Skip to main content
Back to newsletter

NEWSLETTER

Rejected for a Job? If AI Scored You, Colorado Says the Company Has to Explain

November 3, 2026

Cover illustration of a worried woman at a laptop showing a rejection email, beside the headline “What Did the Software Say About You?”

Colorado's SB 26-189, signed May 14, 2026, applies to decisions made on or after January 1, 2027, and requires a company that uses automated decision-making technology to materially influence a major decision, including hiring, to give notice before use and to explain the software's role within thirty days of an adverse decision about a Colorado resident. A settings change that turns on applicant scoring can move an already approved product inside the act, and an AI inventory built from one-time approvals may not notice. The edition sets out four questions, the Thirty-Day Test, on tool inventory, who writes the explanation, notice on the careers page, and record retention.

If you live in Colorado, starting January 1 a company has to tell you what role covered software played in rejecting you. It cannot explain a decision until it knows which software helped make it.

Picture this. It is January 2027, shortly after a new Colorado law takes effect. Everything in the next few minutes is invented: the people, the company, and the software.

Teresa Vance had been out of work since October. She was a scheduler in Denver with nine years of experience, and when a health system posted a remote scheduling job, she applied the same day.

The rejection came on a Wednesday morning, the day after the posting closed. No phone call. No interview. Just a short, polite email.

Hand-drawn illustration of Teresa at her kitchen table, holding a mug and reading a laptop. The screen shows an email headed “REJECTION”. A tag in the corner reads “FICTIONAL SCENARIO”.
Fictional scenario: Teresa gets a short, polite no the morning after the posting closes.

Later that night Teresa got one of those news stories, this one on a new law enacted in Colorado. It said that when an employer uses a computer program to assist in the hiring process and the decision is to deny employment, the program's role in the decision must be conveyed to the applicant. She read it twice. Then, she wrote to the company: "I am in Colorado, and I was rejected on Wednesday. Was this decision aided by a computer program? In particular, what did this program say about me?"

The company could not answer her. But to answer her, a company should have a clear idea of the hiring features it is running. This is where the company went wrong. To understand how, revisit that day they purchased the recruitment software.

The health system had about four thousand employees. Its headquarters were in Ohio but it operated in several other states and used video to provide treatment to patients as well as employing workers in other states remotely, including Colorado, where they treated patients and employed Coloradans but had no office. Everything was handled through one particular piece of software managed from the headquarters; when a candidate's resume was submitted it was stored on the system and recruiters on the system logged on to read what was coming in. The name for that kind of system is an applicant tracking system, but for this story it is just where all those applications go.

Three years earlier the company acquired it and it had been tested by security. The questions they asked were whether hackers could get in; who would be able to log in and where would they store the data. The software passed those checks. It was used to store and display resumes to recruiters. No applicant was scored so there were no questions about artificial intelligence. It passed and the approval ended up in a folder.

Sometime in September 2026 the vendor emailed its customers about what was being offered to them: the Top Match feature. This system read through all resumes and awarded scores to candidates ranging from 0 to 100. The software was not active until an administrator activated it. Priya, an HR administrator at the corporate headquarters and the person responsible for the software settings, had seen how every vacancy had left recruiters lagging. On the vendor's help page they described Top Match as assistive, but also that the system would not deny individuals on its own. One single line mentioned the places where businesses have to notify people of the AI systems. Priya left this line to the lawyers. The law was newly written in Colorado and its rules had not been finalized yet, and she worked in Ohio. One afternoon, the system allowed her to activate this feature for all vacancies. Nothing on the screen asked where people had applied from, though it did ask how long scores should be kept, and she chose ninety days. She remembered always being told to keep as little personal information as possible. No one had ever told her that there might be laws asking for records of what the tool did.

Hand-drawn illustration of Priya, an HR administrator, at her office desk with her hand on a mouse. Her monitor shows a switch set to “TOP MATCH: ON” and a menu reading “KEEP SCORES: 90 DAYS”. A sticky note on the monitor reads “APPROVED 3 YEARS AGO”. A tag reads “FICTIONAL SCENARIO”.
Fictional scenario: one setting, changed in an afternoon, with no one else told.

