A view from DC: Navigating the California and Colorado ADMT rules

New obligations for AI that makes automated decisions go into effect 1 Jan. 2027.

Contributors:
Cobun Zweifel-Keegan
CIPP/US, CIPM
Managing Director, Washington D.C.
IAPP
Editor's note
The IAPP is policy neutral. We publish contributed opinion pieces to enable our members to hear a broad spectrum of views in our domains.
The beginning of 2027 marks the real start of the compliance era for automated decision-making technologies, with two important milestones passing on the same day.
In Colorado, 1 Jan. marks the implementation deadline for this year's updates to the state's AI law, SB 26-189, which practitioners refer to as the Colorado ADMT Act. The hard-fought law repealed and reenacted Part 17 of the Colorado Consumer Protection Act, replacing the never implemented Colorado AI Act, SB 24-205. Though narrower than its predecessor, the law retains operational significance for covered systems, and interplays with California's rules in interesting ways.
California's finalized regulations for ADMT under the California Consumer Privacy Act also go into effect 1 Jan. These long-awaited rules give real substance to what was essentially a placeholder in the state's privacy law.
Vastly different thresholds
Many automated systems may fall out of scope of California's rules by operation of the underlying law's scope. The regulations only apply to covered businesses under the CCPA, so if a company is beneath all three thresholds — USD28 million in annual gross revenue, approximately, as the official 2027 adjusted revenue threshold has not been published; processing personal information of 100,000 or more consumers or households a year; or deriving more than half its revenue from selling or sharing personal information — its ADMT system is not covered.
Colorado offers no such exclusion. If ADMT meets the definitions under the law, the compliance obligations apply, regardless of the size of the organization or its privacy operations. Just like Colorado's privacy law, the ADMT Act also applies to nonprofits.
Qualifying as "ADMT" is a relatively high bar in California. The system must process personal information and use computation to substantially replace human decision-making in a covered domain. "Substantially replace" means the business uses the output to make a significant decision without human involvement. "Human involvement" gets its own three-part test: A human reviewer must know how to interpret and use the output, must review and analyze the output and any other relevant information and must have the authority to make or change the decision based on that analysis.
Colorado takes a broader approach, but with its own set of nesting-doll definitions. Covered ADMT is technology that processes personal data and uses computation to generate an output that is used to make, guide or assist a consequential decision about a person.
Note the verbs: Make, guide or assist. A tool does not need to decide anything on its own to count. Instead, the limiting factor in Colorado is whether the tool "materially influences" a decision, that is, whether the output is a non-de minimis factor in the decision and affects the outcome. Incidental, trivial or clerical uses do not count.
Covered decisions: Significant vs. consequential
The covered domains for decisions under the two sets of rules also create a tangled web of overlaps. Colorado uses the term "consequential decision," while California applies to "significant decisions."
Both state rules apply in the context of these domains: financial or lending services; housing; education enrollment or opportunity; employment and hiring; and healthcare services.
Colorado also covers insurance and essential government services and public benefits.
And there's one more scope distinction because the outcomes of the decisions matter. California's scope is limited to the "provision or denial" of services within covered domains. Colorado applies to decisions that determine access to, eligibility for, selection for, or compensation within a covered domain. It also applies to any "decision about differentiated price, cost sharing, compensation, or other material terms" if the decision is "reasonably likely to materially limit, delay, effectively deny, or otherwise fundamentally alter the consumer's access" to one of the covered domains.
Pre-use and post-decision notices
There are two distinct forms of notice envisioned in both regimes.
California requires a pre-use notice, delivered prominently and conspicuously at or before the point when the business collects the personal information it plans to process with ADMT. The notice must also describe the consumer's right to opt out and how to exercise it, describe the right to access information about the ADMT and how to exercise it, state that retaliation is prohibited and provide additional information about how the ADMT works and how the decision would be made if the consumer opts out. This includes the categories of personal information that affect the output, the type of output and how it is used, whether it is the sole factor or one among others, what the human's role is and what the alternative process is for people who opt out.
Colorado also requires notice before a deployer uses a covered ADMT to materially influence a consequential decision, but it is more limited. This is a clear and conspicuous notice that the deployer used or will use a covered ADMT, with instructions for obtaining additional information. So, the deployer can satisfy that duty with a prominent public notice, reasonably accessible at points of consumer interaction. The more detailed disclosure in Colorado comes later, after an adverse outcome, as part of the process of providing consumers with statutory rights.
Rights and wrongs
