AI can support candidate communication and collect job-relevant information, but it should not decide who advances. A defensible design keeps recruiters accountable, validates every criterion, accommodates candidates, and provides an audit and appeal path.
The right role for AI in candidate screening
AI can help a recruiting team answer process questions, collect a defined set of job-related responses, schedule interviews, and route exceptions. It should not make or effectively make a consequential employment decision. A recruiter should review the underlying information, apply documented criteria, and remain accountable for every advance or rejection.
That boundary matters because a consistent script is not proof of a fair or valid process. A system can repeat the same question for everyone and still exclude qualified people, rely on an irrelevant proxy, misunderstand an answer, or create an inaccessible experience. The EEOC's guidance on employment tests and selection procedures says selection procedures can violate federal law when they disproportionately exclude a protected group and are not job-related and consistent with business necessity. It also makes clear that buying a vendor tool does not transfer the employer's responsibility for how the tool is used.
The practical design principle is simple: use AI as a controlled interaction layer, not as a hiring authority.
Start with tasks that do not determine employment outcomes
The safest pilot separates candidate service from candidate evaluation.
| Task | Appropriate AI role | Human control |
|---|---|---|
| Explain the hiring process | Answer from an approved, versioned knowledge base | Recruiter owns the policy and handles exceptions |
| Schedule or reschedule an interview | Offer available times from an authorized calendar tool | Candidate can reach a person or request another channel |
| Collect minimum qualifications | Ask only approved, job-relevant questions and record the candidate's answer | Recruiter verifies the response and decides what it means |
| Recover an incomplete application | Identify missing fields without inferring their contents | Candidate supplies or corrects the information |
| Route accommodation requests | Provide a private path to the designated team | A trained person handles the interactive accommodation process |
| Communicate status | Read a status already approved in the applicant tracking system (ATS) | The ATS and recruiter remain the source of truth |
Do not ask a language model to rank applicants, infer personality or emotion, score a person's voice or facial expression, predict “culture fit,” or issue a rejection. Those uses turn uncertain model output into a consequential decision and can introduce criteria that were never validated for the job.
Six controls to put in place before the first candidate interaction
1. Keep a person responsible for each decision
“Human in the loop” should mean more than giving someone a button to accept the model's recommendation. The reviewer needs the candidate's actual response, the applicable job criterion, and the authority to disagree. The system should never auto-advance or auto-reject a candidate because of a generated summary, score, missing transcript, accent, speaking style, or detected sentiment.
Write the decision boundary into the workflow. Define which actions the agent may take, which require approval, and which it may never initiate. If an answer is ambiguous or a tool fails, route the record for review rather than treating uncertainty as a negative signal.
2. Validate criteria for the specific job and use
Every screening question should map to a documented requirement for the position. Test whether the procedure measures that requirement and whether a less discriminatory alternative can achieve the same purpose. Revalidate when the job, question, model, prompt, scoring logic, or applicant population changes.
The EEOC advises employers to ensure that selection procedures are properly validated for the positions and purposes for which they are used. Its guidance also notes that a vendor's documentation may be useful, but the employer remains responsible for the procedure. “The model understood the conversation” is not a validation study.
3. Test for discrimination; do not promise that AI removes bias
Standardization can make a process easier to inspect. It does not guarantee nondiscrimination. Before launch and at defined intervals, examine selection outcomes and failure patterns across legally relevant groups with counsel and qualified industrial-organizational or validation expertise. Investigate where transcription, language, access, question design, or reviewer behavior creates unequal effects.
In New York City, a tool within the city's definition of an automated employment decision tool may not be used unless it has a recent bias audit, a public summary of the results, and required notices. The NYC Department of Consumer and Worker Protection's Local Law 144 page links the law, implementing rule, and current FAQ. The scope is definition-dependent, so do not treat a general compliance checklist as a legal determination.
4. Make the process accessible and provide accommodations
A voice or video interface can screen out a qualified applicant whose disability affects speech, hearing, vision, manual input, attention, or interaction with the software. The Department of Justice's guidance on AI and disability discrimination in hiring says hiring technology must not unlawfully screen out qualified people with disabilities and explains that employers may need to provide reasonable accommodations.
Tell candidates what technology they will use and how the interaction will be evaluated. Offer an accessible alternative channel before the interaction begins. Provide a clear accommodation request path, keep the request from lowering the person's prospects, and do not make the conversational agent decide whether an accommodation is warranted. Test the experience with assistive technologies and people with different access needs.
5. Preserve an audit trail and a way to challenge errors
For each interaction, record the approved script version, model and configuration version, tool calls, source data returned, transcript or structured responses, human reviewer, decision rationale, overrides, and corrections. Apply a documented retention and access policy rather than keeping every artifact by default.
Tell candidates how to correct a transcription or data error, request another channel, and ask for human review. An appeal path is operationally useful even where a particular law does not prescribe one: it reveals systematic failures and prevents an uncorrected model error from becoming the final record.
