AI system documentation

Intended purpose#

MeritSlate Recruit is a screening assistant for staffing and recruiting agencies. It is intended to help a recruiter get through the first stage of a search faster: reading a batch of résumés against a job description, conducting a structured first-round interview by voice, and summarising both so the recruiter can decide whom to speak to next. It is intended for early-stage screening only, and only where a human reviews the output before any decision that affects a candidate.

What the system consists of#

• Résumé ranking. A language model reads each résumé and the job description, extracts the skills and experience the job asks for, and ranks candidates by how well the résumé evidences them. Every score comes with the quoted evidence it rests on.
• AI interviewer. A live voice interview, conducted by a language model speaking through a speech service, following a question plan derived from the job. The candidate is told before it starts that the interviewer is an AI, what is recorded, and for how long it is kept.
• Scorecard. A structured summary of the interview against the job's criteria, with the transcript passages it draws on.
• Integrity signals. Neutral observations made in the candidate's own browser during the interview — for example that the interview tab lost focus, or that the camera view changed for a sustained period. They are logged as events and shown as observations. They are not, and must not be presented as, a determination that anything improper occurred.

Data the system uses#

• The job description and screening criteria the agency provides.
• The résumés the agency uploads or imports.
• The candidate's interview: audio, the transcript derived from it, and, where the agency has enabled it and the candidate has consented, the video recording.
• Browser events during the interview (focus, visibility, camera state) used only to produce integrity signals. Analysis of those events happens in the candidate's browser; video is not streamed anywhere for analysis.

Recordings, transcripts and quoted evidence are retained for the window shown to the candidate at consent time and then deleted; the derived judgement (scores and summaries) is retained for the agency's records. Agencies on a California compliance profile keep the underlying records for the longer window that profile requires. The privacy policy sets out the details.

Data the system does not use#

The system does not ask for, infer, or score on protected characteristics — race, colour, religion, sex, national origin, age, disability, genetic information, or any equivalent under local law. It holds no such field. It does not use social media, credit data, criminal history, facial-recognition identity matching, or emotion recognition. The models are not trained on agency or candidate data; that data is used to produce the agency's own outputs and for nothing else.

Outputs and how to read them#

• A ranking score reflects how well a résumé evidences the stated criteria. It is not a prediction of job performance and should be read alongside the quoted evidence.
• A scorecard reflects what the candidate said in the interview against the criteria, in the interviewer's summary. It should be read with the transcript it cites.
• An integrity signal is an observation with an innocent explanation as likely as any other — a phone rang, a child came in, a laptop lid moved. It is information for a conversation, not a verdict.

Known limitations#

• Language models can misread a résumé — an unusual layout, a scanned image, a job title that means different things in different industries. Scores should be spot-checked against the document, especially near the cut-off the recruiter is using.
• The interviewer is English-only. Candidates with strong accents, speech differences, or poor connections may be transcribed less accurately, and a scorecard built on a poor transcript is a poor scorecard. The transcript is always available to check.
• Criteria are only as fair as the job description. If a job description encodes an irrelevant requirement, the ranking will reward it faithfully.
• Integrity signals depend on the candidate's device and surroundings. A shared computer, an older browser, or a permission prompt at the wrong moment can produce a signal with no meaning at all.
• The interview requires a computer with a modern browser and a stable connection. Candidates who cannot meet that should be offered another route, and the system provides no accommodation by itself.

Evaluation and adverse-impact monitoring#

Because the system holds no protected characteristics, it cannot compute selection-rate ratios by group itself, and we do not infer groups to do so. Instead it gives each agency the applicant-flow record needed to run that analysis with the agency's own self-identification data: for every job, an export of all applicants including those screened out, each with a candidate reference, the stage reached, and the scores at each stage. An adverse-impact report can be generated on demand from the workspace's compliance settings and each run is kept. We review ranking and interview behaviour against held-out job sets when models or prompts change, and we publish material changes here.

Human oversight#

• No candidate is advanced or removed by the system. Every stage change is a recruiter's action in the workspace.
• Every score and summary shows the evidence it rests on, so a recruiter can disagree with it in the same view.
• Candidates are told an AI conducts the interview, can decline, and can ask the agency for a human alternative. Candidates can also ask for their interview data to be erased.

What deployers must do#

• Have a person review AI output before any decision that affects a candidate.
• Give candidates the notices the deployer's jurisdiction requires about the use of an automated tool, and keep the records it requires. The workspace's compliance settings and exports exist for this.
• Write job criteria that are job-related, and revisit them when a ranking looks wrong.
• Offer an alternative to candidates who cannot take an AI voice interview.
• Never present an integrity signal to a candidate, a client or a colleague as evidence of wrongdoing.

Where it must not be used#

The system is not currently offered for positions based in New York City, or for candidates located in Illinois; the terms say why. We will update this page and the terms when that changes.

Changes and contact#

This page is updated when the models, the prompts, the data the system uses, or the jurisdictions it is offered in change. The date at the top is the date of the last change. Questions about the system go to hi@meritslate.ai.