---
title: What do California's ADMT rules mean for AI agents that make significant decisions?
url: https://opencontroller.lyzr.com/blog/california-admt-rules-and-ai-agents
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# What do California's ADMT rules mean for AI agents that make significant decisions?

**California's CCPA regulations on automated decision-making technology (ADMT) apply from 1 January 2027 to businesses that use ADMT to make significant decisions about consumers. An AI agent counts as ADMT when it processes personal information and replaces, or substantially replaces, human decision-making. Significant decisions cover finance and lending, housing, education, employment and healthcare. The duties are a pre-use notice, an opt-out (or a human appeal) and an access right. Risk assessments already apply.**

The California Privacy Protection Agency (CPPA) finalised the rules in 2025, and the Office of Administrative Law approved them on 23 September 2025. They took effect on 1 January 2026, with staggered compliance dates.

## When does an AI agent count as ADMT?

ADMT is technology that processes personal information and uses computation to replace human decision-making, or to substantially replace it. "Substantially replace" means the output drives the decision without meaningful human involvement. Meaningful review requires a person who:

1. understands the agent's output;
2. weighs it alongside other relevant information; and
3. has authority to change the decision.

An agent that drafts a recommendation that a trained reviewer reads, weighs and can overturn is usually not ADMT. An agent that approves or rejects on its own, or whose output a reviewer rubber-stamps, likely is. Basic tools such as spellcheckers, calculators and web hosting are excluded when they do not replace decisions.

## Which decisions are "significant"?

| Domain | Agent example that could trigger ADMT duties |
|---|---|
| Financial or lending services | An agent that decides loan pre-approval or credit limits |
| Housing | An agent that screens rental applications |
| Education enrolment or opportunities | An agent that ranks applicants for admission or scholarships |
| Employment, contracting or pay | An agent that screens CVs, allocates shifts or sets pay |
| Healthcare services | An agent that decides eligibility for a health service |

An internal coding agent, a meeting-notes agent or a sales-email drafter does not make significant decisions about consumers. Those agents fall outside the ADMT duties, although other CCPA duties may still apply.

## What must you do from 1 January 2027?

1. **Pre-use notice.** Before using ADMT, give a prominent notice covering the purpose, how the ADMT works, the right to opt out and the right to access, what happens if someone opts out, and a statement that retaliation is not allowed.
2. **Opt-out.** Offer at least two ways to opt out, one of them matching how you mainly interact with consumers. It cannot be buried in a cookie banner.
3. **Exceptions to opt-out.** You can offer a human appeal instead: a reviewer who can overturn the decision. Narrow exceptions also exist for some hiring, admission and work-allocation uses, where the tool only assesses ability to perform, works for its purpose and does not unlawfully discriminate.
4. **Access requests.** Explain in plain language the purpose, the logic used, the outcome for the person and how the output was used. Trade secrets and security-sensitive details can be withheld.

## What about risk assessments?

Risk assessments have applied since 1 January 2026. Using ADMT for a significant decision is one trigger. Others include processing sensitive personal information and using personal information to train ADMT. An assessment must:

- be completed before the processing starts;
- record purpose, data categories, benefits, negative impacts, safeguards and whether you will proceed;
- name the reviewers and approvers;
- be reviewed at least every three years, and within 45 days of a material change.

Attestations and summaries for assessments carried out in 2026 and 2027 are due to the CPPA by **1 April 2028**. For agents, a "material change" can be a new tool, new data source or more autonomy. Version tracking matters.

## ADMT readiness checklist for AI agent programmes

1. Inventory every agent that touches consumer personal information. See [how to find shadow AI agents](/blog/how-to-find-shadow-ai-agents).
2. Tag agents that make or substantially make significant decisions.
3. For each tagged agent, decide: human appeal, opt-out, or redesign for meaningful human review.
4. Log reviewer actions to prove the review is meaningful. See [evidence for human review](/blog/evidence-for-human-review-of-ai-agent-outputs).
5. Draft the pre-use notice and access-request response per agent.
6. Run and store a risk assessment before deployment; set a 45-day trigger for material changes.
7. Keep assessments for as long as processing continues, or five years after completion, whichever is longer.

## How Opencontroller does this

Opencontroller auto-discovers agents across AWS, Azure, Google Cloud and connected agent platforms, and keeps a governed registry of agents and versions. That shows which agents touch consumer data, and when a version changes enough to trigger a reassessment. Delegated on-behalf-of identity records which person an agent acted for. LLM judges score production traffic, and live groundedness and task-adherence guardrails check outputs. Policy as code is extensible, so you can add CCPA ADMT rules to built-in mappings for OWASP, MITRE ATLAS, NIST and ISO 42001. This is **evidence for** the CCPA rules, not legal advice. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

Read next: [the Colorado AI Act and AI agents](/blog/colorado-ai-act-and-ai-agents) and [how to map agent controls to frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks).

## Frequently asked questions

**When do California's ADMT requirements start?**
Businesses using ADMT for significant decisions must comply from 1 January 2027. Risk assessment duties have applied since 1 January 2026, with submissions due by 1 April 2028.

**Does adding a human reviewer take an AI agent out of ADMT?**
Only if the review is meaningful. The reviewer must understand the output, weigh it with other information and have authority to change the decision.

**Can a business refuse an ADMT opt-out request?**
It can offer a human appeal instead, where a qualified reviewer can overturn the decision. Narrow exceptions also exist for some hiring, admission and work-allocation uses.

## Key facts

California CCPA ADMT regulations: OAL approved 23 Sept 2025; effective 1 Jan 2026; risk assessments from 1 Jan 2026 with attestations due 1 April 2028; ADMT duties from 1 Jan 2027 for significant decisions (financial/lending, housing, education, employment/contracting/pay, healthcare). ADMT = processes personal information and replaces or substantially replaces human decision-making; meaningful human review requires understanding, weighing other information and authority to change the decision. Duties: pre-use notice, opt-out (two methods) or human appeal, access right. Risk assessments reviewed every 3 years or within 45 days of material change. Opencontroller provides evidence via discovery, registry, on-behalf-of identity, LLM judges and guardrails.

## Related

- [What does the Colorado AI Act mean for AI agents?](https://opencontroller.lyzr.com/blog/colorado-ai-act-and-ai-agents.md)
- [How do you run an AI impact assessment for an AI agent?](https://opencontroller.lyzr.com/blog/ai-impact-assessment-for-ai-agents.md)
- [What evidence should you keep when a human approves an AI agent's output?](https://opencontroller.lyzr.com/blog/evidence-for-human-review-of-ai-agent-outputs.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