In Colorado, after an adverse outcome — a decision that denies, terminates, revokes or materially reduces access to an opportunity or service, or imposes materially less favorable terms — the deployer must provide a plain-language description of the decision that includes, among other things, the role the covered ADMT played in the decision and an explanation of the consumer's rights and how to exercise them. The consumer may then request instructions for correcting factually incorrect or materially inaccurate personal data used in the decision. Opinions, predictions, scores and protected evaluations are not correctable.
California also gives consumers information about ADMT decisions under a right to access, including the specific purpose of the system, how the ADMT processed their information to generate the output — the parameters and the specific output itself — the outcome and how the output was used, and how the output will be used in future decisions. Though California's law includes a standalone right to correct, the regulations do not mention correction as a right in the ADMT context.
California goes farther by requiring businesses either to provide an opt-out of ADMT for significant decisions using an alternative process or, post-facto, a human appeal to a reviewer with authority to overturn the decision. Importantly, this requirement does not apply in situations where: the ADMT is used solely to assess ability to perform for an admission, acceptance or hiring decision and works for that purpose without unlawfully discriminating; or it is used solely for allocation or assignment of work and compensation under the same conditions.
Colorado does not provide an opt-out right but does require an opportunity for meaningful human review and reconsideration "to the extent commercially reasonable." To qualify as "meaningful," the review must be conducted by a designated individual with authority to approve, modify or override the decision, like in California. Colorado goes further by specifying that the human must consider relevant available primary evidence, be trained to conduct the review, not default to the system output and have access to enough information to understand the output's intended use, material limitations, categories of inputs and principal factors.
Risk assessments and developer disclosures
The gone-but-not-forgotten Colorado AI Act included risk assessment requirements, but its replacement has eschewed them. At the same time, California's regulations have gone in the other direction because ADMT is one of the triggers for the newly applicable risk assessment requirements.
Under the same regulatory package as the updated ADMT rules, using ADMT for a significant decision is a processing activity that presents "significant risk" to consumers' privacy, so a business must conduct a risk assessment before using the ADMT. The duty to conduct a risk assessment can actually kick in even earlier, because processing personal information the business intends to use to train an ADMT for a significant decision is also a trigger.
In either case, for processing that was already underway when the regulations took effect at the beginning of 2026, the assessment must be documented by 31 Dec. 2027, with report-outs submitted to CalPrivacy by 1 April 2028. New processing must be documented before it occurs.
What Colorado does provide for is developer documentation. By the beginning of 2027, a developer must share a packet with each deployer of a covered system that includes a general statement of intended uses and known harmful or inappropriate uses, the categories of data used to train the system, known limitations including known risks and circumstances where the system should not be used, instructions for appropriate use, monitoring and meaningful human review, and information reasonably necessary for the deployer to comply with its disclosure duties. Developers must also give notice of material updates, intentional and substantial modifications and changes to intended use, limitations or risk mitigation within a reasonable time. Both developers and deployers are subject to three-year records retention requirements.
California doesn't provide for these types of duties, though providers may have to assist with access requests or honoring opt outs, if relevant.
Unfinished business
Numerous exceptions and exemptions apply under both regimes, including for businesses subject to existing compliance obligations in highly regulated industries.
And there are two other important caveats for the Colorado law. First, the attorney general must adopt rules by 1 Jan. 2027 clarifying the post-adverse-outcome disclosure requirements and the consumer-rights requirements. Draft rules are posted on the Colorado attorney general website, but final rules have not yet been released.
Second, pending litigation over the Colorado law has not gone away, even after the amended law was passed. xAI, now SpaceXSI, sued over the Colorado AI Act, but the court's initial order in that case bars enforcement of that act "or any legislation replacing or amending it enacted during the 2026 session." Under the court's order, xAI has 28 days after the final adoption of implementing rules to file its preliminary injunction motion. So, the enforceability of the law remains in question until the final rules are adopted.
Even so, AI governance teams should be prepared for these and similar requirements to continue popping up. ADMT rules are not going anywhere.
Please send feedback, updates and significant decisions to cobun@iapp.org.
This content is eligible for Continuing Professional Education credits. Please self-submit according to CPE policy guidelines.
Word count: 1,684

Contributors:
Cobun Zweifel-Keegan
CIPP/US, CIPM
Managing Director, Washington D.C.
IAPP



