On May 9, 2026, Colorado’s legislature passed SB 26-189, repealing the Colorado AI Act and replacing it with a narrower framework governing “automated decision-making technology.” Governor Polis signed the bill into law on May 14, 2026. The new law takes effect January 1, 2027. It represents the end of a two-year arc that began when Colorado enacted the nation’s most ambitious state-level AI regulation in 2024 and ended, under pressure from xAI’s federal lawsuit and the DOJ’s intervention, with something considerably less ambitious.
Most of the commentary on SB 26-189 has focused on what the legislature removed: the mandatory bias audits, the risk management programs, the impact assessments, the deployer obligations that made the original Colorado AI Act the most prescriptive AI governance framework in the country. But here, I examine something the new statute added: an exemption for AI tools used solely to “summarize, organize, translate, draft, route, or present information for human review.” That exemption reveals what the legislature thinks AI does in professional settings. And what the legislature thinks AI does is already two product cycles behind what AI vendors are selling to law firms.
What SB 26-189 kept and what it cut
The original Colorado AI Act, SB 24-205, regulated “high-risk artificial intelligence systems”—any AI system that makes, or serves as a substantial factor in making, a “consequential decision” in one of eight enumerated categories. As I discussed in April, “legal services” was one of those eight categories, which meant that a law firm using AI to generate analysis informing a settlement recommendation or a case-screening decision was a “deployer” subject to the full suite of statutory obligations: risk management policies, annual impact assessments, consumer notifications, public disclosures, and $20,000-per-violation penalties.
SB 26-189 replaces that framework with one built around “covered automated decision-making technology”—a term defined to capture systems that process personal data and generate outputs used to “materially influence” a consequential decision. The covered categories are narrower: employment, education, housing, financial or lending services, insurance, health care, and essential government services. Legal services no longer appears on the list. The deployer obligations shift from governance and auditing to consumer-facing notice: pre-decision disclosure that an automated system is involved, post-adverse-outcome explanations within thirty days, correction rights, and an opportunity for meaningful human review. The bias audits and impact assessments are gone. A sixty-day right-to-cure provision allows deployers to remedy violations before the Attorney General can act.
The net result for law firms operating in Colorado is that the statutory threat I described last in my last post—the prospect of being regulated as a deployer of high-risk AI in a domain where the AI’s contribution to consequential decisions was explicit in the statute’s text—has receded. Lawyers are no longer in a specifically enumerated category under the new law.
The exemption that does the work
The more consequential feature of SB 26-189 for legal practice is not the removal of “legal services” from the covered categories. It is the breadth of the exemption for tools used solely to “summarize, organize, translate, draft, route, or present information for human review of administrative processing.” The statute also excludes chatbots and conversational tools that are not “contracted, advertised, marketed, configured, or intended to be used in a consequential decision” and that are subject to an acceptable use policy prohibiting their outputs from being used in such decisions.
Read those exclusions against the backdrop of how lawyers currently use AI, and the exemption appears to cover most of it. A lawyer who prompts an LLM to summarize a deposition transcript is using a tool to “summarize information for human review.” A lawyer who uses an AI drafting assistant to produce a first draft of a motion is using a tool to “draft information for human review.” A lawyer who uses an AI research tool to locate and organize relevant authorities is using a tool to “organize and present information for human review.” Under SB 26-189, these use cases fall outside the statute’s scope regardless of whether the underlying decision—a case evaluation, a settlement recommendation, a litigation strategy—is consequential.
The exemption’s logic rests on a premise that will be familiar from the early debates about AI in professional settings: that a tool which presents output for human review is not making the decision; the human who reviews the output is. On that theory, the AI is infrastructure. The human is the decision-maker. And because the human retains control, the regulatory obligations attach to the human’s conduct—through professional-responsibility rules, malpractice liability, and the other mechanisms that have always governed professional judgment—rather than to the tool itself.
That theory was plausible when AI tools operated as prompt-and-response systems: the lawyer asks a question, the tool produces an answer, and the lawyer evaluates the answer against her own knowledge, judgment, and understanding of the client’s situation. It is less plausible when the tool is an autonomous agent that decomposes a goal into subtasks, selects among research strategies, decides which authorities to retrieve and which to discard, determines how to structure its analysis, and presents a finished work product—all before the lawyer sees anything at all.
The exemption was drafted for a different tool
Harvey’s Agent Builder, Thomson Reuters’ CoCounsel, and Legora’s aOS each produce output a lawyer will review. Each vendor describes its product with language emphasizing human oversight—“human-in-the-loop checkpoints,” “review-ready work product,” “fiduciary-grade AI.” Under SB 26-189’s exemption, each could plausibly be characterized as a tool that “summarizes, organizes, drafts, or presents information for human review.”
That characterization would be accurate as a description of the output’s format. It would be misleading as a description of the work the system performed to produce it. When an agentic due diligence system reviews a data room, it does not retrieve documents and present them for the lawyer’s consideration. It decides which documents to examine first. It identifies provisions it classifies as risk factors, applying criteria the lawyer did not specify and often cannot inspect. It selects which risks to foreground and which to treat as routine. By the time the lawyer opens the report, the analytical choices that will shape her review have already been made—by the system, not by her. The statutory framework treats “summarize, organize, draft, and present” as neutral, ministerial functions. The tools performing those functions in 2026 embed evaluative judgment at every step of the process.
