Feature AI Risk
Colorado's Draft ADMT Rules Put a 45-Day Clock on Human Review
Colorado's interim ADMT rules would require acknowledging human-review requests in 10 days and finishing in 45. Here's the review queue to build by Jan. 1.
Table of Contents
- Colorado repealed and replaced its AI law. The replacement bans nothing and requires no impact assessment — it requires a review queue with a clock. The October 6 interim draft rules would give you 10 days to acknowledge a human-review request and 45 days to finish it.
- The scope question is now “materially influence,” and the draft rules list six considerations that decide it. Fourth on the list: whether the reviewer actually has authority to change the decision. Third: whether they do independent analysis. A human-in-the-loop that rubber-stamps is not a defense.
- Enforcement is stayed under a April 27 court order and the final rules aren’t adopted, with six weeks to go. Build to the statute, not the rules, and put the ticket clock in now — the clock is the cheapest part and the hardest to retrofit.
The cheapest control is the one nobody has
Ask a lender for its human review workflow and you will get a process map, a policy section, and a name. Ask what happens on day eleven of a review request and the room goes quiet.
That gap is about to get expensive. Colorado’s replacement AI statute, SB 26-189, repealed and reenacted the 2024 Colorado AI Act and takes effect January 1, 2027. The Attorney General’s interim draft rules, released October 6, 2026, would require a deployer to confirm receipt of a human-review request within 10 days and complete the review and response within 45 days.
Ten and forty-five. Not “promptly.” Not “in accordance with policy.” Two numbers, one ticket field, and a clock that starts whether or not your intake route was designed for it.
Most review queues cannot report either number today. That is the specific thing to fix, and it is fixable before the rules are final.
If you already built for the old law, the explainer on what SB 26-189 requires covers the statutory duties. This piece is about the workflow underneath them.
What the statute actually demands, in order
C.R.S. 6-1-1704 and 6-1-1705 sequence four obligations. The sequence matters because each one produces an artifact the next one consumes.
| Section | Duty | Artifact it creates |
|---|---|---|
| 6-1-1704(1)–(2) | Clear, conspicuous notice before the ADMT is used; a prominent public posting near the point of interaction can satisfy it | Dated notice copy, placement screenshot, URL |
| 6-1-1704(3) | Within 30 days of an adverse outcome: plain-language description of the decision and the ADMT’s role, how to request more information, and the consumer’s rights | Adverse-outcome letter template and send log |
| 6-1-1705(1) | On request: correction of inaccurate personal data, and meaningful human review and reconsideration “to the extent commercially reasonable” | Correction ticket, review ticket, reviewer identity, outcome |
| 6-1-1703 | Retain compliance records three years after each consequential decision | The retained file itself |
Note what 6-1-1705(1) does not say. It does not say the reviewer decides in the consumer’s favor. The statute’s definition of meaningful human review, in 6-1-1701, says the reviewer must be trained to conduct the review, must consider relevant available primary evidence, must have authority to approve, modify, or override — and must not default to the system output.
“Must not default” is doing a lot of work in one sentence.
How scope gets decided now
The old law asked whether a system was “high-risk.” The replacement asks whether the ADMT “materially influences” a consequential decision, defined in 6-1-1701 as a non-de minimis factor that affects the outcome by constraining, ranking, scoring, recommending, or classifying.
At the October 26 rulemaking hearing, industry and consumer groups split on this, according to Wiley’s account of the hearing. Several commenters argued material influence should turn on actual reliance and the output’s independence from human oversight, not the existence of an output. Others backed the draft rules’ presumptions.
The draft rules answer it with six considerations for whether a factor is de minimis. They are worth reading as a control checklist, because each one is something you can build evidence for:
- Whether anyone actually reviews the output before the decision.
- Whether that person does independent analysis.
- Whether they have subject-matter understanding of the domain.
- Whether they have authority to change the decision.
- Whether marketing, procurement, or training materials frame the output as authoritative.
- Whether the data volume or complexity exceeds what a human could replicate.
Four of the six are about the human, not the model. If your credit analysts see a score, understand the product, can override it, and write their own reasoning, you have a materially different scope position than a team that clicks approve on a recommendation engine built to be trusted.
Two of the six are organizational honesty tests. Number five means your vendor deck can put you in scope: if procurement’s own evaluation says the model “makes” the decision, the reviewer’s independence claim gets harder. Number six is why “a human reviews every file” fails as a blanket answer in high-volume consumer lending — nobody meaningfully reviews 40,000 monthly decisions one at a time.
Build the queue, not the rhetoric
The workflow below runs on the statute’s own deadlines and survives whatever the final rules say, because the rules can shorten a clock but they cannot repeal the duty.
Week one: find the intake. Most institutions have no dedicated route. Requests arrive as complaints, branch escalations, emails to a servicing inbox, and CFPB portal referrals. Pull ninety days of complaints containing “why was I denied,” “score,” “model,” or “system” and trace where each landed. Every path that dead-ends is a missing intake route.
Week one, same pass: the request register. One table, six columns: received date, decision ID, request type (correction or review), intake channel, assigned reviewer, clock date. The clock date is received date plus 45. Make it a formula so nobody types it.
Week two: name the reviewers. Not a team — named people, per decision type, with a documented delegation of override authority. If the analyst who ran the model is the reviewer, say so and describe the independence control. If the reviewer needs a manager’s approval to change an outcome, the authority to override isn’t theirs and 6-1-1701 says so.
Week three: the acknowledgment. Ten days means most acknowledgments go out before anyone has looked at the file. That is the point. The acknowledgment states receipt, names a reviewer, gives the date by which the consumer will hear back, and explains what is being reviewed. It does not concede error.
