Hook

Employee requests ChatGPT Enterprise: "We need this for customer support." Security asks 15 questions: data retention policy, SOC 2 compliance, sub-processor list, data residency, encryption details. Legal reviews the ToS for 2 weeks. Procurement negotiates pricing. Finance approves the budget. Two months later, ChatGPT is approved. But the requester has already been using the free version for 6 weeks (shadow AI), and the team's process has adapted to it. The approval comes too late to shape behavior. Worse: the team requested Claude for a different use case. Same 2-month process. By the time both tools are approved, the team has given up and gone back to shadow tools. The approval workflow became a bottleneck instead of a governance enabler. Now flip the scenario: tools are categorized by risk. Low-risk tools (Gemini, Claude, both with good privacy policies) are approved in 2 days via a streamlined checklist. High-risk tools get the full security review. The approval time matches the urgency of business need. Security gets the governance they need; business gets speed.

Purpose

An AI tool approval workflow is a structured process: request submission → security/privacy review → risk assessment → approval/denial → provisioning → monitoring. Done right, it enables business innovation while controlling risk. Done poorly, it becomes a bottleneck that drives shadow AI.

This lesson teaches you to design workflows that actually work: risk-based approval (fast-track for low-risk, detailed review for high-risk), clear criteria, human decision points, SLA targets, and feedback loops. You'll learn when to automate (routine checks) and when to require manual review (judgment calls).

Why This Matters

An approval workflow has competing goals:

Speed: Slow approvals drive shadow AI. If the approval process takes 8 weeks, employees use unapproved tools. You can't govern what you can't see.

Security and compliance: Unapproved tools without security vetting introduce risks. A tool with a data retention policy that allows training on customer data is a compliance violation.

Consistency: Approvals should follow consistent criteria, not depend on who reviews the request.

User satisfaction: If the approval process feels arbitrary or unfair (one team gets approval, another doesn't), trust erodes.

A well-designed workflow balances these. It's fast for low-risk requests, thorough for high-risk, and transparent about criteria.

Core Concepts

Key insight: Risk-Based Approval Tiers Determine Process Depth

Not all tools require the same review depth. Tiering speeds the process.

Tier 1 (Fast-Track, 1-2 days)

Applicable to: Low-risk tools with clear security posture. Low data sensitivity.

Criteria:

• Tool has SOC 2 Type II certification (or equivalent).
• Data retention policy explicitly states: customer data is not used for training, not retained after service termination.
• Encryption in transit and at rest.
• Audit logging available.
• No significant security advisories or breaches.
• Use case is low-risk (brainstorming, writing, learning, non-critical analysis).

Examples: Claude (with enterprise agreement), ChatGPT Enterprise (with data non-retention), Copilot with organization account.

Review process:

• Requester submits tool name, use case, team size.
• Automated check: Is tool on pre-approved list? If yes, auto-approve.
• If not pre-approved, send to risk team for 30-min review against checklist.
• Decision: Approve or escalate to Tier 2.

Tier 2 (Standard, 2-3 weeks)

Applicable to: Moderate-risk tools, or low-risk tools used for sensitive data.

Criteria:

• Tool has reasonable security controls (encryption, audit logging).
• Data retention policy is acceptable (not perfect, but defensible).
• No major compliance concerns.
• Use case involves internal or confidential data (not restricted/PII).

Examples: Perplexity, Gemini, open-source models with enterprise support.

Review process:

• Requester submits: tool name, use case, team size, data classification, expected upload frequency.
• Security review (5 days): Assess security controls, data handling, privacy policy.
• Privacy review (5 days): Assess GDPR compliance, data residency, data retention.
• Legal review (5 days): Review ToS for acceptable terms. Any contract negotiation needed?
• Risk classification meeting (2 days): Weigh risks vs. benefits. Approve, approve with conditions, or deny.

Tier 3 (Deep Review, 6-8 weeks)

Applicable to: High-risk tools, tools with sensitive data, custom/internal tools, tools that require contract negotiation.

Criteria:

• Tool has security concerns or unclear policies.
• Use case involves restricted data (PII, health records, financial data).
• Tool requires custom contract or SLA.
• Tool has significant cost or long-term commitment.

