GaplessFrameworks
The security, privacy, and AI frameworks enterprise customers ask about, what each requires in practice, and the specific engineering controls in this system that satisfy them.
This page summarizes the frameworks most often raised by enterprise customers and internal legal teams, and maps each to concrete controls. Use it to plan work and to speak the same language as compliance teams. It is a summary, not a substitute for the frameworks themselves or for legal advice.
SOC 2
An attestation report by an independent auditor on controls related to security, and optionally availability, confidentiality, processing integrity, and privacy. A Type II report covers how controls operated over a period, usually 6 to 12 months, and is what enterprise customers expect.
| Control area | In this system |
|---|---|
| Access control | SSO, role-based permissions, access reviews, least privilege for engineers and services |
| Change management | Reviewed pull requests, CI tests and eval gates, controlled deploys |
| Risk assessment | Threat model, AI risk tiers |
| Monitoring | Alerts, audit logs, incident response |
| Vendor management | Reviews of model providers and other subprocessors |
| Data protection | Encryption, retention, deletion |
| Availability | Backups, disaster recovery tests, SLOs |
AI-specific points auditors increasingly check: how prompt and model changes are controlled (treat them as code changes with review and evaluation), and how model providers are assessed as vendors.
ISO/IEC 27001
An international standard for an information security management system (ISMS), certified by an accredited body. It overlaps heavily with SOC 2, with more emphasis on the management system: documented risk assessment, a statement of applicability for its control set, internal audits, and management review.
If you have SOC 2 controls in place, ISO 27001 is mostly additional documentation and process.
ISO/IEC 42001
A standard for an AI management system, published in 2023. It applies the management system approach to AI: policies, risk and impact assessment, lifecycle controls, data quality, transparency, and oversight, with a set of AI-specific controls.
| Requirement area | In this system |
|---|---|
| AI policy and roles | Intended and prohibited uses, accountability table |
| AI risk assessment | Risk tiers for use cases |
| AI impact assessment | Impact assessments for high-tier use cases |
| Lifecycle controls | Evaluation datasets, CI gates, canary rollouts, monitoring |
| Data for AI | Data inventory, sources, quality checks, lineage |
| Information for interested parties | Transparency notices, system card |
| Third-party relationships | Model provider records and reviews |
The Responsible AI and Governance section describes most of what 42001 expects. Certification is increasingly requested by large customers in regulated industries.
GDPR and similar privacy laws
The EU General Data Protection Regulation applies to personal data of people in the EU, and many other jurisdictions have similar laws.
| Principle or right | What it means here |
|---|---|
| Lawful basis | The company deploying the assistant needs a basis for processing employee or customer data, often legitimate interest with an assessment. Your customers decide this as controllers. You support them as a processor |
| Purpose limitation | Data used for answering questions is not reused for unrelated purposes, such as training models, without a basis |
| Data minimization | Send only needed content to models. Keep logs free of content by default |
| Storage limitation | Retention schedule with deletion |
| Right of access | Export a user's conversations and data on request |
| Right to erasure | Deletion that reaches every copy, including indexes and traces |
| International transfers | Model calls to providers in other countries are transfers. Need appropriate safeguards, or processing in-region |
| Processors | Model providers and other vendors are processors or subprocessors, with DPAs |
| Data protection impact assessment (DPIA) | Usually required for large-scale processing of employee data or new technologies. Your customers will ask for input |
| Automated decision-making | Decisions with legal or similarly significant effects based solely on automated processing are restricted. Human oversight design matters |
EU AI Act
A risk-based regulation for AI systems placed on the market or used in the EU, with obligations phasing in over several years. Check the current timeline and guidance with counsel, as details continue to be clarified.
The broad structure:
| Category | Examples | Obligations, broadly |
|---|---|---|
| Prohibited practices | Social scoring, certain manipulative practices, emotion recognition in workplaces with narrow exceptions | Not allowed |
| High-risk systems | AI used in employment decisions (recruiting, promotion, termination, task allocation, monitoring), credit scoring, access to essential services, and others listed in the Act | Risk management, data governance, technical documentation, logging, transparency to deployers, human oversight, accuracy and robustness, conformity assessment, registration |
| Transparency obligations | Systems interacting with people, generated content | Inform people they are interacting with AI. Label certain generated content |
| General-purpose AI models | Foundation models from providers | Obligations mainly on model providers |
| Minimal risk | Most other uses | No specific obligations, voluntary codes |
For the reference assistant:
- General Q&A and drafting is generally not high-risk, but it has transparency obligations: users must know they are interacting with AI.
- Any use case in the employment area, such as screening candidates or evaluating performance, may be high-risk. This is why the prohibited and restricted uses list keeps those out unless separately assessed.
- If you deploy a third-party model, the provider carries the general-purpose model obligations. You carry obligations for how you use it.
Note that roles matter: a company building and selling the assistant is a provider of the AI system. A customer using it is a deployer. Each has different duties.
HIPAA
The US Health Insurance Portability and Accountability Act applies to protected health information (PHI) handled by covered entities and their business associates.
If customers in healthcare want to use the assistant with PHI:
- You need a Business Associate Agreement (BAA) with the customer, and BAAs with every subprocessor that touches PHI, including model providers. Not every provider or model offers a BAA. Check per model and service.
- Access controls, audit logs, encryption, and breach notification processes must meet HIPAA's security rule.
- Restrict PHI-handling tenants to covered providers and regions through model gateway configuration.
Mapping table
| Control | SOC 2 | ISO 27001 | ISO 42001 | GDPR | EU AI Act |
|---|---|---|---|---|---|
| SSO and RBAC | Yes | Yes | Security of processing | ||
| Permission-filtered retrieval | Yes | Yes | Confidentiality | ||
| Encryption and key management | Yes | Yes | Security of processing | ||
| Retention and deletion | Yes | Yes | Storage limitation, erasure | ||
| Audit logs | Yes | Yes | Yes | Accountability | Logging for high-risk |
| Eval datasets and CI gates | Change management | Yes | Lifecycle controls | Accuracy | Accuracy and robustness |
| Risk tiers and impact assessments | Risk assessment | Risk assessment | Yes | DPIA | Risk management |
| Human confirmation for actions | Human oversight | Automated decisions | Human oversight | ||
| Transparency notices and system card | Yes | Transparency | Transparency, documentation | ||
| Vendor reviews of model providers | Yes | Yes | Third parties | Processor contracts |
Checklist
- Legal has confirmed which frameworks and laws apply.
- SOC 2 Type II or ISO 27001 is planned or in place before selling to large enterprises.
- Prompt and model changes are part of change management evidence.
- Model providers are covered in vendor management, with DPAs and, where needed, BAAs.
- Privacy rights (access, erasure) are supported across all stores.
- Use cases are classified against AI regulation categories.
- Transparency obligations are met in the UI.
Compliance and Legal
The compliance obligations an enterprise AI assistant typically faces, how engineering work maps to them, and how to prepare for customer security reviews and audits.
Data Residency
How to keep customer data and AI processing in required regions, including model calls, embeddings, logs, and support access, and how to design for multiple regions.