Three sets of AI obligations switch on Jan. 1, 2027: Colorado's Chatbot Safety Act, Colorado's rewritten automated decision-making law, and California's ADMT rules. That's 51 days from today (19 left in November + 31 in December + 1 = 51), or 7 weeks and 2 days.
When we wrote about Colorado on Oct. 16, the comment window was open. It's closed now. Here's what changed, what California adds, and one correction.
This is a plain-language summary, not legal advice. Check your situation with your own counsel.
Where Colorado stands today
The rules still aren't final. The Attorney General's office filed proposed rules (4 CCR 904-6) with the Secretary of State on Aug. 11, released an interim draft Oct. 6, posted a regulatory analysis Oct. 21, and held the formal hearing Oct. 26. Written comments closed at 11:59 p.m. that night (Colorado AG). Today the Secretary of State's docket lists the proposal and hearing, and its adopted rule section is empty (Colorado SOS eDocket 2026-00360).
The statute tells the AG to adopt rules on adverse outcome disclosures and consumer rights by Jan. 1, 2027 (SB 26-189, C.R.S. 6-1-1704(4)(b) and 6-1-1705(3)). Anything below citing the interim draft is a draft.
The dates keep moving. The original Colorado AI Act, SB 24-205, had already been pushed to June 30, 2026 by SB 25B-004. SB 26-189, signed May 14, 2026, repealed and reenacted it, takes effect Jan. 1, 2027, and applies to decisions made on or after that date (leg.colorado.gov; signed act, Section 5).
In xAI's federal suit against Attorney General Phil Weiser, an April 27 order says the AG won't start enforcement, investigations included, for violations of SB 24-205 or legislation replacing it that occur on or before 14 days after the court rules on xAI's preliminary injunction motion. xAI files that motion within 28 days after final rules are adopted (D. Colo. minute order). No final rules, no clock. Whether that order reaches the separate chatbot law is a question for your lawyer, not us.
The AG's cover letter to the interim draft adds this (Colorado AG):
While the laws go into effect on January 1, 2027 per the statute, the Attorney General's policy has been to focus on egregious and intentional violations of the law.
That's a policy, not a safe harbor.
The chatbot law: read the exemptions first
From Jan. 1, HB 26-1263 requires an operator to disclose that its chatbot is AI at the start of a user's first interaction each day, at least every three hours in a continuous chat (or with a persistent label), and whenever someone asks. It requires a suicide and self-harm referral protocol and bans presenting output as equivalent to a licensed health care, legal or mental health professional or a dietitian. Operators must also estimate users' ages, with extra duties for known minors, and annual reports to the AG start July 1, 2027 (HB 26-1263, C.R.S. 6-1-1708(2) to (6)). It's a statute. The AG says rulemaking isn't required but is doing it anyway (Colorado AG).
Here's our correction. The definition of a covered chatbot excludes software "primarily designed to provide commerce-related or transactional assistance," naming shopping, ordering, payments, delivery, returns and customer support (HB 26-1263, C.R.S. 6-1-1701(3.5)(b)). It also excludes bots that handle a "narrow and discrete topic" and can't drift into sexual content or self-harm. The interim draft's examples are a purchase status update, a billing question and appointment scheduling (interim draft, Rule 9). Our October checklist treated most site chatbots as in scope. A bot that only answers order and returns questions probably isn't. A general "ask me anything" assistant is the closer call.
Turn the label on anyway. Intercom's Fin has a "Show AI Agent label in Messenger" toggle, and Intercom says it's off by default for newer deployments and that customers are responsible for telling users they're talking to AI (Intercom). Zendesk shows a "Generated by AI" label on the AI agent's name in its Web Widget (Zendesk). HubSpot's customer agent shows "Powered by AI" in the chat header (HubSpot).
