Blog

California ADMT Rules 2027: Does Your AI Agent Count?

From 1 January 2027, California businesses covered by the CCPA that use automated decisionmaking technology for significant decisions owe notices, opt-outs and access rights. Here is how to tell if your AI agent counts.

Atiq Rezaul Hoque Turjo
Atiq Rezaul Hoque Turjo
8 min read
California ADMT Rules 2027: Does Your AI Agent Count?

Your AI agent screens job applicants, ranks rental applications or pre-qualifies loan leads. Does California now treat it as the decisionmaker?

That depends on what the agent decides, how people use its output, and whether your business is covered by the CCPA at all.

California ADMT rules require businesses covered by the CCPA that use automated decisionmaking technology for significant decisions to give consumers specific rights from 1 January 2027. Using AI does not by itself make an agent ADMT. What matters is whether it processes personal information, replaces human decisionmaking and affects a covered decision.

For a small business, the practical job is to inspect each workflow before deciding what to build. This is general information, not legal advice; have California privacy counsel confirm your position.

When do the California ADMT rules take effect?

There are three dates to keep apart. On 23 September 2025, the California Privacy Protection Agency (CPPA) announced that the Office of Administrative Law had approved updated regulations covering the CCPA, cybersecurity audits, risk assessments and automated decisionmaking technology. The regulations took effect on 1 January 2026, but not every obligation started that day.

Date What it means
1 January 2026 The updated regulations took effect, and risk assessment requirements began.
1 January 2027 Businesses using ADMT for significant decisions must comply with the ADMT requirements, including ADMT already in use before that date.
1 April 2028 The risk assessment attestation and summary are due to the CPPA.

So an applicant-screening system running since 2025 is not outside the rules. The official rulemaking materials for the CCPA ADMT regulations are on the CPPA regulations page.

Does the CCPA cover your business in the first place?

Check this before classifying any agent. The ADMT obligations apply to businesses covered by the CCPA, not to every company that uses AI.

According to the CPPA FAQ, the CCPA applies to for-profit businesses that collect consumers' personal information, decide why and how it is processed, do business in California and meet at least one threshold:

  • Gross annual revenue above the threshold for the preceding year. Check the current revenue threshold on the CPPA FAQ.
  • Buying, selling or sharing the personal information of 100,000 or more California residents or households.
  • Deriving 50% or more of annual revenue from selling or sharing California residents' personal information.

Ask whoever handles your privacy compliance to document the coverage decision. If you are covered, move on to the workflow-level analysis.

Need to map where your AI agents make decisions? A NextCTL AI operations audit identifies automated decision points, human handoffs and logging gaps before you redesign a workflow.

What actually counts as automated decisionmaking technology?

Under the approved regulation text, automated decisionmaking technology is any technology that processes personal information and uses computation to replace human decisionmaking or substantially replace it. That is narrower than "uses AI".

To substantially replace human decisionmaking means using the technology's output to make a decision without human involvement. Human involvement has three requirements. The reviewer must:

  1. Know how to interpret and use the output to make the decision.
  2. Review and analyse the output and any other information relevant to making or changing the decision.
  3. Have the authority to make or change the decision based on that analysis.

Take a hiring workflow. An agent reads applications, scores candidates and produces a ranked shortlist. The hiring manager sees only the top names, and everyone else is rejected automatically. It is hard to call that human-reviewed if nobody examines the rejected applications or can change those outcomes.

Now take an agent that summarises applications while a qualified hiring manager reviews the relevant information and makes the selection. That is a different design.

The text also excludes tools such as web hosting, firewalls, spam and robocall filtering, spellchecking, calculators, databases and spreadsheets, provided they do not replace human decisionmaking. A spreadsheet is not ADMT just because it holds applicant data, but a tool's label does not settle the question either. The actual decision process does.

Which AI agent decisions are considered significant?