Week four: the review record. A review that produces one line — “upheld” — will not read as meaningful. The record needs the primary evidence the reviewer considered, the factors they weighed, whether they changed anything, and if not, why the original outcome holds. The override authority test goes deeper on what separates a real review from a signature.
Week five: correction handling. Inaccurate data is correctable; opinions, predictions, and scores are not. That means the intake has to sort fact from judgment at the door. A wrong employer name is a correction. A low propensity score is not. Get this wrong in either direction and you either rewrite the model for one borrower or refuse a legitimate fix.
The two traps
The commercial-reasonableness escape hatch. “To the extent commercially reasonable” is in the statute, and finance will find it. Write down what it means before the first appeal: which decision types get full reconsideration, which get a documented review only, and who signs that scoping. A scope decision made case by case, after the consumer complains, is not a control.
The 60-day cure is not a rehearsal window. Under 6-1-1706(3)(a)–(b), before an action the Attorney General must give a notice of violation and 60 days to cure if a cure is deemed possible. But 6-1-1706(3)(c) lets the AG skip the cure entirely for knowing or repeated violations. Miss your 45-day clock on a whole population rather than one file and you are arguing about repetition, not arithmetic. The cure subsection itself sunsets January 1, 2030 (6-1-1706(3)(f)).
What the exemptions don’t cover
6-1-1708 gives real carve-outs. Insurers subject to C.R.S. 10-3-1104.9 are deemed compliant. HIPAA covered entities are excluded except for employment decisions. GLBA-protected NPI need not be disclosed.
None of those are a pass for a bank’s lending or hiring stack. And there is no small-business exemption in the enacted text, unlike the 2024 version’s under-50-employee carve-out.
Enforcement is also paused, which is a scheduling fact and not a scope fact. The April 27, 2026 minute order in X. AI LLC v. Weiser bars the AG from initiating enforcement, investigations included, for violations occurring on or before 14 days after the court rules on xAI’s preliminary injunction motion — a motion due within 28 days after final rules land. The obligations still start January 1, 2027. When the stay lifts, the decisions your team made last quarter are already on the record.
So what
The final rules will move. Six weeks out from an effective date, that is the worst possible reason to wait. The 10-day and 45-day clocks, the named reviewer with override authority, and the three-year retained file are all buildable from the statute text today, and none of them get cheaper after January 1.
Start with the request register and one quarter of backward-looking intake tracing. That is a two-week job, it needs no vendor, and it will produce the first honest number your program has ever had on review volume. Our AI Risk Assessment Template & Guide includes the use-case scoping and reviewer-authority fields to hang it on.
FAQ
What does Colorado’s ADMT law require for human review?
C.R.S. 6-1-1705(1) requires that, on request after an adverse outcome, a deployer let the consumer correct inaccurate personal data and provide meaningful human review and reconsideration to the extent commercially reasonable. Reviewers must be trained, hold authority to approve, modify, or override, consider primary evidence, and not default to the system output. The October 6 interim draft rules would add a 10-day acknowledgment and 45-day completion clock.
Does a human approving every decision take the system out of scope?
Not automatically. Scope turns on whether the ADMT is a non-de minimis factor under 6-1-1701. The draft rules’ six de minimis considerations weigh whether anyone really reviews output, does independent analysis, understands the domain, and can override — plus whether internal materials frame the output as authoritative. A rubber stamp does not defeat scope. Independent review might.
When does the law take effect, and is it being enforced?
Obligations apply to consequential decisions made on or after January 1, 2027, and the AG must adopt rules before that date. Enforcement is currently stayed: the April 27, 2026 order in X. AI LLC v. Weiser blocks the AG from initiating enforcement, investigations included, for violations through 14 days after a ruling on xAI’s preliminary injunction motion.
How long must deployers keep records?
Three years after each consequential decision, under C.R.S. 6-1-1703. Practically, the retained file holds the vendor documentation package, the notice served, the decision record, the review ticket, and the review outcome.
Are banks and insurers exempt?
Insurers under C.R.S. 10-3-1104.9 are deemed compliant, and HIPAA entities are excluded except for employment decisions. There is no exemption for banks, credit unions, or lenders in the financial or lending services domain, and no small-business carve-out in the enacted text.
◆ Need the working template?
Start with the source guide.
These answer-first guides summarize the required fields, evidence, and implementation steps behind the templates practitioners search for.
◆ Related template
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Immaterial Findings · Weekly
Sharp risk & compliance insights. No fluff.
◆ FAQ
Frequently asked questions.
What does Colorado's ADMT law require for human review?
Does a human approving every decision take an AI system out of scope?
When does the Colorado ADMT law take effect?
Is Colorado enforcing right now?
How long do deployers have to keep records?
Author
Rebecca Leung
Rebecca Leung has 8+ years of risk and compliance experience across first and second line roles at commercial banks, asset managers, and fintechs. Former management consultant advising financial institutions on risk strategy. Founder of RiskTemplates.
◆ Related framework
AI Risk Assessment Template & Guide
Comprehensive AI model governance and risk assessment templates for financial services teams.
◆ Keep reading
Related posts.
AI Risk
AI Risk Tiering Just Flipped: OCC 2026-54 Ranks Consequence, Not Sophistication
OCC Bulletin 2026-54 puts a rules-based payment agent in Tier 1 and your frontier model in Tier 4. Build AI risk tiering on agency and reversibility.
Nov 20, 2026
AI Risk
Your Agent Is Logged In as a Service Account Nobody Owns
SR 26-2 pushed agentic AI out of model risk scope. AI agent access control now needs an entitlement register, JIT credentials, and a named owner.
Nov 18, 2026
AI Risk
Retiring an AI Model Is a Control, Not a Cleanup Task
AI model retirement is a control, not a cleanup task. The decommission memo, retention surface table, and inventory states that survive an exam.
Nov 15, 2026