Governance

Transparency obligations for AI come from consumer protection, privacy, employment, and AI-specific rules, and from the trust of the people affected. This guide covers the notice types, when each is required, how to word them so people understand, where to place them, how they connect to contestability, and the records that show they were given.

Standard security questionnaires miss what matters about AI vendors: whether they train on your data, who their model providers are, how they evaluate and change models, and what they log. This guide provides the AI-specific question set by domain, explains which answers are disqualifying, and shows how to verify answers rather than file them.

California regulates AI through several channels: privacy regulations on automated decision-making and risk assessments, transparency laws for AI interaction and generated content, employment rules on automated screening, deepfake and synthetic media laws, and consumer protection enforcement. This guide maps the channels and sets out a preparation plan for businesses in or serving California.

Colorado enacted the first comprehensive US state law on high-risk AI systems, placing duties on developers and deployers to prevent algorithmic discrimination. This guide explains coverage, the high-risk definition, developer and deployer obligations, impact assessments, notices, enforcement, and a preparation plan for businesses in or serving Colorado.

Explainability requirements arrive from regulators, auditors, customers, and courts, and they differ by audience. This guide covers who needs what kind of explanation, how to produce faithful explanations for classical models and LLM systems, the limits of explanation techniques, and how to build explanation into the design.

Regulators increasingly require meaningful human oversight of AI decisions, and most implementations fail the meaningful test. This guide covers where oversight obligations come from, the oversight modes and when each applies, designing oversight that people can actually exercise, measuring effectiveness, and the evidence to keep.

Regulated industries can deploy AI, and the ones that succeed share a control set: inventory and tiering, independent validation, explainability, human decision rights, data controls, audit trails, and vendor oversight. This guide covers the shared controls, how sectors differ, where to start, and the mistakes that stall regulated AI programs.

AI incidents include data leakage, harmful outputs, wrong automated decisions, and security compromises, and each may trigger disclosure obligations to regulators, affected people, customers, and partners. This guide covers defining incidents, mapping obligations, timelines, internal escalation, communication, and the records that support defensible disclosure.

AI gives insiders faster ways to gather data, new ways to manipulate systems, and new roles with privileged access to prompts and models. This guide covers the threat patterns, the controls from least-privilege retrieval to anomaly detection and change control, monitoring without surveillance overreach, and the response.

Model risk management was built for statistical models with fixed inputs and outputs. LLM applications and agents strain it: they change with prompts and provider updates, and their outputs are open-ended. This guide covers extending inventory, tiering, independent validation, ongoing monitoring, and governance to AI systems, and what must change.

AI penetration testing assesses whether an LLM application or agent can be made to leak data, take unauthorized actions, or bypass controls, through the model, its tools, its retrieval, and its infrastructure. This guide covers scoping, attack classes, methodology, deliverables, cadence, and how it relates to red teaming and conventional testing.

AI systems generate obligations to keep records that conventional software never did: what the system was, what it was tested on, what it decided, and why. This guide covers the record categories, where each obligation comes from, retention considerations, how to produce records automatically from delivery and operation, and common gaps.

The simplest way to make AI systems PCI compliant is to ensure they never see cardholder data. This guide covers tokenization and masking before AI, scoping decisions when an AI component must process card data, the controls that follow, vendor and model provider requirements, and the common leakage paths.

AI in reconciliations, journal entries, estimates, and disclosures becomes part of the internal control environment. This guide covers documenting AI's role in controls, keeping humans accountable, change management for prompts and models, evidence auditors can test, segregation of duties, and control testing.

Zero trust principles apply to AI systems with one addition: the model itself is an untrusted caller. This guide covers verified identity for every request, least-privilege and short-lived permissions for agents and tools, segmentation of retrieval and tools, gateway policy enforcement, logging, and the architecture that contains compromise.

As AI produces more of what organizations publish, knowing what was generated, by which model, with what review, and from what sources becomes a trust and compliance requirement. This guide covers provenance metadata and content credentials, process controls where standards do not apply, disclosure, and defenses against misattribution.

Every AI control depends on knowing what data a system uses, who may see it, where it came from, and how long it stays. This guide extends data governance to AI: classification that follows data into prompts and indexes, permissions in retrieval, lineage to outputs, quality for training and evaluation data, retention across AI stores, and ownership.

AI systems leak data through paths conventional controls never covered: prompts to unapproved tools, outputs revealing retrieved content, raw logs, agent memory, over-broad retrieval, and vendor retention. This guide maps each path to its control, from classification and redaction to permission trimming, output screening, and vendor terms.

Synthetic voice, video, and text impersonating executives, employees, and brands are now used for payment fraud, access, and reputational attacks. This guide covers the threat patterns, verification procedures that do not rely on voice or video, detection tooling, provenance on genuine content, response playbooks, and training.

An AI ethics committee examines effects that compliance alone does not catch: fairness, impacts on people, automation in sensitive contexts, and alignment with values. This guide covers its scope, membership across disciplines and affected stakeholders, its relationship to the governance board, cadence, and how to keep it from becoming ceremonial.

Policies are read once; training changes behavior. This guide covers what AI acceptable use training must teach, approved tools, data rules, verification, disclosure, and reporting, how to design it by role, how to deliver and refresh it, and how to measure it by behavior rather than completion rates.

AI assistants and agents act on behalf of people and processes, and they must never read or do more than those principals are permitted to. This guide covers identity for every request, delegated permissions, retrieval trimming, tool scoping with gates, secrets handling, and audit, with the mistakes that cause leaks.

Personal information flows through prompts, retrieval, memory, and logs, and California privacy rights follow it. This guide covers notice and purpose limitation, honoring access, deletion, correction, and opt-out requests across AI systems, service provider terms for AI vendors, and preparing for automated decision-making rules.

Education records that reach AI tools remain protected. This guide covers when AI vendors may receive student data under the school official exception, the contract terms that make it lawful, data minimization, prohibitions on redisclosure and training, parent and student rights, and the audit trail institutions need.