HireVue: Video Interviewing, Assessments, and AI Governance
On this page 10 sections
HireVue combines live and on-demand video interviewing with assessments and workflow tools. The important buying question is not whether the platform uses AI. It is which decision a feature informs, what evidence supports that use, and what recourse a candidate has when automation affects the process.
What HireVue currently documents
HireVue’s AI explainability and ethics update says the company had supported more than 40 million video interviews and 200 million assessments as of September 2024. Those are vendor-reported cumulative activity figures, not evidence that a specific employer will improve time to hire or quality of hire.
The same document says HireVue publishes model cards, tests adverse impact, and lets candidates opt out of AI-scored assessments. Its broader AI in hiring overview describes structured interviewing, skills assessments, scheduling, and conversational workflows. Product pages establish what the vendor offers; they do not independently validate every accuracy, fairness, or return-on-investment claim.
Video interviews and assessments solve different problems
On-demand interviews can standardize the questions candidates receive and reduce scheduling work. Live video supports remote interviews but does not itself standardize interviewer behavior. Assessments can add job-relevant evidence when their content and scoring are validated for the role.
These components should not be treated as one score. A defensible workflow records:
- whether a response is reviewed by a person or scored by a model;
- which traits or skills are being inferred;
- how the score is used in the hiring decision;
- whether the candidate can request another format;
- how long recordings and derived data are retained.
HireVue has previously explained why it built assessments in its assessment methodology overview. Employers still need evidence for their own jobs, applicant populations, and decision thresholds.
Fairness evidence and its limits
Fairness testing can identify group differences and prompt investigation. It does not prove that a tool is bias-free, lawful in every jurisdiction, or suitable for every role. Results depend on the outcome being predicted, the comparison groups, sample size, accommodations, and how the employer combines the output with other criteria.
The US Equal Employment Opportunity Commission warns that an employer may remain responsible when a software tool screens out applicants with disabilities. Its AI and ADA guidance recommends safeguards such as accessible procedures and a clear accommodation process. Employers should also check state and local rules that apply to recorded interviews and automated employment decisions.
A practical buyer review
Before deployment, ask HireVue and the internal hiring owner to document:
- the exact feature and model version used for each job family;
- the job analysis and validation evidence behind the score;
- adverse-impact results, sample limitations, and retest schedule;
- candidate notice, consent, accommodation, appeal, and deletion routes;
- data processors, retention periods, security controls, and audit logs;
- the human decision owner and conditions for overriding a recommendation.
A pilot should compare the new process with a defined baseline. Track completion, stage conversion, elapsed time, candidate complaints, accommodations, subgroup outcomes, and later job performance. Do not infer causation from a vendor case study without checking the employer’s design and confounding changes.
Map the product to distinct hiring decisions
The label “HireVue” can hide several materially different uses. An on-demand interview may simply collect recorded answers for a recruiter. A structured interview workflow may guide human scoring. An assessment may generate a score that influences progression. Scheduling automation may have no selection function at all. Governance should be based on the configured use, not the vendor name.
| Workflow | Evidence produced | Main benefit to test | Main control to verify |
|---|---|---|---|
| On-demand video | recorded answers to fixed questions | scheduling reach and question consistency | accessibility, reviewer rubric, retention |
| Live interview | conversation and interviewer ratings | remote access and coordinated review | interviewer training, recording notice |
| Game or skills assessment | scored responses | incremental job-relevant signal | validation, subgroup results, accommodation |
| Scheduling assistant | availability and communication records | less administrative delay | escalation path, time-zone and contact accuracy |
This map also clarifies what a candidate should be told. Notice should identify whether recording occurs, whether a model scores any response, what part of the process the output affects, and how to request another format. A generic privacy-policy link is not a substitute for a stage-specific explanation.
Validation evidence buyers need
Employers do not need to accept either a marketing claim or a single global coefficient. Ask for the job analysis, the construct or behavior being measured, the criterion used as the outcome, the study population, sample size, time period, and evidence that the score is reliable. Then ask whether the intended job and applicant pool are similar enough for that evidence to transfer.
The federal agencies’ Uniform Guidelines questions and answers describe validation as demonstrating job relatedness and distinguish criterion-related, content, and construct approaches. They also say a procedure developed elsewhere can be used, while the employer remains responsible for showing that its use for the particular job is consistent with the Guidelines. That is why a vendor study is useful input but not a blanket approval.
For structured interviews, content evidence may come from asking the same job-related questions and applying an anchored rubric. For a scored assessment, the evidentiary chain is longer: the score must measure what it claims, relate to relevant work, and add useful information at the threshold actually used. Revalidate or review when the job, candidate population, question set, model, or decision rule changes.
AI governance is an operating practice
Model cards and adverse-impact testing are useful disclosures only if the customer can connect them to the deployed feature and version. Keep a configuration record showing whether AI scoring is enabled, who may change it, which jobs use it, and when the underlying model or questions were updated. An audit log should make it possible to reconstruct why a candidate was advanced or held.
The voluntary NIST AI Risk Management Framework offers a practical cycle: govern ownership, map the context, measure performance and risk, and manage the resulting controls. Applied to HireVue, that means testing the complete employer workflow rather than treating a vendor assurance as a certification. Monitoring should include errors, accessibility incidents, overrides, candidate challenges, and drift in selection patterns.
Human review must be substantive. A reviewer who sees only a recommendation and routinely accepts it is not an independent control. Give reviewers the underlying job-relevant evidence, train them on when an override is appropriate, and monitor whether overrides themselves produce inconsistent outcomes.
Data, privacy, and accessibility questions
Video and assessment workflows can involve recordings, transcripts, device data, responses, derived scores, and interviewer notes. The contract and implementation design should identify each category, the controller and processors, storage locations, retention schedule, deletion behavior, and downstream exports. Verify what remains in backups and analytics after a candidate requests deletion, subject to applicable retention duties.
Accessibility should be tested with the actual candidate journey. Confirm keyboard operation, captions or transcript options, assistive-technology compatibility, mobile behavior, extra-time settings, and a staffed route to request an alternative. The EEOC guidance is especially relevant because a feature that appears neutral can screen out a person whose disability affects interaction with the test rather than ability to do the job.
A decision-ready pilot
Choose a small set of recurring roles with enough volume to observe the funnel. Freeze the interview questions, assessment configuration, pass thresholds, and downstream process for the pilot window. Record the old baseline and predefine success: perhaps less scheduling delay with no reduction in completion or subgroup progression. Do not change employer brand campaigns and compensation at the same time if the goal is to attribute an effect.
At review, separate operational and selection outcomes. Scheduling time can improve even if an assessment adds no incremental evidence. Candidate completion can rise while later interviewer agreement falls. Make the production decision feature by feature, and document which results are observed, which are vendor reported, and which remain too uncertain to claim.
Bottom line
HireVue offers a mature set of interview and assessment workflows, and it publishes more AI-governance material than many recruiting vendors. That is a useful starting point, not a substitute for employer validation. The strongest implementation uses structured questions, role-specific evidence, transparent candidate choices, and documented human accountability.