Nobody informed security about this, and no guideline required Priya to. The approval covered a product that scored no one. Security has kept a list of all fourteen AI tools used across the company. The recruiting tool had been approved for use as a resume repository. It never appeared on that list. And so nothing in the company's process ever told security that it had become something else. The gap between enabling an AI recruitment feature and asking what that change meant had no owner.

Hand-drawn illustration of a hand holding a clipboard titled “AI TOOLS WE TRACK: 14”, with colored dots beside each line. A folder labeled “HIRING SOFTWARE” sits to the side, tagged “NOT ON THE LIST”, with a dashed arrow and a question mark pointing from it to the clipboard. A tag reads “FICTIONAL SCENARIO”.
Fictional scenario: the security team’s list of fourteen AI tools does not include the hiring software.

In the second week of January, the opening closed on a Tuesday with six hundred resumes for the three posts. As the person who had to determine who would be invited for an interview and who would not, Gil, the recruiter, did not have time to go over six hundred resumes. A program automatically ranked them from highest down and grouped them according to their score, so he simply dialed the telephone numbers of the first forty individuals. The fortieth individual on the list had scored a 71. The remaining five hundred sixty people received a form email the following morning. All calls were made by one human, the software set the order, and Gil stopped with forty.

Beside her name, in a column visible only to the recruiters, her score was written down: 23 out of 100. Her rejection email never indicated that number; the score only showed where she landed, not why.

Hand-drawn illustration of Gil, a recruiter, on the phone while looking at a ranked applicant list on a large monitor. A red line labeled “CALLS STOP HERE” cuts across the list, with a note “40TH: 71”. A highlighted row near the bottom reads “TERESA V. 23”. A tag on the desk reads “560 FORM EMAILS”, and another reads “FICTIONAL SCENARIO”.
Fictional scenario: the software set the order, and the calls stopped at forty.

The following day, her email was passed on to Gil who passed it on to his manager, who in turn passed it on to the general counsel, the company's senior legal figure. The general counsel replied that the health system was subject to a federal health privacy law and that Colorado's law might have an exception for entities like this one, so a courteous acknowledgment would be sufficient for the time being.

However, the last person copied on the email was the CISO (Chief Information Security Officer) and she had a look at the list of AI tools and then decided to add a sentence: There is no hiring tool on it.

"The hiring system gives every applicant a number," replied Gil, writing to the whole thread. "I have been going down the list in order since September."

The answers arrived in the next few days.

Priya called up the vendor and the confirmation came back, a computer algorithm had read and scored every resume sent in. By law, the company should have informed people before using it. But their careers page did not mention anything, and that is because nobody thought the switch needed a notice.

The law required them to have the information about the tool that they were going to tell Teresa: name, version, vendor, data it processed. Priya knew when she had turned the tool on, but she did not know what version was running or what data it processed and there was no screen to show this information.

A separate, longer-term problem was that the law also mandated that the company keep the records showing their compliance for three years following the decision. The scoring system wiped scores off their database after ninety days, so Teresa's score of 23 would be erased in April, and with it the only copy anyone had found of what the tool had had to say about her.

And then Teresa was allowed to request the decision be re-evaluated by someone else, where this was reasonable. Gil had used the ranked list and he had never even read her resume because her score put her below the first forty names.

She was rejected on Wednesday: the company had thirty days from that Wednesday to explain the software's role. By the next Monday there were only five days out of the thirty gone and twenty-five remaining. It was not until then that they found out what there was to explain.

Hand-drawn illustration of a desk calendar headed “30 DAYS” with the first five days crossed out in red ink and day five circled. A tag on the calendar reads “25 LEFT”, and a hand holds a pen at its edge. A tag reads “FICTIONAL SCENARIO”.
Fictional scenario: by the time anyone looked, five of the thirty days were gone.

The software was approved as a filing cabinet for resumes. In September it was changed to set the order for which candidate a call was made to. No separate review was required before she enabled it.

If the scenario sounds too surreal, then the reality is that a version of the above is already in federal court. In Mobley v. Workday, a man over forty years old says he applied for over 100 jobs with companies that use Workday and was rejected every single time. He claims Workday's AI technology scored, sorted and prioritized people and that humans were only consulted later. In May 2025 a US District Court in the Northern District of California (Judge Rita Lin) allowed the case to move ahead as a collective action. In February 2026 it authorized notice to all people over forty who had applied via the platform since September 24, 2020, with a March 7, 2026 deadline to join the litigation. Workday has claimed they just supply the technology and people using it decide whether to hire, but at this stage the court has not decided whether the claims are true. This is the applicant's story: being rejected without an explanation, only to get a litigation notice, years later, about allegations that software screened applicants before human review. The notice still does not explain that person's rejection.

