# Vendor AI Example This example shows how the AI Control Architecture can be applied to a vendor AI feature inside an enterprise SaaS platform. This is a filled example, not a blank template. --- # 1. Use Case Summary ## Use Case Name ```text AI-Enabled Customer Support Case Summary ``` ## Description ```text The enterprise is considering enabling a vendor-provided AI feature inside the customer support SaaS platform. The feature summarizes customer support cases, suggests next steps, drafts internal notes, and recommends possible responses for support agents. The feature is embedded inside an existing SaaS platform already used by the customer support team. ``` ## Business Purpose ```text Reduce time spent reviewing long support case histories, improve support agent productivity, and improve consistency of customer case summaries. ``` ## AI Pattern ```text [ ] Copilot [ ] Internal LLM application [ ] RAG system [x] AI-enabled SaaS [x] Embedded vendor AI [ ] Agent [ ] AI-enabled workflow automation [x] Customer-facing AI, if drafted responses are sent externally [ ] Employee-facing AI [ ] Developer AI tool [ ] Security operations AI [x] Decision-supporting AI [ ] Action-capable AI, unless auto-send or workflow actions are enabled ``` ## Lifecycle Status ```text Vendor review / pre-enablement assessment ``` --- # 2. Ownership ## Business Owner ```text Name: Head of Customer Support Function: Customer Operations Responsibility: Owns business purpose, support workflow impact, agent adoption, and customer communication risk. ``` ## Technical Owner ```text Name: Customer Support Platform Owner Function: Business Applications Responsibility: Owns SaaS configuration, feature enablement, access controls, integration settings, and technical remediation. ``` ## Data Owner ```text Name: Customer Data Owner Function: Data Governance / Customer Operations Responsibility: Owns customer data classification, access approval, retention expectations, and data exposure review. ``` ## Vendor Owner ```text Name: Customer Support SaaS Vendor Manager Function: Procurement / Vendor Management Responsibility: Owns vendor relationship, vendor AI evidence, contract review, and vendor remediation tracking. ``` ## Privacy / Legal Owner ```text Name: Privacy Counsel Function: Legal / Privacy Responsibility: Reviews customer personal data processing, retention, cross-border transfer, contractual terms, and notification obligations. ``` ## Incident Contact ```text Name: Customer Operations Incident Lead Function: Customer Operations Responsibility: Coordinates response if AI-generated summaries or responses cause customer impact, privacy exposure, or evidence issues. ``` --- # 3. Initial Risk Tier ## Assigned Tier ```text Tier 3: Decision-supporting AI ``` ## Tier Rationale ```text The vendor AI feature processes customer support case data and generates summaries, next-step suggestions, and draft customer responses. In the proposed configuration, the feature does not automatically send customer communications, close cases, issue refunds, change entitlements, or trigger workflows. The AI output may influence support agent judgment and customer communication, so Tier 3 applies. Risk increases to Tier 4 if the feature is allowed to send responses, update case status, trigger workflows, issue credits, change customer records, or perform other actions. ``` ## Risk Drivers ```text [x] Vendor AI [x] Customer data [x] Customer support records [x] Personal data possible [x] Draft customer communication [x] Decision support [x] Evidence dependency on vendor [ ] Autonomous action [ ] Financial transaction [ ] Production system change ``` --- # 4. Vendor AI Feature Description ## Vendor Feature ```text AI case summarization and response suggestion inside customer support SaaS platform. ``` ## Vendor-Provided Capabilities ```text Summarize case history Summarize customer sentiment Identify possible next steps Draft internal case notes Draft customer response suggestions Suggest knowledge articles Suggest case category Suggest case priority ``` ## Proposed Configuration ```text Feature enabled for limited pilot group only. AI may generate summaries and draft responses. Human support agent must review, edit, and approve any customer-facing response. Auto-send is disabled. Automatic case closure is disabled. Automatic refund, credit, entitlement, or account change actions are disabled. ``` --- # 5. What Can AI See? The vendor AI may process: ```text Customer support case text Customer messages Support agent replies Internal case notes Case metadata Customer account identifiers Product information Knowledge base articles Case history Attachments if enabled ``` The vendor AI must not process unless separately approved: ```text Payment card data Government identifiers Authentication secrets Passwords Access tokens Highly sensitive personal data Legal privileged content Security incident details Customer contractual documents Medical or safety data ``` --- # 6. Data Boundary ## Data Classification ```text Highest expected classification: Confidential / Customer data / Personal data possible ``` ## Approved Data Use ```text Generate case summaries Generate internal draft notes Suggest support response drafts Suggest relevant knowledge articles Suggest category or priority for agent review ``` ## Prohibited Data Use ```text Vendor training on enterprise customer data unless explicitly approved Vendor product improvement using prompts, outputs, or case data unless explicitly approved Retention beyond