I examined this supervision gap in detail in a recent post on agentic tools—the structural difficulty of overseeing a system that makes intermediate decisions the reviewer cannot observe or reconstruct. The Colorado exemption raises a related but distinct problem: the statute’s drafters built a regulatory boundary around the task/judgment distinction without acknowledging that the current generation of legal AI tools has blurred that boundary beyond easy recognition. The exemption assumes that “summarizing information for human review” is a describable, bounded activity. For the tools law firms are buying, it is not.
The human review assumption
The exemption’s reliance on “human review” as the regulatory boundary reflects a broader assumption in AI governance: that a human who reviews an AI system’s output exercises independent judgment over the decision the output informs. Colorado is not alone in this assumption. The EU AI Act’s framework for high-risk systems requires “human oversight” as a safeguard. California’s proposed Rule 1.1 amendments require lawyers to “independently review, verify, and exercise professional judgment” over AI outputs. The assumption is pervasive.
But the Colorado exemption treats “human review” as a binary: if a human reviews the output before acting on it, the tool falls outside the statute’s scope. A lawyer who reads an AI-generated contract review catches the mischaracterized clause—but she cannot catch what the system chose not to include. If the agent scanned forty contracts in a data room and flagged twelve as containing material risks, the reviewing lawyer does not know what criteria the agent applied to the other twenty-eight. A three-page summary of a five-hundred-page data room invites judgment over the summary’s conclusions, not over the choices that determined which five hundred pages produced which three pages. The exemption assumes that review of the output is sufficient oversight of the process. Whether that is true depends on what the tool did upstream—and the exemption does not ask. The strongest version of the opposing view is that a competent, engaged lawyer can interrogate the agent’s path, demand criteria for what was flagged, and examine what was omitted; on that account, what the statute requires is the lawyer’s diligence, not the tool’s transparency. The economic logic of agentic tools cuts the other way: a lawyer who reconstructs each upstream choice captures none of the efficiency the tool was bought to provide, and the exemption draws no distinction between the lawyer who interrogates the process and the lawyer who reviews the output and signs off.
What happens next
Colorado’s legislative retreat from the original AI Act’s deployer obligations will not go unnoticed by other states—in part because it is not the first. Connecticut’s SB 2, introduced in 2025, tracked Colorado’s structure closely: broad deployer obligations for high-risk AI systems used in consequential decisions across multiple categories, including education, employment, financial services, and healthcare. Governor Lamont threatened to veto it, and the bill died without a House vote. Connecticut returned in 2026 with SB 5, a narrower omnibus that passed both chambers with large bipartisan margins and that Lamont has indicated he will sign. But SB 5 is not the Colorado model. It regulates employment-decision AI, chatbot safety, social media protections for minors, and synthetic-content provenance. It does not impose broad deployer obligations. It does not cover legal services. Two states have now attempted comprehensive audit-and-governance AI frameworks and retreated from them—Colorado under a federal lawsuit and DOJ intervention, Connecticut under a gubernatorial veto threat—and both replaced the broad framework with something considerably narrower. States that had been watching either jurisdiction before committing to their own frameworks received a clear signal about the political viability of the approach.
For lawyers, the practical takeaway from SB 26-189 is counterintuitive. The removal of “legal services” from the covered categories and the breadth of the human-review exemption mean that most AI use in legal practice now falls outside Colorado’s statutory framework. Even AI applications that fall within the new framework face limited near-term enforcement exposure: Colorado’s Attorney General has publicly stated that his office will not enforce SB 24-205 or its replacement until rulemaking under SB 26-189 concludes. That should not be read as a green light. The professional-responsibility obligations that predate SB 26-189—the duty of competence under Rule 1.1, the duty of supervision under Rules 5.1 and 5.3, the communication obligations under Rule 1.4, the requirements of ABA Formal Opinion 512—remain in full force. They apply regardless of whether a statute independently regulates the same conduct. And they are, in certain respects, more demanding than what the statute would have required: the duty of competence requires the lawyer to understand the technology before relying on it, not merely to disclose that she used it. The duty of supervision requires the lawyer to maintain reasonable oversight of the AI’s contribution to the work product, not merely to review the final output.
The Colorado AI Act was, for all its flaws, an attempt to build a statutory framework that acknowledged AI’s role in consequential decisions and imposed structured obligations on the entities using it. Its replacement exempts the fastest-growing category of AI deployment in professional settings—agentic tools that plan, select, and structure before the human ever intervenes—by treating their output as a summary prepared for review. The exemption was drafted for a model of AI use in which the lawyer asks a question and evaluates the answer. The tools law firms are adopting in 2026 decompose the goal, select the strategy, execute the workflow, and present the finished product—all before the lawyer sees anything at all. The statute carved out a safe harbor for a tool that no longer describes what firms are buying.
This post draws on the text of Colorado SB 26-189 (passed May 9, 2026; signed by Governor Polis on May 14, 2026); the Colorado Sun’s reporting on the bill’s passage; EPIC’s analysis of the SB 26-189 amendments; The Sum and Substance on the signing; the DOJ’s intervention in the xAI lawsuit; Connecticut SB 5 and the Transparency Coalition’s bill guide for the Connecticut Artificial Intelligence Responsibility and Transparency Act; and the Future of Privacy Forum’s analysis of Connecticut SB 2. It extends the analysis of deployer obligations from a prior post on the original Colorado AI Act and the supervisory analysis from The Agent Is Not Exercising Judgment.