Hand-drawn illustration of a hallway table holding a large envelope stamped “COURT NOTICE” and a worn, yellowed printout reading “REJECTED” and “NO REASON GIVEN”. A tag between them reads “YEARS LATER”. No person is shown.
Illustration: years after a rejection, a court notice arrives about software that allegedly screened applicants. This is based on the claims in Mobley v. Workday, which the court has not decided.

This health system is hypothetical, but the rule is actual. The thirty-day duty applies only in Colorado, to benefit those who reside there.

What Colorado passed

On May 14, 2026, Governor Jared Polis signed Senate Bill 26-189, Chapter 131 of the 2026 Session Laws of Colorado. This bill repeals and replaces the state's 2024 AI law, which was due to be implemented on June 30, 2026. The principal provisions of this new act apply to decisions made on or after January 1, 2027, fifty-nine days from the date of this issue.

The act regulates a technology it refers to as automated decision-making technology, i.e. any technology that "processes personal data and uses computation to generate output." When that output is used to materially influence a major decision, the tool is covered. The duties fall on any "person doing business in Colorado" that uses one, and the people protected include employees and "a job applicant who is a Colorado resident." Ogletree Deakins describes the employment decisions it covers as hiring, compensation, promotion and similar decisions about current or prospective employees. The act names seven areas it treats as major decisions, among them employment, lending, insurance and health care. Whether a company with no Colorado office counts as "doing business in Colorado" is a question for counsel, because the definition says nothing more than those words.

What decides whether a tool counts is what its output does. The act excludes tools used only to summarize, organize or present information for a person to review. That exclusion stops applying when the system produces a score, ranking, recommendation, classification or prediction. Hiring software that only stores resumes produces no score. The same software, after someone switches on a feature that gives every applicant a number, may fall inside the act. A settings change can move a product across a legal line, and nothing in an approval that looked at the product once will notice. Greenhouse's support documentation describes the same pattern: its AI match scoring stays off until a site administrator enables it, says some jurisdictions may require disclosing AI tools, and calls the feature "assistive AI, not automated-decision-making." A vendor's label does not settle coverage under the act.

What you can ask for, if you live in Colorado

If you are the applicant, this is the part for you.

If the decision goes against an applicant who lives in Colorado, the company has thirty days after making it to provide three things: a plain-language description of the decision and the role the software played, instructions for asking for more information about the tool, and an explanation of the applicant's rights. The tool information includes its name and version, its developer, and the kinds of personal data it used, to the extent the company receives that from the developer. Before using such a tool at all, the company must give notice, and the act says a "prominent public notice that is reasonably accessible at points of consumer interaction" meets that duty. For a hiring tool, the careers page is the obvious place. You do not have to guess whether software was involved.

On request the applicant can ask to have factually incorrect personal data corrected, and can ask for "meaningful human review and reconsideration," in the act's words "to the extent commercially reasonable." The act does not require a company to correct an opinion, prediction or score, only factually incorrect personal data used in the decision. The company must keep the records needed to show it complied for not less than three years after the date of the decision. The act does not give you the job back, and it does not let you sue under it. Only the attorney general enforces it.

The general counsel in the story was half right. The act does say its main sections do not apply to organizations covered by the federal health privacy law, known as HIPAA. In the same sentence it adds one exception: "except for a consequential decision related to employment or an employment opportunity." A hospital's patient decisions may sit outside the act. The hospital's hiring does not.

What nobody can tell you yet

There are two dates that matter. The first is the beginning of duties, January 1, 2027. The second is the day on which companies can be subject to penalties. That, at least, remains uncertain. The attorney general drafts the rules that flesh out the act. Colorado's Department of Law issued proposed rules on August 11, 2026, published an interim draft on October 6 and held a formal hearing on October 26. The attorney general's office maintains that the act requires that final rules be adopted before January 1, 2027. On October 23, Jones Walker, a law firm that has been watching the process, stated that the term "materially influence" may be one of the most consequential open questions. Under the test in the act, for a tool's output to "materially influence" a decision is for the output to be a "non-de minimis factor" in the decision, and the act excludes "incidental, trivial, or clerical uses." Deciding who gets a phone call based on a score can hardly be deemed to be trivial. That reading is a judgment, and it is a matter of interpretation to be resolved by final rules.