approved contractual period Processing of payment card data, secrets, or highly sensitive personal data Use of AI output as final customer commitment without agent approval ``` ## Data Boundary Controls ```text [x] Vendor data processing review required [x] Retention review required [x] Training/reuse review required [x] Customer data classification required [x] Admin setting review required [x] Pilot user group restriction required [ ] Attachment processing disabled or separately reviewed [ ] Sensitive data detection tested ``` --- # 7. What Can AI Decide? The feature may influence: ```text Support agent understanding of case history Suggested next step Suggested response wording Suggested case category Suggested priority Suggested knowledge article ``` The feature must not decide: ```text Final customer response Refund approval Credit approval Contractual commitment Entitlement change Case closure Legal position Complaint resolution outcome Regulatory response ``` ## Decision Boundary ```text AI-generated summaries and suggestions are advisory. A human support agent remains responsible for reviewing, editing, approving, and sending customer communications. Support managers remain responsible for escalations, refunds, credits, complaints, and customer-impacting decisions. ``` --- # 8. What Can AI Do? ## Approved Capabilities ```text Generate case summary Generate internal draft note Suggest response Suggest priority Suggest category Suggest knowledge article ``` ## Not Approved ```text Auto-send response to customer Auto-close case Auto-issue refund or credit Auto-change entitlement Auto-update customer record Auto-trigger escalation workflow Auto-commit to service terms Auto-delete case data ``` ## Tool / Action Capability ```text Current status: Draft and recommendation only. No autonomous or write-capable action is approved for pilot. ``` --- # 9. Vendor Assessment ## Vendor Processing Questions | Question | Status | |---|---| | Does vendor process customer case data? | Yes | | Does vendor process prompts? | Yes | | Does vendor process AI outputs? | Yes | | Does vendor retain prompts or outputs? | Pending vendor confirmation | | Does vendor use data for model training? | Must be disabled or contractually prohibited | | Does vendor use data for product improvement? | Pending vendor confirmation | | Are subprocessors involved? | Pending vendor confirmation | | Is processing region known? | Pending vendor confirmation | | Are admin controls available? | Yes, under review | | Are usage logs available? | Partial, under review | | Are prompt/output logs available? | Unknown | | Can feature be disabled by admin? | Yes | | Is incident support path defined? | Pending | ## Vendor Evidence Required ```text Vendor AI feature documentation Data processing terms Retention terms Training/reuse terms Subprocessor list Processing location Admin configuration guide Logging documentation Evidence export capability Incident support process Contractual AI terms Security assurance report where available ``` --- # 10. Prompt and Input Controls ## Allowed Inputs ```text Routine customer support case content Product support questions Troubleshooting notes Approved knowledge base content Support agent prompts related to active case ``` ## Prohibited Inputs ```text Payment card data Passwords API keys Access tokens Government identifiers Highly sensitive personal data Legal privileged content Security incident details Contractual negotiations Medical or safety data Prompt injection instructions ``` ## Input Controls ```text [x] User guidance required [x] Prohibited input categories defined [x] Attachment processing disabled until reviewed [x] Prompt injection risk acknowledged [ ] Sensitive data detection tested [ ] Prompt injection testing completed ``` --- # 11. Output Controls ## Output Types ```text Case summary Internal note draft Customer response draft Next-step recommendation Category suggestion Priority suggestion Knowledge article suggestion ``` ## Output Rules ```text AI-generated summaries must be reviewed before being copied into official case notes. AI-generated customer responses must be reviewed and edited by a human support agent before sending. AI must not make contractual commitments, refund commitments, legal statements, or regulated claims. AI must escalate ambiguous, complaint-related, legal, privacy, security, or high-impact customer issues. ``` ## Output Risks ```text Incorrect case summary Missing important customer detail Overconfident customer response Unauthorized commitment Sensitive data included in draft response Incorrect priority or category Agent overreliance ``` ## Output Controls ```text [x] Human review required [x] Customer response auto-send disabled [x] Case closure disabled [x] Refund/credit actions disabled [x] Output use restrictions included in guidance [ ] Output quality sampling required during pilot [ ] Escalation trigger testing required ``` --- # 12. Human Accountability ## Accountability Statement ```text The vendor AI feature may summarize, suggest, or draft, but it does not own the customer outcome. Support agents remain accountable for reviewing and approving customer communications. Support managers remain accountable for escalations, complaints, refunds, credits, and customer-impacting decisions. The business owner remains accountable for customer operations impact. The vendor owner remains accountable for vendor evidence and remediation tracking. ``` ## Human Review Model ```text Human-in-the-loop for customer-facing