
- Use-case risk review
- Approved-tool configuration
- Microsoft and Google cloud integration
- SSO and MFA configuration
- Role and access review
- Data classification guidance
- Endpoint and network controls
- Backup and recovery planning
- Security monitoring
- Policy development
- Employee training
- Incident response support
- Periodic policy and configuration reviews
These services can be aligned with existing X-Tek business IT services.
The initial rollout should use low-risk workflows.
Examples
- Internal document drafting
- Meeting summaries
- Approved knowledge-base search
- Nonproduction code assistance
- Administrative scheduling
Higher-risk workflows should require additional testing and approval.
Businesses needing an assessment can submit the Business Solutions Information Request.
Implementation checklist
Before deploying an AI assistant
- Assign a business owner
- Document the use case
- Classify the data
- Review the vendor
- Confirm retention settings
- Require SSO and MFA
- Limit permissions
- Configure logging
- Define human approval points
- Test prompt injection risks
- Add the tool to the registry
- Create an incident procedure
- Confirm backup requirements
- Train users
- Schedule a review date
AI assistants should be managed as business systems.
They should not be treated as informal browser utilities.
Policy, identity, data governance, monitoring, and recovery controls should be established before company data or business actions are connected.
Contact Information
Business Solutions Information Request:
https://xtekit.com/business-solutions-information-request/
815-516-8075