There are also some enforcement terms of the act that need mention. Only the attorney general may enforce the act. A violation is a deceptive trade practice. No private right of action is created by the act. The recipient of a notice of violation has sixty days to cure but only so long as it is not knowing or repeated and that right expires on January 1, 2030.

Legal note on timing: a federal court order from April 27, 2026, entered in a lawsuit xAI filed against the attorney general, says the attorney general "shall not initiate enforcement" for violations that occur up to fourteen days after the judge rules on a preliminary injunction motion that xAI must file within twenty-eight days after the final rules are adopted. Whether the order shields every company or only the plaintiff is a question for counsel.

The responsibilities start from a date of January 1. There remains the question of when enforcement may start. This question is, to date, unanswered. But the work the act asks for does not get shorter while that date stays unknown.

The Thirty-Day Test

Four questions are based on the duties which exist in the act and decide whether the organization can respond to Teresa by day thirty.

One: Do you have a list of every tool which evaluates, scores, ranks, or recommends people for hiring, promotion, and pay, including functionality that someone turned on after the sale? Start with the configuration panels and release notes of all the hiring vendors and ask the administrators what they turned on in the last twelve months.

Two: If a Colorado resident were turned down tomorrow, who would write the explanation in plain language, who would sign it, and where would the writer get the name of the tool, its version, the names of its developer and the categories of data to which it applies? The act tells a company to pass on any information received from its developer, so it begins in the contract between the tool's developer and the company.

Three: Is there a notice where applicants apply and from what date? The act requires notice before the tool is used, so when it is put up matters.

Four: How long does the tool keep what it produced about each person, and could you show an attorney general what the tool did on the day of the decision, three years later? In the story the setting was ninety days.

Question three can be answered by Friday by looking at the careers page. Question one requires asking each administrator which features they turned on.

Count the AI-based products in the inventory maintained by your security team. Then ask HR to list all the tools that score a person, and if you find an item they have that you do not, find out why it is missing from your inventory. Begin with looking for the approval that looked at the product once.

Before a recruitment team starts relying on a feature that ranks applicants, someone will have to be accountable for reviewing the switch that turned it on. But if you cannot name that person, then the switch has no owner.

The health system, people, product features, scores and settings in the story are made-up for the purposes of illustrating what may happen. The law, dates, court order and Mobley case are derived from the sources below. This is general information and not legal advice.

Sources: Colorado General Assembly, Senate Bill 26-189, “Automated Decision-Making Technology,” Chapter 131, Session Laws of Colorado 2026, approved May 14, 2026; read October 31, 2026. Colorado Attorney General, AI rulemaking page, last modified October 27, 2026. Andrew R. Lee, Jason Loring, Christina Baptista and Graham H. Ryan, Jones Walker LLP, “The Colorado AI Law You Prepared For Is Gone, And The Rulemaking Comment Window Closes October 26,” October 23, 2026 (republished by Legal 500 on October 26, 2026). Order of April 27, 2026, X.AI LLC v. Weiser, No. 1:26-cv-01515 (D. Colo.), as quoted by AI Challenge Watch, read October 31, 2026. Holland & Knight, “Colorado Governor Signs SB 189, Significantly Amending the State’s AI Law,” May 18, 2026. Benjamin W. Perry and Lauren N. Watson, Ogletree Deakins, “Colorado’s New AI Act Targets Automated Decision-Making for Consequential Decisions,” May 26, 2026. Holland & Knight, “Federal Court Allows Collective Action Lawsuit Over Alleged AI Hiring Bias,” May 27, 2025. Wiggins Childs, “Federal Court Authorizes Notice in Lawsuit Challenging AI Hiring Software for Potential Age Discrimination,” February 17, 2026. Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal.). Greenhouse Support, “Talent Matching,” last updated October 5, 2026, read November 3, 2026.

Last week: Give Your AI Agent the Bank Login and the Loss Is Yours.