The regulation defines a significant decision as one that results in providing or denying:

  • Financial or lending services.
  • Housing.
  • Education enrollment or opportunities.
  • Employment or independent-contracting opportunities or compensation.
  • Healthcare services.

Advertising to a consumer is not a significant decision under the final text. An agent drafting ads is doing a different job from one that denies a loan application. But an agent that starts in marketing and drifts into eligibility screening can cross the line.

This table is a starting point for classification, not a legal determination.

AI agent task Touches a significant decision? ADMT, or human involvement? What to build
Screening job applicants Likely Depends; check with counsel Documented human review, decision logs and ADMT request workflows
Ranking rental applications Likely Depends; check with counsel Reviewer authority, written criteria and request handling
Pre-qualifying loan leads Depends; check with counsel Depends; check with counsel Separate sales prioritisation from lending decisions; log both
Setting contractor pay Likely Depends; check with counsel Human compensation review, audit trail and notices
Scheduling appointments No, for ordinary scheduling No, for ordinary scheduling Activity logs and escalation if scheduling starts denying services
Writing ad copy No No Editorial review and normal marketing controls
Filtering spam No No, while it only filters Document its narrow function and watch for scope changes

A lead-scoring agent that helps a salesperson order callbacks is not the same as one that decides whether someone gets a lending service. The useful question is not "does this agent use AI?" but "what decision does this agent cause the business to make?"

Miniature sorting table with house, briefcase and schoolbag trays representing California ADMT decision categories.

Housing, employment and education are among the significant-decision categories covered by the regulations.

What must covered businesses provide from January 2027?

For covered businesses using ADMT to make significant decisions, the regulations set four main obligations.

A Pre-use Notice. Consumers must get a Pre-use Notice before the business uses ADMT for a significant decision about them. Write it around the actual workflow, not a generic AI disclosure.

A right to opt out. Consumers can opt out of ADMT used for a significant decision, with listed exceptions. The best known is where the business offers an appeal to a human reviewer who meets the regulation's conditions: the reviewer must consider the consumer's information, know how to interpret the output and have authority to overturn the decision, and the appeal must be easy to submit. A nominal review button does not meet that bar.

A right to access. Consumers can request meaningful information about how the ADMT worked and how it affected them. That is an engineering question: if someone asks, can you pull the relevant inputs and outputs for their case? A final status field alone may not be enough.

A risk assessment. Using ADMT for a significant decision is one of the activities that requires a risk assessment, and those requirements began in 2026.

On timing, the CPPA FAQ says businesses must confirm receipt of requests to delete, correct or know within 10 business days and respond within 45 calendar days, with a possible 45-day extension if they notify the consumer. Confirm with counsel which timelines apply to each ADMT request type.

How should you design AI agents for California ADMT?

The most useful design decision is to separate automation from decision authority. An agent can collect information, draft summaries, flag missing fields and organise cases without making the final significant decision, and that separation has to exist in the software, not only in a policy.

Map every consequential decision. Mark the point where an agent's output changes someone's eligibility, access, opportunity or pay. For each, record the inputs, the output, the downstream action and who decides. Use the significant-decision list as the checklist.

Give reviewers real authority. Where human review of AI decisions is the design, the reviewer needs enough information to analyse the recommendation, the knowledge to understand it and the power to change it. Record what they reviewed and what they decided. A screen with only an Approve button is unlikely to meet the standard. Our guide to AI agent governance in production covers wider operating controls, and our guide to human approval gates for AI agents covers handoff design.

Build consumer rights into the workflow. Where ADMT obligations apply, design the notice, opt-out, access and appeal paths before launch. A request should link to the right decision record and start the right review. Opt-out, access and appeal are different actions; the system must tell them apart.

Keep a usable decision history. Retain scoped records of the inputs used, the output, the reviewer and what they considered, the final decision and whether it differed from the recommendation, and any consumer requests. Logs must follow your own privacy, security and retention rules; do not hoard personal information just to build an audit trail.