Colorado ADMT: notice up front, rights after a no
SB 26-189 covers tech that "materially influences" decisions in education, employment, housing, lending, insurance (pricing included), health care and government services. Pricing counts when less favorable terms materially limit access in one of those areas. Customer service triage, advertising, and tools that only summarize or draft for a human are carved out (SB 26-189, C.R.S. 6-1-1701).
If you're in, the deployer duties are a clear notice before use (a prominent posting near the point of interaction qualifies) and a plain-language explanation within 30 days of an adverse outcome. After an adverse outcome, on request, you owe a way to correct factually incorrect data plus "meaningful human review," the review to the extent commercially reasonable. Keep records three years (SB 26-189, C.R.S. 6-1-1703 to 6-1-1705). Before enforcing, the AG must give notice and 60 days to cure where it deems a cure possible, except for knowing or repeated violations, until that provision sunsets Jan. 1, 2030 (C.R.S. 6-1-1706(3)). SB 26-189 doesn't add an opt-out right. The interim draft would require confirming a review request within 10 days and finishing within 45 (interim draft, Rule 8).
The interim draft's own example: a manager sees ADMT scores for internal candidates, interviews them all, and picks the top score. The score is presumed to have materially influenced the decision (interim draft, Rule 6). Resume scoring in an applicant tracking system looks a lot like that.
California: narrower list, bigger businesses
First, check whether the CCPA covers you. It applies to for-profit businesses with over $26,625,000 in prior-year revenue, or that buy, sell or share personal information of 100,000 or more consumers or households, or that earn half their revenue selling or sharing it (CPPA; Civil Code 1798.140(d)). That figure is inflation-adjusted every odd-numbered year, so watch for a January update.
The rules were approved Sept. 22, 2025 and took effect Jan. 1, 2026, with ADMT compliance starting Jan. 1, 2027 (CPPA; CPPA announcement). ADMT here means tech that replaces or substantially replaces human decisionmaking. A human counts only if they can interpret the output, review it, and have authority to change the decision. A "significant decision" covers lending and financial services, housing, education, employment and health care. Insurance isn't on the list, general pricing isn't, and advertising is expressly out (CCPA regulations, §7001).
On Jan. 1 you owe a pre-use notice, an opt-out, and an access response explaining the purpose and logic (§§7200, 7220 to 7222). You can skip the opt-out by offering an appeal to a human with authority to overturn, and hiring tools used solely to assess ability to do the job, that work and don't unlawfully discriminate, get an exception too. Opt-outs need at least two methods. A cookie banner alone doesn't count (§7221).
Risk assessments aren't a Jan. 1, 2027 item. They've been required before starting covered processing, including ADMT for significant decisions, since Jan. 1, 2026. Processing that began earlier must be assessed by Dec. 31, 2027, and summaries go to the CPPA by April 1, 2028 (§§7150, 7155, 7157). If you're covered and started that kind of processing this year without one, you're already behind.
Seven weeks, by week
Week 1 (Nov. 11 to 17). List every AI tool that talks to the public or scores people. Note your customers' and applicants' states.
Week 2 (Nov. 18 to 24). Decide whether your chatbot is support-only or general purpose. Turn on the vendor's AI label either way.
Week 3 (Nov. 25 to Dec. 1). Thanksgiving week. Draft the Colorado point-of-interaction notice and, if covered, the California pre-use notice.
Week 4 (Dec. 2 to 8). Build the intake: correction and human review for Colorado, opt-out or human appeal plus access for California.
Week 5 (Dec. 9 to 15). Ask vendors for the documentation Colorado requires developers to provide: intended uses, training data categories, known limits, and review instructions (SB 26-189, C.R.S. 6-1-1702).
Week 6 (Dec. 16 to 22). Name the human reviewers and confirm they can actually overturn a decision. That's the review time we warned costs real hours.
Week 7 (Dec. 23 to 29). Check the SOS docket and coag.gov for final rules and adjust. Dec. 30 and 31 are your two days of slack.
Want us to look at whether your chatbot or screening tool is in scope? Write to us: team@lirium.ai.