The NIST AI Risk Management Framework is voluntary, not an employment-law safe harbor. Its Govern, Map, Measure, and Manage functions are nevertheless a useful way to assign ownership, document the context of use, test risks, and respond to findings.
6. Minimize data and lock down integrations
Collect only information needed for the approved task. Do not invite medical information, infer protected characteristics, or enrich an application with unrelated personal data. Before recording a call or retaining a transcript, determine which notices and consents apply in the relevant jurisdictions.
Give the agent the least access it needs. A scheduling agent may need free/busy calendar slots but not interview feedback. A status agent may need a read-only application state but not authority to change it. Separate test and production credentials, restrict available tools, log writes, and make repeated requests idempotent so a retry cannot duplicate an action.
A reference workflow for a human-governed screening call
- The ATS starts the interaction. It passes a minimum identifier and the approved workflow version, not the entire applicant record.
- The agent explains its role. It states that it is an AI system, describes the purpose of the interaction, and offers a human or accessible alternative.
- The candidate chooses how to continue. Accommodation requests and channel changes route to a trained person without penalty.
- The agent asks approved questions. It collects verbatim or structured responses and may clarify format, but it does not score, infer, or recommend.
- Tool access stays narrow. Calendar and ATS calls are authenticated, allowlisted, and limited to the fields and actions required for the pilot.
- A recruiter reviews the record. The reviewer sees the original response, resolves ambiguity, applies validated job criteria, and makes the decision.
- The candidate can correct the record. A correction or appeal creates a new review event; it does not silently overwrite the audit trail.
- The team audits the system. It reviews completion, failures, accommodations, overrides, complaints, and outcome differences, then pauses or changes the workflow when a threshold is breached.
Where Dasha fits—and where it does not
We can provide the real-time voice interaction layer for an inbound or outbound candidate call. Our documented Model Context Protocol connections can expose allowlisted external tools, while webhooks can send call lifecycle events, transcripts, failures, and transfer requests to your systems. Call history and inspection can support testing and review.
Those capabilities do not make an employment workflow valid or compliant on their own. Your team must supply the job analysis, approved questions, ATS and calendar integrations, access controls, notices, consent and retention rules, accommodation process, validation evidence, human review, and appeal route. If a connector cannot return authoritative data or enforce the required permission, the agent should not guess or take the action.
For a first Dasha pilot, limit the agent to candidate FAQs and scheduling, or to collecting a small set of clearly job-related responses for later human review. Do not enable ranking, recommendation, advancement, rejection, or autonomous status changes.
How to evaluate the pilot
Measure the system as a controlled process, not as a promise that AI is “accurate” or “unbiased.” Useful pilot measures include:
- task completion and human-transfer rates, broken down by workflow step;
- transcription corrections and ambiguous-answer reviews;
- tool errors, duplicate actions, and attempts to use an unauthorized tool;
- accommodation requests, alternative-channel completion, and accessibility defects;
- recruiter overrides and the reasons for them;
- candidate complaints, corrections, and appeals;
- selection and failure patterns across legally relevant groups, assessed with appropriate expertise; and
- time to detect, contain, and remediate a material issue.
Set pause conditions before launch. A serious accessibility failure, unexplained outcome disparity, unauthorized write, missing audit record, or inability to reach a person should stop the affected workflow until the cause is understood.
Frequently asked questions
Can AI make the final candidate-screening decision?
It should not. Keep consequential employment decisions with an accountable recruiter who can examine the original information, apply validated criteria, and correct errors. Do not make a generated score or recommendation the default that a reviewer merely rubber-stamps.
Does asking every candidate the same questions eliminate bias?
No. A uniform question can still be irrelevant to the job, inaccessible, misunderstood, or disproportionately exclusionary. Standardization can help an audit, but nondiscrimination requires job-related design, validation, outcome testing, accommodations, and human review.
Should a screening agent analyze a candidate's voice or emotion?
No. Accent, prosody, pace, facial movement, and similar signals can reflect disability, language background, culture, equipment, or environment rather than ability to do the job. A safer voice workflow records the candidate's answer to an approved question and routes that answer to a person.
Is compliance with NYC Local Law 144 enough everywhere?
No. The law has a specific jurisdiction and definition, while federal, state, local, privacy, recording, and accessibility requirements may also apply. Determine the rules for each role, location, employer, applicant, and technology with qualified counsel.
Build the service layer, not an automated hiring authority
AI can make candidate communication easier to access and operate when the scope is narrow and the controls are real. Start with FAQs, scheduling, or structured information collection. Keep evaluation criteria job-related, give candidates an accommodation and correction path, audit the workflow, and require a person to make every consequential decision.
If that is the architecture you want to test, evaluate Dasha's voice AI backend as the conversation layer and design the human, legal, and integration controls around it before any candidate enters the workflow.