Examples: Specialized medical AI tools, custom LLMs, vendor AI products without standard terms.

Review process:

• Requester submits: detailed requirements, use case, data flows, security requirements.
• Security deep-dive (2 weeks): Penetration testing considerations, vulnerability assessment, threat modeling.
• Privacy assessment (2 weeks): GDPR/CCPA impact assessment, data processing agreement needed.
• Legal negotiation (2 weeks): Contract review, custom terms, liability, indemnification.
• Executive review (1 week): Cost-benefit analysis. Is the tool worth the risk/cost?

Key insight: Checklist-Driven Reviews Are Faster and More Consistent

Structured checklists reduce review time and bias.

Sample security checklist (Tier 1 fast-track):

Tool: _______________ Date: _______________

Security Assessment:
[ ] Tool has SOC 2 Type II certification (or ISO 27001 or equivalent)
Evidence: ________________
[ ] Encryption in transit: TLS 1.2+ or higher
Evidence: ________________
[ ] Encryption at rest: AES-256 or equivalent
Evidence: ________________
[ ] Audit logging available to customers
Evidence: ________________
[ ] Data retention policy publicly available
Evidence: ________________
[ ] Customer data is NOT used for AI model training
Evidence: ________________
[ ] No breaches reported in past 3 years
Evidence: ________________
[ ] No high-severity security advisories
Evidence: ________________

Privacy Assessment:
[ ] GDPR-compliant (if EU customers)
Evidence: ________________
[ ] Data processing agreement (DPA) available
Evidence: ________________
[ ] Data residency: acceptable region(s)
Evidence: ________________

Operational Assessment:
[ ] Uptime SLA defined (99.5% or better)
Evidence: ________________
[ ] Support available during business hours
Evidence: ________________
[ ] API available for integration (if needed)
Evidence: ________________

Use Case Assessment:
[ ] Intended data classification: _____
[ ] Intended team size: _____
[ ] Expected data volume: _____
[ ] Risk level acceptable for stated use case: YES / NO

Decision: [ ] Approve [ ] Escalate to Tier 2 [ ] Deny

Reviewer: ________________ Date: ________________

Benefits of checklists:

• Reviewers don't forget questions.
• Decisions are documented and auditable.
• Requester knows what's being evaluated.
• Consistency across approvals (same checklist used for all Tier 1 requests).

Key insight: Pre-Approved Lists Accelerate Low-Risk Decisions

Many tools are low-risk and commonly requested. Establish a pre-approved list.

Pre-approved AI tools (example list):

Tool: Claude (Anthropic)
├─ Tier: 1 (Fast-track)
├─ Approval date: 2024-01-15
├─ Approved for: General writing, brainstorming, code review, learning
├─ NOT approved for: Customer PII, financial data without encryption, proprietary algorithms
├─ License: Enterprise account with data non-retention agreement
├─ SLA: 99.9% uptime, support during business hours
├─ Data retention: Customer data not retained after session; not used for model training
├─ Cost: $30/month per seat
├─ Renewal date: 2025-01-15
├─ Provisioning: Anthropic SSO integration, email domains: @company.com
└─ Notes: Preferred tool for enterprise use. Low data risk. Encourage adoption over unapproved alternatives.

Tool: GitHub Copilot
├─ Tier: 1 (Fast-track)
├─ Approved for: Code completion, coding assistance, learning new libraries
├─ NOT approved for: Non-code documents, customer-facing content
├─ Cost: $20/month per engineer (includes Copilot, Copilot Chat, Copilot for Docs)
├─ Provisioning: GitHub organization account with SSO
└─ Notes: Specialized for code. Approved for Engineering team. Other teams may request if needed.

Tool: Gemini (Google)
├─ Tier: 1 (Fast-track)
├─ Approval date: 2024-02-01
├─ Approved for: Similar to Claude
├─ NOT approved for: Same restrictions
├─ License: Google Workspace integration
├─ Data retention: Google does not retain data for model training; complies with GDPR
└─ Notes: Good alternative to Claude. Deploy via Google Workspace if org uses Google products.