responses. Human-in-the-loop for generated case notes before they become official record. Human-on-the-loop for summary and category suggestion monitoring during pilot. ``` --- # 13. Monitoring and Evidence ## Evidence Required Before Pilot ```text AI inventory record Risk assessment Vendor assessment Vendor data processing evidence Retention and training/reuse evidence Admin configuration record Pilot user list Output use guidance Logging availability review Incident escalation path Feature disablement path Pilot approval ``` ## Evidence Required During Pilot ```text Usage metrics Sample AI summaries Sample AI draft responses Reviewer feedback Output correction records Customer complaint records if any Policy violation records Vendor issue records Support agent feedback Open findings and remediation actions ``` ## Logging Approach ```text Usage metadata and feature activity logs should be enabled. Prompt/output content logging requires privacy/legal/data approval. If vendor cannot provide prompt/output logs, the evidence limitation must be documented and considered in the risk decision. ``` --- # 14. Assurance Tests ## Required Tests Before Pilot ```text [ ] Admin configuration review [ ] Feature disablement test [ ] Pilot user restriction test [ ] Auto-send disabled test [ ] Auto-close disabled test [ ] Refund/credit action disabled test [ ] Sensitive data handling test [ ] Prompt injection test [ ] Output quality sample test [ ] Escalation trigger test [ ] Logging availability test [ ] Vendor evidence export test ``` ## Example Test Cases | Test Case | Expected Result | |---|---| | AI drafts customer response with refund promise | Response is flagged or requires human edit; no auto-send. | | Agent asks AI to close case | Feature does not close case. | | Case contains payment card data | Data is masked, blocked, or escalated according to policy. | | Prompt asks AI to ignore policy | AI does not follow policy-bypass instruction. | | Admin disables AI feature | Feature becomes unavailable to users. | | Audit requests usage evidence | Usage evidence can be exported or evidence gap is documented. | --- # 15. Incident Containment ## Potential Incident Scenarios ```text AI generates incorrect customer response AI includes sensitive data in draft output AI summary omits critical customer detail AI suggests unauthorized refund or commitment Vendor retains customer prompts unexpectedly Vendor logs unavailable during incident Feature enabled for unapproved users Customer complaint references AI-generated response ``` ## Containment Actions ```text Disable vendor AI feature Restrict pilot user group Stop use for affected case types Preserve available logs Collect sample outputs Notify business owner Notify privacy/legal/security where required Contact vendor support Correct customer communication where required Review affected cases Update guidance and controls Retest before restart ``` ## Restart Criteria ```text Incident contained Affected customers identified where required Evidence preserved Root cause understood Vendor response received where required Controls updated Assurance retest completed Business owner approval obtained Privacy/legal approval obtained where required ``` --- # 16. Open Issues | Issue | Owner | Due Date | Status | |---|---|---|---| | Confirm prompt/output retention | Vendor Owner | TBD | Open | | Confirm training/reuse settings | Vendor Owner | TBD | Open | | Confirm subprocessor involvement | Vendor Owner | TBD | Open | | Confirm prompt/output log availability | Vendor Owner | TBD | Open | | Test feature disablement | Technical Owner | TBD | Open | | Complete privacy/legal review | Privacy / Legal Owner | TBD | Open | | Finalize support agent guidance | Business Owner | TBD | Open | | Complete output quality sampling plan | Assurance Owner | TBD | Open | --- # 17. Decision ## Pilot Decision ```text Not yet approved. Pending vendor evidence and privacy/legal review. ``` ## Conditions for Pilot Approval ```text 1. Vendor must confirm whether prompts, outputs, and case data are retained. 2. Vendor must confirm training and product improvement use are disabled or contractually prohibited. 3. Processing location and subprocessors must be reviewed. 4. Admin configuration must be reviewed and documented. 5. Auto-send, auto-close, refund, credit, and entitlement actions must be disabled. 6. Feature disablement must be tested. 7. Pilot user group must be restricted. 8. Support agent guidance must be published. 9. Logging and evidence limitations must be documented. 10. Privacy/legal approval must be obtained before pilot. ``` --- # 18. Summary ```text Use case: AI-Enabled Customer Support Case Summary Pattern: AI-enabled SaaS / Embedded vendor AI / Decision-supporting AI Risk tier: Tier 3 initially; Tier 4 if action capability is enabled Business owner: Head of Customer Support Technical owner: Customer Support Platform Owner Vendor owner: Customer Support SaaS Vendor Manager Data owner: Customer Data Owner Data boundary: Customer support case data only; sensitive categories restricted Decision impact: Support agent decision support and draft customer communication Tool/action capability: Draft/recommendation only; no auto-send or case closure Key risks: Vendor retention, training/reuse, customer data exposure, incorrect response, missing logs Key controls: Vendor assessment, admin config review, human approval, output restrictions, evidence review, incident path Decision: Not approved until vendor evidence and privacy/legal review are complete ```