Keep low-risk tasks separate. Where you can, limit agents to drafting, sorting and summarising, with consequential decisions in a distinct review step. This is worth planning from the start of any custom AI agent development. It makes the system easier to inspect and test, though it does not remove ADMT obligations on its own.

Tiny figure pulling a paper folder from an automated conveyor onto a separate table for human review.

A human review path needs a reviewer who can examine the information and change the decision.

What should your 2027 readiness plan include?

Start with an inventory, not a compliance dashboard. AI agent compliance begins with knowing which systems process personal information and which outputs feed decisions.

  • [ ] Inventory: list current and planned AI agents, scoring systems and decision workflows.
  • [ ] Confirm coverage: check whether the business is subject to the CCPA.
  • [ ] Classify: map each workflow against the significant-decision categories.
  • [ ] Trace decisions: find where automated output becomes a business decision.
  • [ ] Check human involvement: verify reviewer knowledge, information and authority.
  • [ ] Choose the path: decide whether real human involvement exists or ADMT obligations apply.
  • [ ] Prepare notices: draft Pre-use Notices and assess exceptions with counsel.
  • [ ] Build request handling: opt-out, access and human-appeal workflows.
  • [ ] Build logging: capture outputs, decisions, reviewer actions and requests with privacy controls.
  • [ ] Test end to end: run rejected applicants, disputes, opt-outs, access requests and overrides.

Test a complete case, from sample application to simulated consumer request. Can you reconstruct the decision, name the reviewer and answer the request? Repeat the review whenever an agent's job changes.

Good AI workflow automation governance makes these boundaries visible and maintainable instead of relying on assumptions about how the system behaves.

The next step: inspect the decision, not the AI label

The California ADMT rules turn on one question for 2027: does automated output help a qualified person decide, or does it decide for them?

NextCTL LLC helps businesses assess technical design, find decision boundaries and build the controls their legal team specifies. Explore our AI governance services, or contact NextCTL for a free consultation.

FAQ

Frequently Asked Questions

FAQ

What is California ADMT?

California ADMT means automated decisionmaking technology under the CCPA regulations: technology that processes personal information and uses computation to replace or substantially replace human decisionmaking. The definition is about what the system does, not whether it is marketed as AI. Using ADMT for significant decisions triggers notice, opt-out and access rights, subject to the regulation's conditions and exceptions.

When do the California ADMT rules start?

The updated regulations took effect on 1 January 2026, when risk assessment requirements also began; the attestation and summary are due by 1 April 2028. Businesses using ADMT for significant decisions must comply with the ADMT requirements from 1 January 2027. That includes ADMT already in use before 2027, so existing workflows need review too.

Does a human-in-the-loop take my AI agent out of the ADMT rules?

Not automatically. Output used without human involvement counts as substantially replacing human decisionmaking. Human involvement means a reviewer who knows how to interpret the output, reviews it with other relevant information and has authority to make or change the decision. A person clicking approve on every recommendation is unlikely to qualify. The actual workflow decides it.

Do the ADMT rules apply to marketing and ads?

Advertising to a consumer is not a significant decision under the final regulation text, so an agent that writes ad copy or runs ordinary campaigns does not become significant-decision ADMT just because it uses AI. A marketing workflow that also decides access to lending, housing or another covered category needs separate analysis, and other CCPA obligations may still apply.

Continue Reading

Independent check

Decide with evidence, not a pitch.

Open a neutral research prompt in the AI you already use. It asks where NextCTL fits, where a freelancer or an in house hire may be enough, and what to verify before you choose.

AI answers can be incomplete. Check nextctl.com/methodology before deciding.

See where your operations
are leaking revenue.

Tell us what you want to fix. We show where missed calls and slow follow up cost you most, with a written plan for what to fix first.

Or run a free tool first:Website AuditAI Readiness Scan