Tool: ChatGPT
├─ Tier: 2+ (Requires detailed review; free tier not approved)
├─ Approved variant: ChatGPT Enterprise
├─ Approval conditions:
├─ Data non-retention agreement required
├─ DLP integration must block uploads of PII, financial data
├─ Usage monitoring dashboards required
└─ Quarterly compliance audits required
├─ Cost: $30/month per seat (enterprise tier)
├─ Notes: Free tier has less favorable data policies. Enterprise tier has defensible terms.

Benefits:

• Requester sees tool status instantly.
• Common tools don't go through full review process.
• But list is not open-ended (provides governance).

Key insight: Clear Approval Criteria Reduce Judgment Calls

Vague criteria ("Is this secure?") lead to inconsistent decisions. Clear criteria reduce back-and-forth.

Good approval criteria:

• Testable (yes/no checks).
• Transparent (requester knows what will be evaluated).
• Consistent (same tool always judged by same criteria).

Bad criteria:

• Subjective ("Does this feel safe?").
• Vague ("Reasonable security controls").
• Hidden (requester doesn't know what's being evaluated).

Example: Bad criteria

"The tool must be secure and compliant."

Requester: "Is Claude compliant?"

Reviewer: "Define 'compliant.' For what industry? What regulations?"

Requester: "GDPR."

Reviewer: "It has a DPA. But does it cover all sub-processors?"

Requester: "Sub-processors? I don't know."

Reviewer: "You need to know this before we can decide."

Requester: "Can you tell me what I need to check?"

Reviewer: "Ask the vendor."

[Process stalls.]

Good criteria

"Tool must have SOC 2 Type II certification, GDPR-compliant DPA, and data non-retention clause."

Requester: "Is Claude approved?"

Reviewer: [Checks certification site] "Yes, has SOC 2 Type II. [Checks Anthropic website] Yes, DPA available. [Checks ToS] Yes, data non-retention clause in enterprise agreement. Approved."

[Process completes in 30 minutes.]

Key insight: Conditions Can Approve High-Risk Requests Without Blocking Them

Sometimes a tool is high-risk for unrestricted use, but acceptable with conditions.

Conditional approval examples:

• "ChatGPT is approved for marketing team only. DLP integration required to prevent accidental uploads of customer data. Monthly audits required."
• "Perplexity approved for brainstorming; prohibited for proprietary data analysis."
• "Custom LLM approved for this project. Data residency must be US. Quarterly security audits required."

Conditions allow teams to use the tool while reducing risk.

Key insight: Feedback Loops Improve the Process

Track approval metrics:

• Average approval time by tier.
• Approval/denial rates.
• Tools that are approved but rarely used (why?).
• Tools that are denied but requested multiple times (should the criteria change?).

Use feedback to optimize:

• Is Tier 1 fast-tracking actually fast? If reviews are taking 5 days instead of 1 day, process is broken.
• Are denials justified? If 80% of denied tools are later used as shadow AI, criteria might be too strict.
• Is the pre-approved list helpful? If 70% of requests are for pre-approved tools, expand the list.

Practical Use Cases

Before/After: Slow vs. Fast Approval Process

Before (No Risk-Based Tiers):

• Employee requests ChatGPT Enterprise for customer support.
• All requests go through same 6-week process:
• Week 1: Security review.
• Week 2: Privacy review.
• Week 3: Legal review.
• Week 4: Procurement.
• Week 5: Budgeting.
• Week 6: Final approval.
• By week 2: Employee has already set up personal account and started using ChatGPT free tier (shadow AI).
• Week 6: ChatGPT Enterprise is approved, but team is already adapted to free tier. "We'll migrate next month" (they don't).
• Result: Approval came too late to prevent shadow AI.

After (Risk-Based Tiers):

• Employee requests ChatGPT Enterprise.
• Recognized as Tier 2 (moderate-risk, some data sensitivity).
• Security review (5 days): Checks SOC 2, DPA, data retention. Passes.
• Privacy review (5 days): Checks GDPR compliance, sub-processors. Passes.
• Legal review (5 days): ToS is standard. No negotiation needed. Passes.
• Total: 2-3 weeks.
• Decision point: Before employee sets up shadow AI.
• ChatGPT Enterprise is approved and provisioned within 3 weeks.
• Team adopts approved tool because it's available when needed.

