CCPA CPRA AI Compliance Checklist: Ten ADMT Controls Before January 2027
A working checklist for businesses running AI over California consumer data before ADMT-specific requirements begin January 1, 2027. Each item names the duty, the concrete action, and the evidence it produces, so the list functions as an audit-readiness pass rather than a statement of intent. Ordered the way a CPPA review moves: scope and identity first, then notice and opt-out, then access, appeal and security.
TL;DR
- Inventory every AI use and document which systems substantially replace human judgment in significant decisions.
- Bind identity, notice version, personal-information classification, policy, prompt, and output to every ADMT request.
- Enforce consumer opt-outs before a prompt reaches the model, and link each appeal through human review to the original decision.
- Keep signed, tamper-evident records outside the application and make breach scope queryable by consumer, prompt, data type, and endpoint.
- Complete these controls before ADMT duties attach on January 1, 2027; statutory breach damages are $100 to $750 per consumer per incident.
California's ADMT clock is running. The California Privacy Protection Agency's regulations were approved on 23 September 2025 and took effect on 1 January 2026, with automated decisionmaking technology obligations attaching from January 1, 2027, initial risk assessments due by 31 December 2027, and the first CPPA submission on 1 April 2028.
This checklist follows the order a CPPA review tends to take: scope and identity records first, then pre-use notice and opt-out, then access, appeal, security, and breach readiness. Each item pairs the action with the evidence it produces, because the regulations test what happened to a specific consumer's decision rather than what a programme document intended.
1. Decide which AI uses are ADMT for significant decisions
Inventory every AI call and mark the ones where computation replaces or substantially replaces human decisionmaking about financial or lending services, housing, education enrollment or opportunity, employment or independent contracting, or healthcare. Evidence is a scoping record naming each system, the decision it drives, and the reasoning, since advertising and most internal productivity uses fall outside.
2. Bind identity to every AI request
Route AI traffic so a named user, agent, or consumer context, and not a shared service key, is attached to each prompt. Evidence is a per-request record showing which identity sent which prompt to which model, which is what every downstream ADMT right depends on.
3. Bind the pre-use notice version to the request
Record which notice and which policy version were in force at the moment of each ADMT decision. Evidence is a per-request policy-state stamp, which is what proves a published notice actually governed the decisions a consumer is asking about.
4. Enforce opt-out at the boundary
Refuse or reroute requests from consumers who have opted out of ADMT for significant decisions, before the prompt reaches the model. Evidence is the refusal record. An opt-out honored only inside application code leaves nothing an auditor can distinguish from an opt-out that was ignored.
5. Classify the personal information in the prompt
Evaluate the classification of the context window before it leaves, separating ordinary personal information from the sensitive personal information categories the CPRA created. Evidence is a per-request classification decision recorded at the boundary.
6. Produce a per-decision record for every ADMT output
Write a record binding the input prompt to the output and the applied policy for each significant decision. Evidence is decision lineage, the operational backing for the consumer's right to information about the logic, and it mirrors GDPR Article 22 handling.
7. Link the appeal chain to the original decision
Flag decisions that must support appeal and link the original decision record, the consumer's appeal, and the human reviewer's outcome into one chain. Evidence is that chain, retrievable by consumer and timestamp, which is the piece that breaks when decisions live in application logs and appeals live in a ticketing system.
8. Replace static API keys with per-role, per-route policy
Remove shared credentials granting permanent full model access to any caller, which is where CCPA section 1798.100(e) reasonable security meets the cybersecurity audit phasing in from 1 April 2028 by revenue tier. Evidence is an access policy plus the enforcement log showing which roles reached which models.
9. Keep the record independent of the application
Ensure the compliance log is written by a system separate from the one making the AI call, so it survives selective logging, suppression, and crash-loss. Evidence is a signed, tamper-evident log outside application custody, as argued in You Own the AI Liability, Not the Vendor.
10. Pre-build the breach-scoping path
Confirm you can reconstruct which prompts carried which personal information to which endpoint, because CCPA section 1798.150 gives consumers a private right of action with statutory damages of $100 to $750 per consumer per incident. Evidence is a query against the independent request log that produces the affected class. My honest take: that per-consumer math will move California AI logging budgets faster than any CPPA enforcement action.
How the checklist maps to the controls
Items 1 through 10 resolve to a handful of enforcement points at the request boundary, laid out in the CCPA controls mapping. The audit-evidence view covers what each artifact looks like when the CPPA asks. Together they turn this list into a running system with evidence attached.
DeepInspect
Nine of these ten items produce evidence at one place: the boundary between your users or agents and the LLM APIs they call. DeepInspect sits inline at that boundary. It binds identity to each request, reads opt-out state, classifies the prompt, stamps the policy version in force, flags decisions for appeal, and writes a signed per-decision record before the response returns. The checklist becomes a query rather than a scramble.
For a business counting down to January 1, 2027, items 4 and 7 are the two that a consumer complaint will reach first. Book a technical deep dive at deepinspect.ai.
Frequently asked questions
- Is this checklist enough for CCPA compliance?
It covers the AI-traffic obligations a CPPA review reaches for: ADMT scoping, notice, opt-out, access to the logic, appeal, security, and breach reconstruction. It does not replace a full CCPA programme covering collection notices, sale and sharing opt-outs, sensitive personal information limits, retention disclosures, and consumer rights across every channel. Treat it as the AI-specific layer.
- What are the exact ADMT deadlines?
The California Privacy Protection Agency says ADMT-specific requirements begin January 1, 2027. Other requirements in the regulations have separate phased timelines, including risk-assessment and cybersecurity-audit obligations. Confirm which dates apply to your organization with counsel, then retain evidence of scope, notices, opt-out handling, and risk-assessment work so the program is auditable.
- Does the CCPA apply if we are not based in California?
The CCPA applies to businesses doing business in California that meet the revenue, data volume, or data sale thresholds, wherever they are headquartered. A company in Denver making significant decisions about California residents through ADMT falls inside the rules, so scope by where the consumers are rather than where the infrastructure sits.
- How does this compare with other state AI rules?
Colorado's AI Act and the Texas Responsible AI Governance Act take a risk-and-duty approach rather than California's consumer-rights approach, but the underlying artifacts overlap heavily: identity binding, prompt classification, per-decision records, and independent logs. Building for the California ADMT rules produces most of what the other regimes ask for, which is the practical argument for treating California as the floor.