Should employers use AI for first-round interviews, and how can they do it without harming candidates or creating legal risk?

Start with the least consequential capability

“AI first interview” can mean several very different systems. Conflating them hides the real decision. A calendar assistant that offers times is not doing the same job as a model that marks a candidate as unsuitable after analyzing facial expression, speech cadence, or an answer to a recorded prompt.

The practical rule is simple: use automation to reduce coordination work before using it to judge people. For an early experiment, an employer can let a candidate choose a time, answer a few logistical questions, and receive a plain-language explanation of the role. A trained interviewer should still ask the same job-related questions, take responsibility for the evaluation, and make the advancement decision.

This is not merely a matter of tone. An applicant can be disadvantaged by a poor transcription, unreliable internet, an assistive technology conflict, a language difference, a disability-related communication pattern, or a tool that treats an arbitrary proxy as evidence of competence. The more the system influences the disposition, the stronger the employer’s evidence and safeguards must be.

What Colorado and federal law mean in practice

This is general information, not legal advice. Hiring counsel should review the job, locations, vendor contract, applicant flow, data handling, and applicable state and local rules before launch.

For Colorado, the current statute is SB26-189, enacted in May 2026. It repealed and reenacted the earlier Colorado AI Act framework. It covers an automated decision-making technology that makes, guides, or assists a consequential decision, including an employment decision. Key deployer and developer obligations, including employment-related notice, information, correction, and meaningful human-review provisions, are scheduled to begin on January 1, 2027. The bill also requires specified records to be kept for at least three years. A company should build these protections before that date rather than trying to retrofit them after a complaint or failed rollout.

At the point where a candidate interacts with a material hiring tool, provide a clear explanation that the technology is used, what it does in the process, what information it uses, and how to ask a person for help. If an adverse decision is materially assisted by the tool, prepare to give the individual a plain-language description of the tool’s role, a route to correct factual inaccuracies, and a meaningful human reconsideration path. A person who can only repeat a vendor score is not meaningful review.

Federal civil-rights obligations do not wait for Colorado’s 2027 implementation date. Title VII can reach selection procedures with unlawful disparate impact, and the Uniform Guidelines on Employee Selection Procedures explain the longstanding expectation that an employer maintain evidence for a selection procedure that causes adverse impact. The EEOC’s Title VII guidance and its AI and ADA resource hub make the central point: the employer remains responsible when it uses a vendor’s tool.

Under the ADA, a qualified applicant may need a reasonable accommodation to have an equal opportunity in a selection process. A recorded video interview, voice analysis, timed chatbot, or inaccessible assessment can create a barrier. The EEOC gives the concrete example of an applicant with a visual disability being disadvantaged by a tool that requires visual interaction or does not provide an accessible alternative. See Visual Disabilities in the Workplace and the ADA. Put an accommodation request link and a real contact in every invitation, train recruiters to act quickly, and do not require the applicant to disclose a diagnosis to obtain a conventional interview format.

Candidate protections that should be designed in

Notice is necessary, but it is not the same as consent and it does not cure an unfair selection method. Make the candidate’s choice real whenever recorded video or optional AI assistance is proposed. The invitation should say whether the system schedules, transcribes, summarizes, scores, ranks, or recommends, whether audio or video is retained, who can see it, and how long it is retained. It should also link to an equivalent alternative that does not penalize the applicant.

Colorado’s Privacy Act identifies biometric data used to identify an individual as sensitive data. Whether and how the Act applies to a particular employer and applicant record needs legal analysis, but it is a strong reason not to treat a face or voice recording as ordinary interview paperwork. See the Colorado Attorney General’s Colorado Privacy Act resource.

Ban emotion and personality inference

Adopt a written prohibition on using AI to infer emotion, honesty, enthusiasm, personality, culture fit, mental state, or employability from facial expression, voice, typing rhythm, eye contact, accent, or similar signals. These are not reliable substitutes for demonstrated job-related capabilities, and they can convert disability, language, culture, or presentation differences into a hidden ranking signal.

The prohibition should cover both deliberate product features and hidden vendor settings. Ask the vendor to identify every input feature, derived feature, model output, and default configuration. Contractually prohibit biometric identification, face analysis, voiceprint creation, affect analysis, and personality scoring unless the employer has separately approved a lawful, necessary, job-related use after specialist review. For ordinary first-round hiring, the safer answer is to keep all of them off.

Prove the process is job-related before it filters applicants

A hiring process starts with a job analysis, not a model demonstration. Define the essential requirements of this role, the observable evidence of each requirement, and the structured questions that elicit that evidence. For example, a customer-support role may require explaining a process clearly, documenting a case accurately, and handling a routine upset customer. It does not require a particular facial expression, home background, camera quality, or a vendor’s opaque “communication potential” score.

For each proposed criterion, record why it matters to the job, how the interviewer will recognize evidence of it, what answer range is acceptable, and what would be disqualifying. Have the same rubric applied to the structured human baseline and to any pilot. A qualified industrial-organizational psychologist or selection-validation expert can help determine whether the tool measures a job-related construct and whether the evidence supports the intended use.

