Last updated: October 30, 2026
Cohera builds and operates an AI-assisted recruiting platform. Employers use it to interview and assess candidates; the AI's output informs the employer's hiring decision. This statement describes how that AI works in plain language. Candidates taking an interview should also read How the AI Interview Works; our Privacy Policy and Subprocessors list cover data handling.
1. What the system does
- Conducts interviews. An AI voice interviewer asks candidates questions about their background and the role, and, in hands-on rounds (coding, machine learning, system design), talks with the candidate while they work.
- Authors assessments. For hands-on rounds, an AI model drafts the problem and its answer key from the employer's role description; the employer can edit them before use.
- Evaluates rounds. After each round, an AI model reads what the candidate said and produced and scores it against the rubric the employer configured for the role. It returns per-competency scores, a written summary, strengths, concerns and suggested follow-up questions.
- Summarises across rounds so the hiring team can compare candidates on one page.
- Scores public professional profiles for role fit when an employer sources candidates. Nobody is contacted automatically unless the employer turns on Autopilot for a role, in which case the employer has instructed Cohera to email the highest-ranked candidates for that role on its behalf, up to a weekly number the employer sets, to answer their questions about the role from the employer's brief where the AI is confident (anything else waits for the employer), and to invite those who are interested to the first interview round; otherwise a recruiter decides whether to reach out.
The system is intended to support employment decisions. It is not designed for identity verification, proctoring, personality or emotion assessment, or lie detection.
2. What is analysed
- The transcript of the candidate's spoken answers (speech is transcribed to text).
- Code, notes, whiteboard content and design documents produced in an assessment workspace.
- Files or a public repository submitted for a take-home assignment.
- In AI-Assisted assessments, how the candidate used the AI coding tool: their prompts and a summary of its responses, the commands they ran, the changes they made, and the results of automated checks, some of them hidden from the candidate.
- The resume the candidate supplied when applying.
- The employer's rubric, must-haves and instructions for the role.
3. What is not analysed
- No facial analysis. The camera feed is recorded so the hiring team can watch the interview. It is never sent to an AI model; only the microphone audio is.
- No emotion, sentiment or personality inference. Scoring is about the content and structure of answers and work, never about how a candidate looks or sounds. The evaluator is instructed to ignore accent, dialect, second-language grammar and transcription errors.
- No voice biometrics. Speech is converted to text; the voice itself is not used to identify anyone or infer anything about them.
- No training on candidate data. Cohera does not train AI models, and candidate data is sent to our providers for processing only, under contracts that restrict its use.
4. Model providers
Cohera does not build its own AI models. It uses:
- OpenAI for the realtime voice interviewer (listening, turn-taking and what it says), speech-to-text transcription, and the evaluation of voice interview rounds.
- ElevenLabs for the interviewer's voice in some rounds: it turns the interviewer's written replies into speech. It receives only that text, never the candidate's audio or transcript.
- Anthropic for authoring and grading hands-on assessments (coding, machine learning, system design and take-home rounds).
- In AI-Assisted assessments, candidates work with Claude Code (Anthropic) or Codex (OpenAI). These are the candidate's tools, not Cohera's evaluator. Their traffic passes through Cohera, which records it for the hiring team and the evaluation and caps its cost.
Specific model versions change over time; each stored evaluation records the model and rubric version that produced it, so any result can be traced back. Recordings, transcripts and evaluations are stored in Google Cloud. The full list of providers and what each handles is on our Subprocessors page.
5. How scoring works
- Each competency in the employer's rubric is scored 0 to 10 against the bar for the role's level: 7 to 8 means the candidate meets the bar, 9 to 10 exceeds it, and lower scores describe how far below it the candidate fell.
- A competency the interview never reached is marked not assessed and left out; it does not count against the candidate.
- The overall score (0 to 100) is the weighted average of the assessed competencies, recomputed by our servers rather than taken from the model's arithmetic.
- The default passing score is 70, which is exactly what a candidate who meets the bar on every assessed competency scores. Employers can adjust the threshold for a round.
- Results are labelled Go, Maybe or No Go. A result within five points below the threshold, or one where too little of the rubric could be assessed, is a Maybe and goes to a human. A result that contradicts itself (for example, a "Go" with a clearly failing score) is also sent to a human instead of being acted on.
- Employers can set must-haves for the first screening round. These are checked at save time and cannot refer to protected characteristics or salary history; the AI interviewer is instructed never to ask about them.
6. Human oversight
- The hiring team reviews the recording, transcript, scorecard and summary before advancing, shortlisting, making an offer to, or rejecting a candidate. The AI evaluation is an input to that decision, not the decision itself.
- Borderline, low-evidence or contradictory results are always held for a person.
- Depending on the employer's settings, a clearly below-the-bar result may either be held for the hiring team to confirm or generate a rejection email automatically. The employer remains responsible for that outcome, and the candidate can request a human review at any time (section 8).
- If our systems fail during an evaluation, the candidate is never rejected because of it; the round is flagged for us to fix and re-run.
7. Known limitations
We want candidates and employers to know where the system can fall short.
- AI outputs vary. Two evaluations of the same interview can differ. We reduce this with anchored rubrics and server-side checks, but we cannot eliminate it.
- Transcription is imperfect, especially with strong accents, non-native speech, background noise, poor microphones or weak connections. A garbled transcript gives the evaluator less evidence to work with.
- No independent bias audit yet. We have not yet commissioned a study of outcomes by protected class, and our candidate volumes are still small. Until one exists, Cohera is not offered for positions in New York City.
- Rubrics are written by employers. Our checks catch explicit unlawful criteria, not every proxy for them.
- Silence handling. If a candidate goes quiet for an extended period and does not respond to check-ins, a live interview can be ended by the system. A microphone failure or a long pause can trigger this; whatever was captured is still evaluated, and a round with a recording but no usable transcript is sent to a human rather than rejected.
- Device requirements. Live rounds need a modern browser, a working microphone and camera, and a stable connection. Typed-answer and extra-time options are not yet available; see the accommodation route in How the AI Interview Works.
8. Requesting a review or correction
Candidates may ask for a human review of an AI-assisted evaluation, a correction of inaccurate information, a copy of the information used, or deletion of their recording and evaluation. Email privacy@coheraai.com from the address you interviewed with, naming the employer and the role, or use the "Need help" link on any interview page. Timelines and details are in How the AI Interview Works.
Employers with questions about their own obligations when using AI in hiring, or who need a copy of our detailed system documentation, can write to the same address.
9. Changes to this statement
We update this page when the way we use AI changes materially: a new kind of analysis, a change in what is or is not analysed, a change in how results are acted on, or a change of provider.