← Back to blog

California’s 2026 ADMT Rules Require a Decision Boundary, Not an AI Label

California’s CCPA rules define rights for certain automated decisions. Product teams should map system decisions, human review, and the consumer requests they must support.

Calling a feature “AI” does not determine whether California’s automated decision rules apply. The important questions are what personal information the system processes, whether computation replaces or substantially replaces human decision-making, and whether the result drives a covered significant decision.

The California Privacy Protection Agency adopted regulations covering automated decisionmaking technology (ADMT), risk assessments, and cybersecurity audits in July 2025. The Office of Administrative Law approved them on 22 September 2025, and the regulations became effective 1 January 2026. For product teams, ADMT is now a design and data-flow question—not merely a future policy review.

Define the decision the software makes

The regulations define ADMT as technology that processes personal information and uses computation to replace or substantially replace human decision-making. They include profiling when it performs that replacement. The definition also excludes ordinary technical functions such as hosting, caching, storage, databases, and spreadsheets when they do not replace a human decision.

That distinction matters for analytics. A dashboard that counts pageviews or gives an analyst a report is not automatically ADMT just because it uses statistics or machine learning. A system that uses behavioral data to accept or reject an applicant, rank a worker, determine access to housing or education, or make another covered significant decision needs a different analysis.

Map the full decision path: input data, model or rule, output, human reviewer, and final action. “Human in the loop” is not a sufficient description if a reviewer cannot understand the output or routinely approves it without meaningful independent judgment. Record what the person actually decides and whether the system’s output controls the outcome in practice.

Build access and opt-out into the product

For covered uses, the regulations establish consumer rights to access information about a business’s use of ADMT and to opt out of certain uses. The opt-out right is tied to specified ADMT uses, including technology used to make significant decisions; it is not a blanket right to disable every algorithm, report, or analytics query.

Treat these requests as product operations. Keep an inventory linking each ADMT use to its purpose, input categories, decision type, and affected workflow. Provide a clear route to submit requests, verify only what is needed to act on them, propagate the result to systems that use the same data, and retain evidence of completion. Avoid forcing a person to create an account or provide extra information when the business can honor an opt-out without it.

Access responses need explainable system behavior, not a marketing description. Be able to identify the purpose of the use, the output or result about the consumer, and the information needed to explain how the technology processed personal information and how the output affected the significant decision. Do not log every input indefinitely just to make a future request answerable; document the decision path and retain only justified evidence.

Keep measurement out of eligibility logic

An analytics event stream should not silently become a decision system. Keep product metrics separate from eligibility, employment, credit, housing, education, or other consequential workflows. If a team wants to feed analytics into a decision model, treat that as a new processing purpose: reassess the legal basis, notices, risk assessment, data quality, consumer rights, and the model’s human-review design before launch.

Do not use a dashboard’s “conversion” or “engagement” score as a proxy for a decision rule without examining its effect. A metric can become consequential when it determines who receives an offer, service, or opportunity. Conversely, a model label alone does not make ordinary descriptive analytics ADMT under the regulation’s definition.

The implementation task is concrete: maintain a decision inventory, mark the point where a human judgment is replaced, connect that use to access and opt-out handling, and test that the request reaches every downstream system. That boundary is more useful than trying to classify an entire product as either “AI” or “not AI.”

Sources

Comments

Loading comments…