Before/After: Consistency Issues in Approvals

Before (No Checklist):

• Marketing team requests Gemini. Reviewer says "No, unclear data policy." Denied.
• Engineering team requests Gemini 2 weeks later. Different reviewer. Same tool. "Looks fine to me. Approved."
• Marketing team is frustrated. "Why did Engineering get approved and not us?"
• Reviewer 1 is more cautious; Reviewer 2 is more lenient. Decisions are inconsistent.

After (Structured Checklist):

• Both requests go through same checklist.
• Both reviewers check: SOC 2 status, data retention, encryption, GDPR compliance.
• Both find the same facts about Gemini.
• Both make the same decision: Approved.
• If one reviewer thinks data policy is unclear, they document it and escalate. Both requests get escalated together.
• Consistency improves. Fairness improves.

Before/After: Conditional Approval Enables Business Need

Before (Binary Approve/Deny):

• Customer support team requests ChatGPT free tier.
• Security team: "Data retention policy allows training on customer data. Unacceptable. Denied."
• Team: "But ChatGPT is amazing for customer support. Speeds up response time 3x."
• Security team: "Too risky. Denied."
• Team: "Fine, we'll use it anyway" [Shadow AI].
• Result: Team is frustrated. Shadow AI is used anyway. Governance fails.

After (Conditional Approval):

• Same request.
• Security team: "Free tier is risky due to data retention policy. But ChatGPT Enterprise has better terms. Approved with conditions: (1) Must use ChatGPT Enterprise, not free tier. (2) DLP integration to prevent accidental uploads of customer data. (3) Team training on data handling. (4) Quarterly compliance audits."
• Team: "We can work with that. Approve ChatGPT Enterprise."
• Business need is met. Risk is reduced. No shadow AI.

Examples

Example 1: Tier 1 Fast-Track Approval (Same-Day Decision)

Request: Approval for Claude Enterprise

Requester: Sarah Chen (Marketing Manager)
Request date: 2024-03-15
Tool: Claude (Anthropic) - Enterprise plan
Team size: 6 people
Use case: Content writing, brainstorming, marketing copy review
Data classification: Internal/public only (no customer data)
Expected usage: 10-20 conversations per person per week
Budget: $180/month ($30/person)

Automated pre-approval check:
├─ Is Claude on pre-approved list? YES
├─ Pre-approved for: General writing, brainstorming, code review, learning
├─ Is use case consistent with pre-approved uses? YES (writing, brainstorming)
├─ Is data classification acceptable? YES (internal/public)
├─ Risk tier: TIER 1 (Fast-track)

Automated approval:
└─ Status: APPROVED
Approved on: 2024-03-15, 9:45 AM (same-day)
Approver: Automated policy engine
Notes: Tool is pre-approved. Use case and data classification are within approved scope.

Provisioning:
├─ Email to Anthropic: Add 6 users with email domain @marketingcompany.com to Claude Enterprise organization account
├─ Email to team: "Claude Enterprise approved. Access will be available by EOB tomorrow. Training materials: [link]"
└─ Update license tracking: 6 seats, $180/month, renewal date 2025-03-15

Timeline:
├─ Approval: Same day
├─ Provisioning: Within 24 hours
├─ Team can use tool: Next day
└─ Total time from request to use: ~24 hours

Result: Speed. Security (tool was vetted before pre-approval). User satisfaction (tool available when needed).

Example 2: Tier 2 Standard Approval (3-Week Process)

Request: Approval for Perplexity Pro

Requester: Jordan Kim (Research Manager)
Request date: 2024-03-15
Tool: Perplexity Pro
Team size: 12 people
Use case: Research, competitive analysis, market trends
Data classification: Confidential (company strategy, competitive positioning)
Expected usage: 20-30 searches per person per week
Budget: $240/month ($20/person)

Initial assessment:
├─ Tool not on pre-approved list (Tier 1)
├─ Data classification is "Confidential" (requires Tier 2 review)
├─ Risk tier: TIER 2 (Standard review)

Security Review (Days 1-5, assigned to Jake Wong, Security Analyst):
├─ Check certifications: "No SOC 2, but has ISO 27001 certification (acceptable)"
├─ Check encryption: "TLS 1.3 for transit, AES-256 for at-rest (acceptable)"
├─ Check audit logging: "Available via API. Customer can pull logs. (acceptable)"
├─ Check data retention: "Perplexity does not retain customer data after session. Not used for model training. (acceptable)"
├─ Check breaches: "No major breaches. One minor incident 2022, resolved. (low risk)"
├─ Decision: PASS
├─ Summary: "Security controls are comparable to Tier 1 tools. No major concerns."

