---
title: Do California's automated decision rules apply to my AI product?
description: California's CCPA rules on automated decisionmaking technology take effect for significant decisions on January 1, 2027. Who they cover, what they require, how they reach AI vendors, and how to check your product.
category: AI features
service: ai-feature-review
updated: 2026-11-17
---

## Short answer

They apply if a business covered by California's privacy law (the CCPA) uses automated decisionmaking technology to make significant decisions about people: decisions about lending, housing, education, employment, or health care. From January 1, 2027, that business must give people notice before using the technology, let them opt out or appeal to a human, and explain the decision on request. If you sell an AI tool used for those decisions, the rules reach you through your customers, who will need your product to support notice, human review, and explanations.

## Who the rules cover

The rules apply to "businesses" under the CCPA. Generally, that's a for-profit company that does business in California and meets one of these tests:

- Annual gross revenue above $25 million, adjusted for inflation ($26,625,000 as of 2025).
- Buys, sells, or shares the personal information of 100,000 or more California consumers or households a year.
- Gets half or more of its annual revenue from selling or sharing personal information.

Many early-stage companies don't meet any of these tests. But AI vendors often sell to companies that do, and as a service provider you'll be asked to support their obligations by contract.

## What counts as automated decisionmaking

Automated decisionmaking technology (ADMT) is technology that processes personal information and uses computation to replace, or substantially replace, human decisionmaking. A model that scores applicants, ranks candidates, or decides who gets an offer is ADMT if its output decides the outcome without meaningful human involvement.

Human involvement counts only if a person who can understand and interpret the output reviews it and has the authority to change the decision. A reviewer who rubber-stamps the model's output doesn't take the system outside the rules.

The rules focus on **significant decisions**: decisions that result in providing or denying financial or lending services, housing, education enrollment or opportunities, employment or independent contracting opportunities or compensation, or health care services. Earlier drafts also covered behavioral advertising. The final rules don't.

## What covered businesses must do

From January 1, 2027, a business that uses ADMT for significant decisions must:

- **Give a pre-use notice** before using the technology on someone, explaining what it's used for and how it works, in plain language.
- **Offer an opt-out**, or a recognized alternative, such as letting the person appeal the decision to a qualified human reviewer with authority to change it.
- **Answer access requests** about the technology's use, including the logic involved and how it affected the outcome for that person.

Separate requirements started earlier:

- **Risk assessments.** Since January 1, 2026, covered businesses must assess certain higher-risk processing, including using ADMT for significant decisions and using personal information to train ADMT for those decisions. New processing needs an assessment before it starts; processing already under way when the rules took effect has a later deadline. Attestations and summaries are due to the California Privacy Protection Agency by April 1, 2028.
- **Cybersecurity audits.** Businesses whose processing poses a significant security risk must complete annual independent audits, with first certifications due from April 1, 2028 to April 1, 2030 depending on revenue.

## If you sell the AI tool

Most of these duties fall on the business making the decision, not on the vendor whose model it uses. In practice, though, your customers can't comply unless your product supports them, and their contracts will say so. Build for it:

- **Explanations.** A way to show, for each decision, the main factors behind the output in terms a person can understand.
- **Human review.** A workflow where a qualified reviewer can see the output, add context, and change the decision, with that change recorded.
- **Opt-out and appeal routing.** A way to send a person's case to a human process instead of the model.
- **Records.** Logs of inputs, outputs, model version, and reviewer actions, kept long enough to answer access requests and support risk assessments.
- **Documentation.** A plain description of what the model does, what it was trained on, and its known limits, which customers will need for their notices and risk assessments.

Colorado has its own automated decision law with similar themes and different details. See [Does Colorado's automated decision-making law apply to my product?](/guides/colorado-automated-decision-law)

## Checklist an agent can run against the codebase

- [ ] Find every place model output decides, ranks, scores, or filters people for lending, housing, education, employment, contracting, or health care.
- [ ] For each, check whether a person reviews the output before the decision, and whether that reviewer can change it. Note where the model's output is final.
- [ ] Check for a pre-use notice shown before the technology is applied, and what it says.
- [ ] Check for an opt-out or appeal path that routes a person to human review, and that the request is recorded.
- [ ] Check whether the system can produce, for one person, the main factors behind their outcome.
- [ ] Check logging of inputs, outputs, model version, and reviewer actions, and how long those records are kept.
- [ ] Check whether personal information is used to train or fine-tune the decision model, and whether a risk assessment covers it.
- [ ] If the product is sold to other businesses, check that contracts and documentation describe these features.

## Sources

- [California Privacy Protection Agency: approval of the ADMT, risk assessment, and cybersecurity audit regulations](https://www.cppa.ca.gov/announcements/2025/20250923.html)
- [California Privacy Protection Agency: regulations](https://cppa.ca.gov/regulations/)

## If this applies to you

An **AI feature review** works out whether these rules reach your product, directly or through your customers, and produces the notices, human-review design, and build spec to meet them. For one change, such as adding automated scoring to an existing workflow, a **pre-build feature check** covers that change before it ships.
