Scenario: Pharma/Validated Systems

1. GOVERNANCE / AUDITABILITY (Critical — 98% Weight)

Why Dominant: FDA 21 CFR Part 11 mandates electronic record integrity, audit trails, and validation. Any system that generates or processes regulatory data must be validated (IQ/OQ/PQ). Lack of audit trail = non-compliance = potential Warning Letter or consent decree.

Implication: LLM system must be treated as a validated computerized system. On-premise gives you control over validation documentation. API vendors rarely provide validation support packages.

2. DATA LOCALITY / PRIVACY (Critical — 95% Weight)

Why Dominant: PHI under HIPAA requires Business Associate Agreements (BAA). Clinical trial data often subject to confidentiality agreements. EU GDPR adds data residency constraints. Unauthorized disclosure = civil penalties + criminal liability.

Implication: API vendors must sign BAAs and prove HIPAA compliance. On-premise eliminates third-party risk but requires internal HIPAA controls.

3. OPERATIONAL COMPLEXITY (High — 75% Weight)

Why Dominant: Validation lifecycle requires strict change control. Every model update = revalidation (or documented risk assessment under change control). Ops must be GxP-compliant (SOPs, training records, deviation handling).

Implication: On-premise ops must follow validated procedures. API models updating without notice create validation nightmares. Frozen model versions may be required.

4. LATENCY CONTROL (Moderate — 40% Weight)

Why Considered: Most GxP use cases are human-in-the-loop (document review, Q&A). Sub-second latency not typically required. However, batch processing (e.g., 1000 case reports) benefits from predictability.

Implication: API latency variability is tolerable if within validated bounds (e.g., < 5 sec P99). On-premise offers more control but not always necessary for latency alone.

5. COST PREDICTABILITY (Moderate — 50% Weight)

Why Considered: Regulatory compliance budgets are significant but finite. Budget predictability helps long-term planning for validation maintenance and revalidation cycles.

Implication: On-premise CapEx can be amortized. API costs are easier to justify initially but may escalate. Validation costs (IQ/OQ/PQ) apply regardless of deployment model.

1. Unvalidated System Use

Scenario: LLM deployed in production without IQ/OQ/PQ. Used to generate regulatory submissions or batch records.
Consequence: FDA Form 483 observation, Warning Letter, product recalls if data integrity questioned.
Mitigation: Treat LLM as GxP-critical system from day 1. Perform validation before production use. Maintain validation master file.

2. Uncontrolled Model Updates

Scenario: API vendor updates model version without notice. Output format changes. Validation basis invalidated.
Consequence: System out of validated state. Must halt use until revalidation. Delays to submissions/operations.
Mitigation: Negotiate frozen model versions in contract, or implement on-premise with strict version control and change management.

3. PHI/PII Exposure

Scenario: LLM processes patient data without adequate de-identification. Data sent to API without BAA. HIPAA breach occurs.
Consequence: HHS OCR investigation, civil penalties ($100-$50K per record), mandatory breach notification, reputational damage.
Mitigation: Verify BAA in place before API use. Implement de-identification if required. Conduct privacy impact assessment.

4. Audit Trail Gaps

Scenario: LLM logs do not capture user ID, timestamp, or action details required by Part 11. Auditor cannot trace who generated what output.
Consequence: Data integrity concerns, inspection findings, product disposition challenges.
Mitigation: Design audit logging per Part 11 requirements. Include user authentication, prompt/response capture, timestamp (UTC), system ID.

□ Regulatory Validation

• Validation plan approved (defines IQ/OQ/PQ scope)
• User requirements specification (URS) documents intended use
• Risk assessment completed (FMEA or equivalent)
• Traceability matrix links requirements to test cases
• Validation report signed off by Quality Assurance

□ Data Privacy & Security

• BAA executed if using API vendor (HIPAA)
• Data classification performed (PHI, PII, confidential)
• Encryption at rest and in transit verified
• Access controls implemented (role-based, MFA)
• Privacy impact assessment documented

□ Change Control & Audit Trail

• Change control SOP updated to include AI systems
• Audit trail captures: user, timestamp, action, data
• Audit log review procedure defined (frequency, responsibility)
• Model versioning tracked in configuration management
• Electronic signature capability if generating records

□ Operational Procedures

• Standard Operating Procedures (SOPs) written
• User training completed and documented
• Incident/deviation handling process defined
• Periodic review schedule established (e.g., annual)
• Disaster recovery / business continuity plan includes LLM

On-Premise Only is the most validation-friendly when:
• Full control over model versioning required
• Validation documentation must be comprehensive
• Change control process is already mature
• Data cannot leave validated environment (no BAA available)

Hybrid or API-Only may be viable when:
• Vendor provides validation support package (IQ/OQ/PQ templates)
• Frozen model version guaranteed in contract (no auto-updates)
• BAA executed and vendor is HIPAA-compliant (if PHI involved)
• Risk assessment concludes vendor reliability is acceptable
• Use case is non-GxP-critical (e.g., literature search only)

API-Only is high-risk when:
• Vendor cannot provide validation documentation
• Model updates occur without customer control
• Audit trail does not meet Part 11 requirements
• No BAA available for PHI use cases
• System generates or modifies GxP records directly