Privacy Review (Days 6-10, assigned to Priya Patel, Privacy Officer):
├─ Check GDPR compliance: "Perplexity has Data Processing Agreement. GDPR-compliant. (acceptable)"
├─ Check sub-processors: "Uses OpenAI API internally. Sub-processor is OpenAI. DPA covers sub-processors. (acceptable)"
├─ Check data residency: "Data is processed in US. EU data may be subject to data transfer restrictions. (acceptable with conditions)"
├─ Decision: PASS with conditions
├─ Summary: "GDPR-compliant with standard terms. If team operates in EU, need to review data transfer agreement separately."

Legal Review (Days 11-15, assigned to Michael Chen, General Counsel):
├─ Check ToS: "Standard SaaS terms. No concerning liability limitations. (acceptable)"
├─ Check data ownership: "Customer data is owned by customer. Perplexity does not claim ownership. (acceptable)"
├─ Check termination: "30-day notice to terminate. Data is deleted upon request. (acceptable)"
├─ Check indemnification: "Standard indemnification clause. No concerns. (acceptable)"
├─ Decision: PASS
├─ Summary: "ToS are standard and acceptable. No negotiation needed."

Risk Classification Meeting (Days 16-21, assigned to Executive Risk Committee):
├─ Security assessment: PASS (ISO 27001, encryption, no breaches)
├─ Privacy assessment: PASS with conditions (GDPR-compliant, data transfer needs review)
├─ Legal assessment: PASS (standard ToS)
├─ Business benefit: HIGH (team needs competitive research; Perplexity is effective for this)
├─ Data risk: MODERATE (confidential data, but not PII or financial; risk is manageable with controls)
├─ Cost: LOW ($240/month, negligible in budget)
├─ Decision: APPROVED with conditions
├─ Summary: "Perplexity is approved for competitive research. Conditions: (1) Team training on data sensitivity required before access. (2) DLP integration to prevent accidental uploads of PII or financial data. (3) Quarterly compliance audits. (4) If team operates in EU, complete data transfer agreement review."

Provisioning (Days 22-21):
├─ Request Perplexity: Add 12 seats for Research team
├─ Configure DLP: Set policy to block uploads containing words "financial", "revenue", "salary", "confidential" to Perplexity
├─ Schedule training: 30-min training session on data handling best practices
├─ Update license tracking: 12 seats, $240/month

Timeline:
├─ Approval process: 21 days
├─ Provisioning: 1 day
├─ Training: 1 day
├─ Total time from request to use: ~23 days
└─ Team can use tool: Within 1 month of request

Approval notification:
├─ To requester: "Perplexity has been approved with conditions. Training scheduled for [date]. Access will be available by [date]."
├─ To team: "Perplexity Pro is now approved for competitive research. Training materials and data handling guidelines are [link]. Do not upload financial information, PII, or product roadmaps."
├─ To compliance/monitoring: "Perplexity approved. DLP integration active. Quarterly audits scheduled."

Result: Thorough review. Clear conditions. Security governance. Tool available before team pursues shadow AI.

