What Does the Global Regulatory Landscape Look Like?
California has become the de facto laboratory for US state AI and privacy regulation. The California Consumer Privacy Act (CCPA), as amended by the California Privacy Rights Act (CPRA), has been in full effect since 1 January 2023, and the state's enforcement apparatus is now fully operational: the California Privacy Protection Agency (CPPA) began enforcement on 1 July 2023, alongside continued enforcement by the Attorney General, with civil penalties of up to USD 2,500 for each unintentional violation and USD 7,500 for each intentional one.
2025 marks the next escalation. The CPPA advanced its first batch of revised CCPA regulations, including rules touching automated decision-making technology (ADMT), and its long-awaited ADMT regulations—drafted in November 2023—remain the most consequential pending rulemaking for any business that uses AI to profile, score, or make decisions about Californians. At the same time, the California Delete Act, signed in September 2023, begins imposing deletion obligations on registered data brokers from January 2026, tightening the surrounding ecosystem that AI systems draw from.
For enterprises, the practical message is that California is not one more privacy regime to track—it is the template that other states, and increasingly federal conversations, look to. Businesses that build AI systems touching Californian consumers are effectively building to the most demanding US standard in the market today.
What Do the CCPA and CPRA Actually Require of AI Systems?
The core obligations cluster around consumer rights over personal information, including information processed by AI. Consumers have rights to access, correct, and delete personal information, and to opt out of the sale or sharing of their data for cross-context behavioural advertising. The CPRA added a right to opt out of the use of sensitive personal information, and—critically for AI—a right to opt out of automated decision-making technology that profiles consumers or significantly affects them. None of these rights disappears because the processing happens inside a model rather than a database.
For AI systems, three obligations are doing the heavy lifting. First, notice: before processing personal information for AI purposes, businesses must disclose the categories of information, the purposes, and how consumers can exercise rights. Second, deletion and correction: if a model was trained on personal information, the business must be able to respond to deletion requests—which is why training-data provenance and re-training or exclusion mechanisms matter operationally, not just on paper. Third, the ADMT rules: when they take final form, businesses will need pre-use notices, impact assessments, and opt-out mechanisms for ADMT that profiles or produces significant effects.
There is also a contractual dimension. CCPA duties flow through service-provider and contractor agreements, and AI vendors—from model providers to analytics platforms—must be brought into the compliance chain. A business cannot delegate its CCPA obligations to an AI vendor that refuses to support deletion, correction, or opt-out mechanics; the liability remains with the business that controls the processing.
What Are the Compliance Requirements for Enterprise AI?
- Consumer Rights Infrastructure: Implement access, correction, deletion, and opt-out workflows that cover personal information processed by AI models, not just databases.
- ADMT Disclosures and Opt-Outs: Prepare pre-use notices, opt-out mechanisms, and impact assessments for automated decision-making and profiling.
- Sensitive Information Handling: Honour the right to limit the use of sensitive personal information, including inferences used in AI scoring.
- Vendor and Service-Provider Contracts: Flow CCPA obligations through contracts with model providers, analytics vendors, and data processors.
- Training-Data Governance: Maintain provenance, deletion, and re-training mechanisms so consumer rights can be honoured across the model lifecycle.
Treat this list as a design requirement, not a paperwork exercise. Each item fails gracefully in different ways: rights infrastructure fails when deletion requests hit a model with no retraining path; ADMT rules fail when a vendor refuses to support opt-outs; vendor contracts fail when a processor's privacy practices are opaque. Enterprises that design AI systems with these obligations in mind from the start avoid the retrofitting problem that dominates most 2025 remediation projects.
How Do You Build a Sustainable Compliance Program?
Sustainable compliance requires organisational commitment, technical infrastructure, and regulatory intelligence. Commit a named owner and a cross-functional working group—legal, privacy, data, and product—to CCPA and ADMT readiness, and give them standing authority over AI feature releases that involve personal information. Invest in infrastructure that makes rights fulfilment operational: data inventories that include model inputs and outputs, deletion pipelines that reach training data, and notice systems that keep disclosures current as features change.
Regulatory intelligence is the third pillar because the California regime is still moving. The revised CCPA regulations and the ADMT rulemaking will change concrete requirements through 2025 and 2026, and enforcement priorities will sharpen as the CPPA matures. Organisations that review their AI data flows against California requirements quarterly—rather than annually or reactively—will absorb these changes as updates, while organisations that wait will discover that the ADMT rules, when finalised, assume infrastructure they do not have.
How Do You Construct an Enterprise AI Compliance System?
Beehive Strategy recommends building the compliance system across three dimensions: organisational structure, institutional processes, and technical tools. Designate a privacy and AI compliance owner with direct reporting to the Chief Risk Officer or General Counsel, and establish a working group spanning legal, data, engineering, and business teams. The working group's standing mandate should include reviewing each AI use case against CCPA and CPRA obligations before it reaches production—the same gate that catches privacy problems catches most ADMT problems too.
Processes should cover the full lifecycle of each AI system. At intake, classify the use case: does it involve profiling, sensitive information, or decisions with significant effects on consumers? During development, document data flows, model inputs and outputs, and the mechanisms for consumer rights fulfilment. At deployment, publish notices and operationalise opt-out and deletion paths. In operations, monitor for drift between the declared purpose and actual use—California's purpose-limitation and data-minimisation principles make that drift a legal issue, not just a quality issue.
