Which CCPA duties matter most for AI traffic?

The ADMT package does the heaviest lifting: pre-use notice, opt-out, the right to information about the logic, and the right to appeal to a human, with ADMT-specific requirements beginning January 1, 2027. Section 1798.100(e) reasonable security and the section 1798.150 private right of action apply separately, while risk-assessment and cybersecurity-audit requirements follow their own phased timelines.

How does the ADMT definition affect scope?

The definition is functional, covering technology that processes personal information and uses computation to replace or substantially replace human decisionmaking for a significant decision. That excludes advertising and most internal productivity uses, and it captures credit adjudication, tenant screening, admissions, hiring, and clinical triage. Scope decisions therefore turn on what the decision does, not on which model answers.

Do these controls require modifying the model?

They operate at the HTTP request boundary between the application and the model, so they are model-agnostic and need no retraining. The application continues to supply identity and consumer opt-out context, which is its role in any identity-aware design, while classification, enforcement, and recording happen in the request path.

What about businesses outside California?

The CCPA applies to businesses that do business in California and meet the revenue, data volume, or data sales thresholds, regardless of where they are headquartered. A company in Austin making significant decisions about California residents through ADMT is inside the rules, so the practical scoping question is where the consumers are.

CCPA CPRA AI Controls Mapping: California ADMT Obligations to Enforcement Points

TL;DR

Enforce California ADMT duties at the outbound HTTPS request where identity, opt-out state, prompt classification and policy can be evaluated before a model responds.
Bind every AI call to the notice and policy version in force, and retain per-decision lineage linking input, output, consumer and timestamp.
Refuse opted-out requests or route them to a human before model access, while preserving an independent refusal record.
Link each appeal to the original decision and reviewer outcome, and use signed per-request records to prove access controls, audit compliance and breach scope.

The California Privacy Protection Agency wrote AI rules that describe decisions rather than models. Its regulations on automated decisionmaking technology (ADMT), risk assessments, and cybersecurity audits were approved on 23 September 2025, took effect on 1 January 2026, and impose ADMT obligations from January 1, 2027. The definition turns on function: technology that processes personal information and uses computation to replace or substantially replace human decisionmaking, applied to a significant decision about finances, housing, education, employment, or healthcare.

That framing puts the obligation at the request, which is convenient, because the request is also where a control can sit. California programmes usually stop at a governance policy aligned to the CCPA. What tends to be missing is the mapping from a duty to the control that enforces it where personal information leaves the network. I want to set that out duty by duty.

The control point: one HTTPS request

Every mapping below resolves to an HTTPS POST carrying a context window to a model endpoint. That request is where the consumer's identity is known, where their opt-out state can be read, where the prompt classification is evaluated, and where a decision gets recorded or refused. A control described in a privacy policy documents the duty without enforcing it.

Pre-use notice (ADMT rules, from January 1, 2027)

A business using ADMT for a significant decision must tell consumers before the fact what it is doing and how the technology works. The control is policy versioning bound to the request path, so every AI call records which notice and which policy were in force at that timestamp. The evidence is a per-request record that ties a decision to the disclosure that governed it, which is what turns a published notice into proof of the notice's scope.

Opt-out rights

Consumers may opt out of ADMT for significant decisions. The control is enforcement at the boundary: a request from a consumer who has opted out gets refused or routed to a human path before it reaches the model. The evidence is the refusal record itself. An opt-out honored inside application code with no independent trace is unprovable at audit, which is the practical difference between a preference flag and a control.

Access and the right to the logic

Consumers may request information about the use of ADMT, including its logic and how it applied to them. The control is a per-decision record binding input to output and the applied policy, the same lineage requirement that GDPR Article 22 handling creates. The evidence is decision lineage retrievable by consumer and by timestamp rather than a general description of the system.

Appeal to a human reviewer

Consumers may appeal an ADMT result to a person. The control is a route flag on decisions that must support appeal, plus a linkage between the original decision record, the appeal, and the reviewer's outcome. The evidence is a single chain covering all three, which is exactly what breaks when the AI decision lives in an application log and the appeal lives in a ticketing system.

Risk assessments and reasonable security

Initial risk assessments for ongoing activities are due by 31 December 2027, with submission to the CPPA on 1 April 2028, and cybersecurity audits phase in from 1 April 2028 for businesses over $100 million in 2026 revenue, 1 April 2029 for $50 million to $100 million in 2027 revenue, and 1 April 2030 below that. CCPA section 1798.100(e) requires reasonable security. The control is access governance on the model channel: per-role, per-route policy replacing shared static keys, failing closed on ambiguity. The evidence is a continuous access record rather than a point-in-time attestation.

The mapping at a glance

CCPA / CPRA duty            Where it applies         Control                           Evidence produced
--------------------------  -----------------------  --------------------------------  ------------------------------
Pre-use notice (ADMT)       Every ADMT request       Policy version bound to request   Notice-state decision log
Opt-out of ADMT             Request admission        Boundary refusal or human route   Refusal record per consumer
Access / right to logic     Decision output          Per-decision record               Decision lineage
Appeal to human reviewer    Post-decision            Review flag + linked chain        Decision, appeal, outcome
Risk assessment             Programme level          Classification + per-role policy  Per-request permit record
Sec. 1798.100(e) security   Model channel access     Per-role, per-route enforcement   Continuous access record
Sec. 1798.150 breach        Post-incident            Independent request log           Reconstructable timeline

The breach row

The final row activates only after an incident. CCPA section 1798.150 gives consumers a private right of action for a breach of nonencrypted, nonredacted personal information caused by a failure to maintain reasonable security, with statutory damages of $100 to $750 per consumer per incident. When the AI channel is the exposure path, scoping that class depends on knowing which prompts carried which personal information to which endpoint. The independent per-request record answers it, and it is the same artifact the CCPA audit-evidence view is built around. My frank opinion: the statutory-damages math is what will move California AI logging budgets, well before any CPPA enforcement action does.

DeepInspect

Each row above resolves to an enforcement point, and DeepInspect is that point. It sits inline between your users or agents and the LLM APIs they call, evaluates identity, opt-out state, classification, and policy on every request, enforces per-role and per-route rules, and writes a signed record of the decision before the response returns. The ADMT obligations stop being a set of promises and become a running control with evidence attached.

For a business counting down to January 1, 2027, that means notice, opt-out, access, and appeal each have a per-decision record standing behind them. Book a technical deep dive at deepinspect.ai.