PCC Learning Center · AI Governance
Know What the AI Vendor Controls—and What Your Business Still Owns
Buying an AI service transfers processing to a vendor. It does not transfer accountability for configuration, permissions, employee use, connected data, approval boundaries or incident decisions.
Core Rule
A familiar brand, security certification or paid business plan can support due diligence. None replaces a use-case review, clear contract terms or named incident owners.
The Vendor Owns Its Service
The provider operates the model and platform according to its contract, security controls and service commitments.
Your Business Owns the Use
You decide who may use the tool, what data it handles, what systems it reaches, which actions require approval and how incidents are managed.
Start With the Business Use Case
Do not approve a vendor in the abstract. The same service creates different risks when brainstorming public topics, recording client meetings, updating customer systems or supporting employment decisions.
Purpose and People
Document the business purpose, expected users, business owner and technical owner.
Data and Access
Record data classifications, connected systems, permissions, account type and whether the tool can take actions.
Impact and Oversight
Define required human review, approval boundaries and the effect of failure, unavailability or misuse.
Add the use case to the AI Tool Inventory before approval.
Ten Areas to Review
Ask questions appropriate to the actual data, access, authority and consequences—not merely the vendor’s reputation.
Data Use and Model Training
Determine whether prompts, files, outputs, connected data or feedback train or improve models—and whether rules differ by account, feature or model.
Storage, Retention and Deletion
Identify what is retained, configuration options, backup treatment, deletion timing and whether records can be exported.
Access and Account Security
Review SSO, MFA, administration, provisioning, agent identities, secret management, sessions and activity logs.
Permissions and Connected Systems
Inventory every connector, plugin, application and data source. Separate read from create, update, send and delete authority.
Security Evidence
Request current security documentation, appropriate independent assessments, vulnerability practices, encryption and continuity information.
Models and Supply Chain
Identify model providers, subprocessors, hosting dependencies, processing locations, notification practices and service dependencies.
Output and Action Controls
Review prompt-injection defenses, harmful-output controls, tool restrictions, approval, rate limits, rollback and capability shutdown.
Contract and Ownership
Clarify inputs, outputs, confidentiality, liability, intellectual property, security obligations, audit rights and regulated uses.
Incident Notification and Support
Define covered events, notification timing, emergency support, log availability, disabling access and remediation coordination.
Exit, Portability and Fallback
Plan how to export records, remove identities, revoke permissions, confirm deletion and replace the workflow if the service fails or changes.
NIST: AI Risk Management Framework → CISA: Secure by Design → NIST: Generative AI Profile →
Make a Documented Vendor Decision
Approve the use case, account type and configuration—not merely the vendor.
Approved
The reviewed use is acceptable under current controls and terms.
Approved With Conditions
Use is allowed only with documented limits, configuration, training or approval.
Pilot Only
Use is limited to defined users, data and time while evidence is gathered.
Do Not Approve
Risk, uncertainty or contract terms exceed the business’s tolerance.
Retire
Stop existing use and remove access, integrations and retained business data.
Record the outcome, conditions, evidence, approver and next review date in the AI Tool Inventory.
Assign Incident Ownership Before an Incident
A contract describes responsibilities between organizations. It does not automatically assign your internal decision-makers. One person may fill several roles, but every role must be explicit.
Incident Coordinator
Organizes the response and maintains the timeline.
Technical Owner
Disables access, preserves logs and investigates systems.
Business Owner
Assesses operational, employee and client impact.
Security or Privacy Lead
Evaluates data exposure and security obligations.
Legal or Compliance Contact
Evaluates notification, contract and regulatory requirements when appropriate.
Communications Owner
Coordinates accurate internal, client or public messaging.
Vendor Contact
Escalates with the provider and tracks its response.
What Counts as an AI Incident?
An incident is not limited to a vendor announcing a breach. Encourage employees to report uncertainty quickly.
- Sensitive data entered into an unapproved tool
- Unexpected exposure through AI search or retrieval
- An agent acting outside its intended scope
- Compromised accounts, tokens, connectors or agent identities
- Harmful output used in a business decision
- Fabricated information causing client, financial or legal impact
- Unauthorized model, vendor or configuration changes
- Loss of logging, approval or safety controls
- Vendor outage disrupting a critical process
- Violation of client terms, confidentiality or company policy
First-Response Questions
Connect AI incidents involving personal information to the business’s existing breach-response process.
- What happened, and when?
- Which tool, account, model, agent or integration was involved?
- Which user initiated the activity?
- What data was entered, accessed, generated, disclosed or changed?
- Which people, clients, systems or records may be affected?
- Is the activity still occurring, and can it be disabled safely?
- What prompts, outputs, logs and other evidence must be preserved?
- Which vendor, insurer, legal or regulatory contacts may be required?
- What temporary business process will replace the affected tool?
Preserve Evidence
Do not delete prompts, outputs or logs to make the problem disappear. Preserve evidence while limiting continued exposure.
Review Vendors Continuously
Annual review may be enough for a low-impact tool. Higher-risk or rapidly changing services need more frequent and event-driven review.
Vendor Changes
Terms, retention, training practices, models, agents, connectors, subprocessors or independent evidence changes.
Business Changes
Users, data, permissions, owners, critical workflows or action capabilities expand or change.
Risk Events
A security or privacy incident occurs, the service becomes difficult to replace or existing assumptions fail.
Make AI Vendor Decisions With Clear Ownership
Vendor review should produce a documented business decision—not a pile of unanswered questionnaires. PCC helps Bay Area businesses evaluate the identity, permissions, integration and security foundations behind responsible AI use.