Colorado's Automated Decision-Making Technology Act takes effect on 1 January 2027. Senate Bill 26-189 repeals and re-enacts the state's 2024 AI Act with new requirements for any business that uses automated decision-making technology (ADMT) in a consequential decision about a Colorado resident. The draft rules illustrate the covered domains with examples from employment, lending, housing, insurance and education. The Attorney General must adopt rules to implement it on or before the same date, and the rulemaking is nearly done. Proposed rules were filed on 11 August, an interim draft followed on 6 October, a regulatory analysis on 21 October, and the formal hearing was held, and written comments closed, on 26 October.
The final text is not out yet, and details may still move. The proposed rules of 11 August put the point in one sentence: a deployer that is not able to explain how the technology materially influenced the decision, or how it used the person's data, has not complied with the disclosure duty. The 6 October interim draft deleted that sentence, but not the duty behind it. The statute still requires a plain-language description of the role the technology played, and the interim draft adds that the description must be accurate. You cannot describe a role accurately if you never recorded what it was. That is not a paperwork obligation. It is a test of whether a record exists.
The presumption decides your scope, not your intent
Most organisations will begin by asking whether their system is in scope. The draft answers for them. Under Rule 6.1, an ADMT output is presumed to materially influence a consequential decision if it constrains an option set, sets a threshold, or produces a rank, score, classification, recommendation, prediction or other inference that pertains to the individual, is reviewed by the decision maker (or screens the data made available to them) before or during the decision, and is consistent with the outcome.
Read the third limb carefully. Every time a person who reviewed the output reaches the same outcome, the presumption applies. The draft's own employment example makes the point: a hiring manager who interviewed every candidate, read every review and work sample, and also saw the score, is still presumed to have been influenced by it.
The presumption is rebuttable. The evidence that rebuts it is the useful part, because it reads as a specification:
The decision maker did not consult the output. Provable only if access was logged.
A hard cut-off or threshold was applied, consistently. Provable only if you can show the consistency, which means every decision, not the one in dispute.
An independent review of substantial other evidence, by someone with the authority and subject knowledge to make the call, reaching an outcome consistent with that evidence.
The decision went against the output.
Other persuasive evidence of independent judgement, contemporaneous evidence included.
Every route out of the presumption runs through evidence, and the draft names contemporaneous evidence for a reason. A statement drafted after a complaint arrives is reconstruction, and it will carry far less weight than a record made at the time.
The presumption is not about whether the model decided. It is about whether you can show it did not.
One more detail deserves attention from anyone who buys these systems. Among the factors the draft lists for judging whether an output was minor are your own marketing, procurement contracts, internal policies and training materials, and whether they tell staff to treat the output as authoritative. The business case you wrote to justify the purchase is evidence about how the tool is used. If it promised that the score would drive decisions, it will be read as saying so.
Thirty days to explain a decision in writing
Where a covered system materially influences a consequential decision with an adverse outcome, the deployer has thirty days to tell the person, in writing. The draft is specific about what that letter must hold, and specific about what will not do:
What was decided, in plain language, with the effective date where an existing benefit or arrangement is reduced or ended.
What the system did, with specificity. "Automated review of your personal data" is given as insufficient. "Automated review of credit report and employment verification information" is given as sufficient. The role of any human reviewer has to be described as well.
Any automatic denial. If something in the person's data is an automatic denial under your standards, the description has to say the applicant was automatically denied. The draft's examples name the factor, such as a credit score of 580. The August proposal went further and required the principal reasons, down to disclosing a risk score of 82 out of 100. The interim draft removed that, but reasons return at the review stage, where they must be specific to the person's evidence.
How to get more: an online route to request the system's name, version, developer and the categories and sources of data used, and an explanation of the right to correct data and to request human review.
None of this can be written at the end of the month from memory. Each letter is a query against a decision record: which model, which version, which inputs, which score, which reviewer. If that record does not exist at the point of decision, the description of the system's role cannot be accurate, and the draft requires that it is.
Human review that could change the answer
The draft defines a meaningful human review as one that could change the adverse outcome. A sign-off that only ever confirms is not one. The reviewer needs authority to override, must not be steered by management, and must not use ADMT to assist the review (tools that only summarise, organise, translate or route information are excluded from the definition and remain usable). Receipt is acknowledged within 10 days, and the review and response are due within 45. The statute grants the review to the extent it is commercially reasonable, and the draft presumes it is when the outcome is a severe and irreversible denial of a basic need such as housing.
Each review must leave a record: the reviewer's identity, authority and training, timestamps, the evidence in front of them, their access to the system's intended use, limitations, inputs and principal factors, whether they approved, modified or overrode the decision, and a written justification. The response to the person must give reasons specific to their evidence, "not a recitation of the Covered ADMT's general logic".
Outsourcing the tool does not outsource the duty
Deployers and developers carry different obligations, and the draft is clear about where the line falls. A developer documents intended and known harmful uses, the data the system needs and was trained on, how to monitor it, and how to support meaningful human review. A deployer makes the decision, sends the letter and runs the review.
If a staffing firm runs the model and hands you a recommended shortlist, the draft's example is unambiguous: you are the deployer and you hold every deployer obligation. You may contract for help. You may not contract out of the duty. That makes the developer's documentation the input to your own compliance, and a supplier who cannot tell you what their system weighs has just become your disclosure risk.
What to have in place by 1 January
The rules are not final, so build to the structure rather than to the wording. Everything below is read from the statute and the 6 October interim draft, and the Attorney General may still change the rules before adoption.
1. An inventory of decisions, not of models. List every consequential decision about a Colorado resident where a score, rank, recommendation or threshold is visible to the person deciding. Apply the presumption to each and record your conclusion and its basis.
2. A decision record captured at decision time. Model and version, inputs, output, who saw it, when, and what they decided. This one artefact serves the presumption rebuttal, the thirty-day letter and the human review.
3. Developer documentation on file for every covered system. If a supplier cannot provide it, record that you asked. The draft expects your notice to give as much information as you can from what the developer supplied.
4. A letter template tested against the draft's examples. Run it on a real past decision. If you cannot describe the system's role with specificity, the problem is the record, not the template.
5. A review route with real authority. Named reviewers with the training and the power to override, a 10-day acknowledgement, a 45-day response, and a record of every outcome.
6. Procurement and policy language that says what you mean. If the output is one input among several, make sure your own documents do not say otherwise.
This is a practical reading of a draft, not legal advice, and counsel should confirm scope for your organisation. What counsel cannot do is produce a record that was never kept.
Where this sits
In You cannot staff twenty-seven interpretations we argued that the demands under the US state laws converge on a small set: keep a human meaningfully in the loop for consequential decisions, be able to explain how an outcome was reached, and be able to show it later. Colorado is now writing the operating detail of that sentence, down to the response window and the fields in the review log. And as with the EU, the deadline is not the test. The record is.
We help organisations find where automated outputs reach consequential decisions, design the decision record that makes each of these duties answerable from one source, and put human review in place that can genuinely change an outcome. If you deploy scoring, ranking or recommendation in hiring, lending, housing or insurance decisions that reach Colorado, talk to us before the first adverse outcome letter is due.