Example 3: Tier 3 Deep Review and Conditional Approval

Request: Custom Medical AI Tool

Requester: Dr. Lisa Martinez (Chief Medical Officer)
Request date: 2024-03-01
Tool: MedicalAI Pro (custom LLM for patient diagnosis support)
Team size: 25 doctors
Use case: Assist in patient diagnosis (advisory, not decision-making)
Data classification: RESTRICTED (patient PHI, HIPAA-protected)
Expected usage: 10-50 consultations per doctor per week
Budget: $50K/year

Initial Assessment:
├─ Custom tool not on pre-approved list
├─ Data classification is RESTRICTED (patient PHI)
├─ HIPAA regulated (healthcare)
├─ High business risk (patient care depends on tool)
├─ High data risk (PHI data)
├─ Risk tier: TIER 3 (Deep review, 6-8 weeks)

Security Deep-Dive (Weeks 1-2, assigned to Security Team):
├─ Vendor security assessment:
├─ Penetration testing: Vendor agrees to share 3rd-party pen test from 2023. Remediation documented. (acceptable)
├─ Vulnerability scanning: Vendor uses third-party vulnerability scanner monthly. No critical vulns. (acceptable)
├─ Threat modeling: Vendor has threat model for patient data. Reviewed and acceptable.
└─ Disaster recovery: Vendor has automated backups, tested restoration, documented RTO/RPO. (acceptable)
├─ Encryption assessment:
├─ In-transit: TLS 1.3 (excellent)
├─ At-rest: AES-256 with customer-managed keys (excellent)
├─ Key management: HSM (Hardware Security Module) for key storage. (excellent)
└─ Data isolation: Multi-tenant but with strong isolation controls. (acceptable)
├─ Audit logging:
├─ All PHI access is logged (who accessed, when, what data, why)
├─ Logs are tamper-proof and encrypted
├─ Logs are retained for 7 years (HIPAA requirement)
└─ Customer can access logs in real-time via API (excellent)
├─ Decision: CONDITIONAL PASS
├─ Summary: "Security controls meet HIPAA standards. Some areas exceed standards. Recommend approval with ongoing monitoring."

Privacy & HIPAA Assessment (Weeks 3-4, assigned to Privacy Officer + HIPAA consultant):
├─ PHI handling:
├─ Vendor is HIPAA-covered entity (medical software provider)
├─ Vendor is not a Business Associate (yet); needs to become one
├─ Business Associate Agreement (BAA) is required
├─ Standard BAA template reviewed; vendor agrees to sign (acceptable)
├─ Data retention:
├─ Patient data is retained only for current patient care episode
├─ After 90 days, data is deleted from vendor's systems
├─ But 7-year audit logs are retained (compliant with HIPAA)
├─ Sub-processors:
├─ Vendor uses AWS for hosting (AWS is Business Associate)
├─ Vendor uses third-party backup vendor (also Business Associate)
├─ BAA covers sub-processor requirements
├─ Data residency:
├─ Data is processed and stored in US (compliant for US healthcare)
├─ Specifically, data never leaves the VPC in a single AWS region
├─ Decision: CONDITIONAL PASS
├─ Summary: "Privacy compliance requires Business Associate Agreement. Once BAA is signed, HIPAA requirements are met."

Legal Deep-Dive (Weeks 5-6, assigned to General Counsel + Healthcare Attorney):
├─ Vendor contract review:
├─ Standard healthcare SaaS contract
├─ Liability cap: limited to contract value (not ideal, but standard in medical software)
├─ Indemnification: Vendor indemnifies against IP claims and HIPAA violations (acceptable)
├─ Warranty: Vendor warrants HIPAA compliance and 99.9% uptime (acceptable)
├─ Termination: 30-day notice. Data deletion upon termination with witness attestation. (acceptable)
├─ Custom negotiation needed:
├─ Security: Vendor agrees to quarterly pen testing and annual SOC 2 audit (custom requirement)
├─ Subpoenas: Vendor commits to notify customer of subpoenas within 24 hours, allows customer to seek protective order (HIPAA requirement)
├─ Incident response: Vendor commits to notify customer of breaches within 24 hours, assist with HIPAA breach notification (HIPAA requirement)
├─ Insurance: Vendor has professional liability insurance ($2M). Recommend increase to $5M for peace of mind.
└─ Decision: Negotiation recommended on insurance and audit frequency
├─ Summary: "Standard contract with custom clauses for medical/HIPAA requirements. Recommend negotiation on insurance and audit frequency."

