Compliance for AI Applications
Create a practical governance foundation for privacy, safety, and accountable AI use.
Inventory the system
Document models, providers, data sources, users, regions, tools, outputs, retention, and decisions. Include hidden components such as OCR, embeddings, analytics, support exports, and human review queues. A system map turns an abstract AI risk conversation into specific responsibilities. Update it when a model, data flow, user group, or decision purpose changes.
Classify the use case
Risk depends on context, not only model capability. A writing assistant and a system influencing employment, credit, healthcare, education, or access to services need different controls. Describe the intended benefit, affected people, possible harms, level of autonomy, and required human oversight. Do not expand a low-risk prototype into a high-impact decision workflow without a new assessment.
{
"use_case": "invoice_extraction",
"data": ["business_documents"],
"human_review": true,
"retention_days": 30,
"owner": "finance-ops",
"next_review": "2026-10-18"
}Privacy and data controls
Define purpose, lawful basis where relevant, collection limits, access, retention, deletion, provider processing, transfers, and user rights. Keep personal data out of prompts and logs when it is not necessary. Test deletion across raw files, caches, indexes, embeddings, backups, and exports. Document exceptions and assign an owner who can answer how a user’s data is handled.
Safety and accountability
Set policies for harmful content, uncertainty, human review, appeals, incident reporting, and model changes. Identify who approves the use case, who operates it, and who can stop it. Give users understandable explanations of limitations and meaningful routes to challenge an outcome. Accountability should remain with the organization deploying the system, not be shifted to an opaque model.
Vendor and model review
Assess provider security, retention, training use, subprocessors, service regions, availability, audit evidence, licensing, and change-notification practices. Record the exact model and version used for important workflows. Review open-source licenses and downloaded model provenance as carefully as hosted services. A vendor checklist should connect to actual controls rather than become a one-time procurement document.
Evidence and audits
Keep versioned policies, risk assessments, evaluation results, approvals, data-flow diagrams, training records, incident tickets, and release decisions. Store only the data needed for the audit and protect it appropriately. Evidence should show what the system was expected to do, how it was tested, what changed, and how exceptions were handled. Make records searchable before an incident occurs.
Continuous governance
Compliance is maintained through monitoring, periodic review, user feedback, incident response, and change management. Reassess when the model, prompt, retrieval data, tool permissions, geography, or decision impact changes. Schedule owners and review dates, but also create triggers for urgent reassessment. Good governance makes safe use easier to sustain because decisions and responsibilities are visible.
Worked example: reviewing invoice extraction
An invoice extractor may look low risk until its output triggers payment. The review should document the source data, purpose, people with access, retention, human approval, provider processing, and failure response. Keep extraction separate from payment authorization, and require a person or deterministic business rule to verify totals and payee changes. The use-case record is valuable because it exposes where a seemingly simple model output becomes consequential.
Code walkthrough
The JSON record captures the use case, data category, human review, retention period, owner, and next review date. Add model and provider versions, regions, subprocessors, data deletion checks, risk classification, and approval status. Validate that the retention value is enforced across source files, indexes, logs, and backups; documenting “30 days” without an automated deletion path is not a control.
Trade-offs to measure
More review and documentation reduce unmanaged risk but can slow low-impact experiments. Global controls simplify operations yet may be excessive for a harmless prototype. Vendor assurances help but do not remove the deployer’s responsibilities. Proportional governance works best: lightweight evidence for low-risk features, stronger approval, monitoring, and appeal paths for decisions affecting people or money.
Practical exercise
Complete a one-page review for a fictional invoice workflow. Map every data flow from upload to deletion, identify two possible harms, assign an owner, and define a human approval rule. Then change the workflow so it can edit bank details and repeat the assessment. Note which controls must change and what evidence you would retain for an audit or incident.
Sources and further reading
These primary or specialist references informed the concepts in this guide. Product details can change, so verify current documentation before implementation.