Do not wait for a final adverse decision to look for disparate impact. Before launch, test the proposed criteria and candidate journey with accessibility reviewers. During the pilot, compare completion, advancement, and rejection patterns across available demographic groups and protected characteristics where collection, use, and sample size permit lawful, statistically responsible analysis. Investigate a material gap, technical failure pattern, or unexplained score difference before expanding the tool. A vendor’s aggregate benchmark is not validation for this employer’s role, applicant population, or workflow.

A pilot that can answer the right question

Run the pilot on one job family, one location or clearly defined hiring population, and one limited recruiting period. Do not begin with a role where a false rejection has unusually high consequences, such as a scarce specialist role, a public-facing position with an urgent fill deadline, or a job that depends heavily on accessibility accommodations.

Use a structured human first screen as the comparison baseline. The pilot should keep the same job description, eligibility criteria, structured questions, recruiter training, and advancement rubric. If a candidate enters an AI-assisted path, give them an equally prompt human or accessible route. In an initial safety pilot, the AI can schedule, provide a transcript, or summarize a candidate’s answers for the interviewer, but cannot reject, rank, or silently block a candidate.

Define success and stop conditions before the first invitation. The following measures make the tradeoff visible:

Small samples can be noisy, so do not call a pilot fair because a few comparisons look similar. The decision should consider the whole record, especially an identifiable false rejection, a serious accessibility problem, or a candidate experience result that shows people felt misled or unable to participate. Pause rather than tune a model on candidate data when the issue is unclear.

Example

Hypothetical example: an employer hires entry-level support specialists. The job analysis identifies three first-round competencies: explaining a billing rule in plain language, recording a customer interaction accurately, and deciding when to escalate a problem. It writes six structured questions and a behavioral rubric before it purchases any AI interview product.

The candidate invitation offers a choice of a 20-minute human phone or video screen, or a written asynchronous version with the same questions and the same evaluation rubric. Scheduling automation offers available times, and a transcription service makes notes from live interviews that the recruiter can correct. The recruiter sees the original answer, not an emotion score or a single automated recommendation, and makes the advancement decision against the rubric.

The takeaway is that the employer can remove back-and-forth scheduling and note-taking while preserving a fair chance to demonstrate the actual work. If it later considers AI summaries, it compares those summaries with the candidate record and the human screen. It does not turn a convenience tool into a hidden gatekeeper.

Vendor controls and audit rights

Procurement should be able to answer a skeptical candidate’s basic questions without calling sales support. Require the vendor to provide the intended use, known limitations, input data, whether customer data is used to train the product, security controls, subcontractors, retention settings, deletion process, and material model or feature changes. SB26-189 specifically establishes developer documentation and update duties that make this information increasingly important in Colorado. Colorado SB26-189

The contract should give the employer practical audit rights, not merely a marketing assurance. At minimum, seek the right to review validation materials relevant to the actual job use, obtain logs and exports needed for an individual reconsideration, receive prompt notice of material changes or incidents, assess accessibility, prohibit secondary use of applicant data, require secure deletion when the service ends, and terminate the risky feature without losing the applicant record. The employer should also be able to test the system with realistic edge cases before a candidate sees it.

Do not accept vague claims such as “bias-free,” “human in the loop,” or “compliant” as a substitute for evidence. Ask who the human is, what information they see, whether they can override, how often overrides occur, and whether the score has already determined the outcome in practice.

Common failure modes

• A score becomes a hard cutoff by default. Recruiters may believe they are reviewing a recommendation when the workflow has already filtered out everyone below a number. Remove automatic disposition during the pilot and audit the actual queue.

• The alternative path is slower or stigmatizing. A live or accessible route that takes weeks longer is not equivalent in practice. Track response time and advancement by path.

• A transcription error becomes a factual error. Let the candidate and reviewer flag errors, retain the original answer long enough for meaningful review, and correct the decision record.

• The vendor changes the product silently. Require change notice, version logging, re-testing, and approval before a new scoring model or feature is used.

• Video data expands beyond the stated purpose. Disable biometric and affect features, restrict access, separate raw media from hiring notes, and enforce deletion.

• The pilot chases speed alone. A reduction in recruiter minutes is not success if candidates abandon the process, qualified people are missed, or downstream performance falls.

Better alternatives for most employers

For many organizations, a structured human screen is the better investment. Give every interviewer the same job-related questions, behaviorally anchored scoring guidance, a short training session, and a documented decision record. This often improves consistency without creating a new opaque data-processing system.

A short recruiter screen can also be redesigned rather than replaced. Let candidates choose a live phone, accessible video, or written route. Send questions or evaluation themes in advance where appropriate. Use a calendar link for logistics and a human-reviewed transcript for notes. These changes address the common pain points of scheduling, uneven notes, and long backlogs while keeping the first meaningful interaction human.

Rule-based screening can be appropriate when it checks an objective, necessary condition, such as an active legally required credential, a candidate-confirmed work authorization question that counsel has approved, or a role’s genuinely fixed work hours. The rule should be visible, accurate, easy to correct, and reviewed when it excludes someone. If the rule cannot be explained plainly, it should not be an automatic gate.