Clinical Risk Assessment (Week 7, assigned to Chief Medical Officer + Compliance Committee):
├─ Use case review:
├─ Tool is advisory (doctor makes final diagnosis decision, not AI)
├─ Tool does not make autonomous decisions
├─ Doctors are trained on tool limitations and when not to rely on it
├─ Clinical workflow is redesigned to ensure human oversight
├─ Second-opinion process is defined (doctor consults another doctor if AI recommendation differs from clinical judgment)
├─ Liability assessment:
├─ If AI makes wrong recommendation but doctor catches it: no patient harm, no liability
├─ If AI makes wrong recommendation and doctor misses it: patient harm. Who is liable? (Doctor is liable; tool is advisory)
├─ Standard medical negligence applies (tool is a tool, not a decision-maker)
├─ Training plan:
├─ All doctors are trained on tool capabilities, limitations, and proper use
├─ Training is documented and certified
├─ Competency assessment for doctors using the tool
├─ Decision: APPROVE
├─ Summary: "Clinical workflow appropriately uses tool as advisory. Human oversight is maintained. Liability is appropriate for advisory use."

Executive Risk Committee Approval (Week 8):
├─ Risk summary:
├─ Security: PASS (exceeds HIPAA standards)
├─ Privacy: PASS with Business Associate Agreement
├─ Legal: PASS with custom negotiation on insurance and audits
├─ Clinical: PASS with appropriate training and oversight
├─ Data: RESTRICTED PHI (high risk, but mitigated by controls)
├─ Cost: $50K/year (reasonable for 25 doctors)
├─ Business value: HIGH (doctors report 20% improvement in diagnostic accuracy in pilot)
├─ Overall decision: APPROVED with conditions
├─ Summary: "MedicalAI Pro is approved for clinical diagnosis support. Conditions listed below."

Conditions for Approval:
├─ Mandatory:
├─ (1) Vendor signs Business Associate Agreement before go-live
├─ (2) Vendor completes penetration testing and provides report
├─ (3) Vendor increases professional liability insurance to $5M
├─ (4) All 25 doctors complete mandatory training and certification
├─ (5) Clinical workflow changes are reviewed and approved by Chief Medical Officer
├─ (6) Incident response plan is documented and tested
└─ (7) Quarterly compliance audits are scheduled
├─ Recommended (optional but strongly encouraged):
├─ (8) Independent security audit by third-party firm (annual)
├─ (9) Peer review publication of tool's clinical performance and limitations
└─ (10) Patient consent process for AI-assisted diagnoses (if not already in clinical workflow)

Provisioning (Timeline: Weeks 9-12):
├─ Prepare infrastructure:
├─ Set up isolated AWS environment for MedicalAI Pro
├─ Configure encryption keys and HSM
├─ Deploy audit logging
├─ Negotiate and sign agreements:
├─ Business Associate Agreement (vendor)
├─ Professional liability insurance amendment (vendor's insurance)
├─ Custom clauses on security audits (vendor)
├─ Complete conditions:
├─ Vendor provides pen testing report
├─ Vendor increases insurance
├─ Train all 25 doctors (12 hours per doctor, spread over 4 weeks)
├─ Validate clinical workflow changes
├─ Go-live:
├─ Pilot phase: 5 doctors use tool for 2 weeks; monitor and gather feedback
├─ Full rollout: 25 doctors go live; ongoing support and monitoring
├─ Ongoing monitoring:
├─ Monthly usage reports (which doctors use the tool, how often, any errors)
├─ Quarterly compliance audits
├─ Annual security audits
├─ Incident tracking (any patient harm related to tool misuse)

Timeline:
├─ Initial deep review: 8 weeks
├─ Provisioning and implementation: 4 weeks
├─ Total time from request to go-live: ~12 weeks
├─ Pilot phase: 2 weeks
├─ Full rollout: Ongoing

Final Approval Notification:
├─ To CMO: "MedicalAI Pro approved. Go-live in 12 weeks pending completion of conditions. Training plan is [link]."
├─ To doctors: "MedicalAI Pro will soon be available. Mandatory training scheduled for [dates]. Tool is advisory; clinical judgment is final."
├─ To vendors: "Contract approved. Proceed with Business Associate Agreement negotiation and insurance amendment."
├─ To compliance: "Tool approved for deployment. Quarterly audits scheduled. Incident tracking dashboard enabled."