For enterprises using AI vendors, the technical-tools dimension includes contractual hygiene: verify that each vendor can support deletion, correction, and opt-out obligations, and require evidence of it before contracting. Beehive Strategy integrates this vendor diligence into conversational BI and analytics deployments as standard practice—our engagements document where personal information flows, how consumer rights will be honoured, and what each vendor is contractually required to support, so the compliance story holds together when a regulator asks.
How Do the 2025 ADMT Regulations Reshape Your AI Roadmap?
The CPPA's proposed ADMT regulations, first drafted in November 2023 and refined through 2024 and 2025, are the single most important US rulemaking for AI-enabled businesses to watch. As drafted, they would require pre-use notices for ADMT, a consumer right to access information about ADMT use, opt-out mechanisms for certain uses, and impact assessments for high-risk processing. The scope is broader than classic profiling: it reaches systems that meaningfully affect consumers, including many analytics, personalisation, and decision-support applications.
The prudent posture is to build now for the rules as proposed. Map every AI use case that involves Californian personal information, identify which would qualify as ADMT, and design the notice, opt-out, and impact-assessment artefacts while the rulemaking is still in motion. Because the CPPA's rulemaking has a long gestation, organisations that prepare early gain a genuine competitive advantage: when the final regulations arrive, they will be updating a working system rather than constructing one under deadline. That is the difference between treating California as a compliance burden and treating it as the benchmark that makes AI defensible everywhere in the United States.
How Do You Operationalise Consumer Rights for AI?
CCPA and CPRA give consumers the right to know, delete, correct, and opt out of the sale or sharing of their personal information, and AI systems make operationalising those rights harder because personal data hides inside training sets, embeddings, and model outputs. The practical answer is to treat the right as a data-flow problem: maintain a map of where a consumer's data lives, including the derived artifacts, so that a deletion or opt-out request can be honoured at the source and propagated to the models that learned from it on the next retraining cycle.
For businesses, the discipline that pays off is linkage between the consent record and the AI inventory. When a new model is proposed, the team should be able to answer whether it uses personal information, under which lawful basis, and how a consumer could exercise a right against it. We embed these checks into the model-governance gate so a model cannot reach production without a recorded answer. That is what turns a compliance obligation that scares legal teams into a routine engineering step, and it is the difference between a defensible programme and a spreadsheet of good intentions.
What Should Small Businesses Do About CCPA and CPRA?
Smaller businesses often assume the exemptions shelter them, but the thresholds are higher than many realise and the reputational cost of a lapse is disproportionate. The pragmatic starting point is an inventory: what consumer data do you collect, where does it flow, and do you sell or share it — including sharing with ad and analytics partners, which the CPRA treats broadly. Most small businesses discover they are closer to "sharing" than they thought, and that a simple opt-out mechanism and a published notice close most of the gap.
For AI specifically, the low-cost move is to avoid training on customer data you have not explicitly justified, and to keep a record of which systems do. You do not need a fortress; you need a defensible position and the ability to answer a consumer request within the statutory window. Businesses that treated CCPA and CPRA as a prompt to tidy their data house in 2025 entered 2026 with fewer surprises and a clearer story for customers who now ask, rightly, what happens to their information when an AI is involved.
How Do You Prepare for an Enforcement Inquiry?
An inquiry is won or lost before it arrives. The organisation that can answer a regulator's first letter in days — here is our AI inventory, here is the lawful basis for each system, here is the consumer-rights mechanism, here is the audit trail — is in a different position from one that needs months to assemble the same. Preparation means the governance records are real, current, and connected to the systems, so the response is a query, not a project.
We run clients through a mock inquiry: a regulator-style request for a specific model, answered from the live inventory and lineage graph. The exercise exposes the gaps — a model with no recorded basis, a data flow with no owner, a deletion that cannot be proven — while there is still time to close them. Treating the mock inquiry as a routine drill, on a fixed cadence, is what turns a frightening hypothetical into a manageable operational task, and it is the single best preparation for the enforcement environment tightening through 2025 and 2026.
How Do You Balance Innovation with Compliance?
The false choice is innovation versus compliance. The real practice is designing compliance into the build so it never becomes a gate that blocks release. When the model-governance gate runs in CI — checking lawful basis, data sensitivity, and consumer-rights coverage as part of the pipeline — a compliant model ships as fast as a non-compliant one would have, because the checks are seconds, not weeks of review.
The balance also comes from scoping. Teams move fastest on use cases that need little personal data, because the obligation is light; they slow down only where the data is sensitive, and there the slowdown is warranted. We help clients map their AI roadmap onto this risk gradient so innovation flows where it is cheap and is appropriately careful where it is not. Organisations that did this in 2025 shipped more AI, not less, because compliance stopped being the thing that said no and became the thing that said yes, safely.
How Do You Handle the Opt-Out at Scale?
The opt-out is where CCPA and CPRA meet engineering reality: a consumer clicks "do not sell or share", and every system that used their data for AI must honour it. At scale this only works if the consent signal is a single, queryable record that flows to the AI inventory, so a model built on that person's data is excluded from the next retraining and any real-time sharing stops immediately. The organisations that handle this well treat consent as a first-class data attribute, not a checkbox in a privacy portal nobody connects to the models.
The harder case is derived data — embeddings and model weights that learned from the person. Full honoured deletion there means excluding the record from the next training cycle and being able to show it. We wire the opt-out to the same governance gate as the DPIA and the inventory, so the signal is operational, not aspirational. Businesses that built this in 2025 answered consumer requests within the statutory window without a fire drill; those that had not discovered the gap was exactly what a regulator tested first. At scale, the opt-out is an architecture, not a process.