Result: Thorough, multi-stakeholder review. Clear conditions. High-risk tool is approved with appropriate controls, not blocked. Business need is met. Risk is managed.

Anti-Patterns

Anti-Pattern 1: Approval Process Is Slower Than Shadow AI Adoption

Approval takes 8 weeks. Employees adopt unapproved tools in week 2. By the time tool is approved, new shadow AI has emerged. Governance becomes reactive instead of proactive.

Fix: Set SLA targets. Tier 1 approvals ≤ 2 days. Tier 2 ≤ 3 weeks. Tier 3 ≤ 8 weeks. If SLAs are missed consistently, streamline process.

Anti-Pattern 2: All Requests Go Through Deep Review

Every tool request goes through 6-week process, regardless of risk. ChatGPT (enterprise, low-risk) takes 6 weeks. Gemini (low-risk) takes 6 weeks. Even pre-approved tools go through full review again.

Fix: Risk-based tiering. Pre-approved list. Fast-track for low-risk.

Anti-Pattern 3: Arbitrary Approvals Without Clear Criteria

Reviewer 1 approves ChatGPT. Reviewer 2 denies Gemini (same tier, similar risk). Requester has no idea why one was approved and the other denied.

Fix: Publish clear criteria. Use checklists. Document decisions. Make appeals process transparent.

Anti-Pattern 4: Approval Process Doesn't Communicate Requirements

Tool is denied, but requester doesn't know why. "Just not secure enough." What would make it secure? Should vendor change something? Should requester use tool differently?

Fix: Provide detailed feedback. "Tool is denied because data retention policy allows training on customer data. This violates our GDPR requirements. Tool could be approved if vendor signs DPA prohibiting model training."

Anti-Pattern 5: No Monitoring After Approval

Tool is approved. Vendor's security posture deteriorates (breach, negligence). Approver doesn't notice. Tool remains approved.

Fix: Ongoing monitoring. Annual re-assessments. Incident tracking. If vendor has security incident, re-evaluate approval.

Human Judgment Checkpoints

• 
  Have you defined risk tiers that match your actual risk appetite? If you approve nothing, governance fails (shadow AI wins). If you approve everything, security fails. Find the balance.
• 
  Are your criteria measurable and transparent? "Secure" is vague. "SOC 2 Type II certified" is measurable.
• 
  Does your approval process compete with shadow AI adoption? If approval takes 8 weeks and shadow AI adoption takes 2 weeks, you'll lose. Speed matters.
• 
  Have you involved business stakeholders in approval? If IT alone decides which tools are approved, business will route around approval. Security and business should both have a voice.
• 
  Is there an appeals process? If a tool is denied, can the requester appeal? With what evidence?
• 
  Are you monitoring approved tools over time? A tool approved 2 years ago may no longer be secure. Annual re-assessments are reasonable.

Key Takeaways

• 
  Risk-based tiers speed the process: Fast-track for low-risk (1-2 days). Standard review for moderate-risk (2-3 weeks). Deep review for high-risk (6-8 weeks).
• 
  Pre-approved lists eliminate routine reviews: Establish a list of commonly-requested, low-risk tools that bypass full review. Expand list as you gain confidence.
• 
  Clear criteria reduce subjective decisions: "SOC 2 certified, GDPR-compliant DPA, data non-retention clause" is measurable. "Secure" is vague.
• 
  Checklists ensure consistency and speed: Same tool, same questions, same decision. Reviewers don't forget items.
• 
  Conditional approval enables business needs without compromising security: Don't just approve or deny. Approve with conditions (DLP integration, training, audit frequency).
• 
  SLA targets keep approval process from becoming a bottleneck: If Tier 1 takes 10 days instead of 2, process is broken. Monitor and fix.
• 
  Communication matters as much as decision: Approval feedback should be clear. If denied, explain why and what would make it approvable.
• 
  Feedback loops improve the process over time: Track approval metrics. Are Tier 1 tools actually fast? Are denials justified? Use data to improve.
• 
  Ongoing monitoring keeps approved tools safe: Annual re-assessments. Incident tracking. If vendor has security breach, re-evaluate.
• 
  Balance speed and security: Slow approval drives shadow AI. No approval drives risk. Risk-based tiers find the balance.