LL144, EU AI Act, NDPR — ready out of the box.
Three jurisdictions look at AI screening three different ways. We built the audit log, the candidate notice, and the human-in-the-loop default that all three of them want — so compliance shouldn’t cost six figures.
Per-session audit log · Content-only scoring · Candidate notice built in · Human-in-the-loop by default
Three jurisdictions
The same audit log satisfies New York, Brussels, and Abuja.
Different regulators ask for different things, but the underlying evidence is the same: what was said, what was scored, by which model, with what consent.
Local Law 144 (AEDT)
Requires a bias audit within the prior year, a public summary, and 10-business-day candidate notice before an Automated Employment Decision Tool is used.
What we provide
Per-session audit log capturing rubric criteria, per-criterion scores, justifications, and the model + prompt versions that produced the score.
Planned: exportable bias-audit summary in the format DCWP expects from independent auditors.
Candidate notice shown on the apply page before the interview, including the qualifications and characteristics evaluated.
EU AI Act
Classifies AI used in employment as high-risk. Requires transparency to the candidate, human oversight, and record-keeping of system behavior.
What we provide
Risk classification doc mapping Vettika to the high-risk employment category (currently in draft).
Transparency notice shown to every candidate before the interview (what the AI does, what it does not, how to decline).
No emotion-recognition component anywhere in the product. The AI Act prohibits workplace emotion inference (Article 5(1)(f), in force since February 2025) — we satisfy it by architecture: scoring input is transcript text only.
Session records — turns, scores, justifications, and the model + prompt versions behind each score — kept for the retention window you choose, exportable per session and per role as JSON or an LL144-shaped audit summary.
Human-in-the-loop by default: the recruiter is the final decision-maker. We never auto-reject.
NDPR + NDPA 2023
Data minimization, lawful basis for processing, and a clear path to the data controller. African candidate data should not be shipped where it does not need to go.
What we provide
Audio destroyed on hangup by default — we keep the transcript, not the recording.
Planned: region-aware processing for traffic originating from Nigeria where the underlying providers support it.
DPO contact path published in the candidate consent step and on the privacy policy.
Bias audit log
Everything a regulator (or your candidate) might ask, logged.
Every session writes a structured record — the evidence an independent bias auditor needs to produce their report. Candidates can export their own data, and recruiters can download each interview (or a whole role) as JSON or an LL144-shaped audit summary, including the human decision trail.
Per-session record fields
Rubric criteria the recruiter defined (per-role).
Per-criterion score (0–5) and the natural-language justification the model produced.
Candidate consent record — what was disclosed, what was accepted, timestamped.
Full turn-by-turn transcript with speaker, start time, and end time.
Recruiter identity + organization for chain-of-custody.
Model versions captured per session (LLM, STT, TTS providers + revisions).
System + scoring prompt versions and content hashes, never edited in place.
Scoring
How we score — and what we explicitly don’t do.
Content-only: the transcript is the only scoring input.
The scoring model receives the interview transcript text and your rubric. That is the entire input. No emotion recognition, no facial analysis, no voice-tone or prosody scoring — there are no such systems anywhere in our stack to switch off.
You bring the rubric.
Criteria are recruiter-supplied. We do not impose generic templates. If your role weighs ownership at 40% and communication at 10%, that is the weighting the score uses.
Per-criterion, not single-number.
Every criterion gets its own score and its own justification quote pulled from the transcript. No mystery composite.
Model + prompt versions captured per session.
Every completed interview records the exact LLM, speech-to-text, and text-to-speech model versions plus the system and scoring prompt versions (with content hashes) that produced its score — so "what produced this score?" is answerable from the stored record.
Recruiter is the decision-maker.
We summarize what was said against your rubric. The hire / no-hire call belongs to a human — your human.
Anti-claim
We don’t make hiring decisions. We summarize what was said against your rubric.
Data handling
We keep the transcript. We do not keep the audio.
Audio destroyed on hangup
Default behavior. Voice frames are processed for speech-to-text, then discarded. No raw audio or video is retained, and no audio features (tone, pitch, prosody) feed scoring.
Transcript stored, candidate can download
The artifact both sides actually need. Candidate gets a copy in their inbox.
Exports for both sides
Candidates can export their data today. Recruiters can export any interview or role as JSON or an LL144-shaped audit summary — decision trail included.
Candidate consent in the loop
Disclosed before the agent joins. Declines end the call cleanly, no scoring.
What we don’t cover
We are not a law firm. This is not legal advice.
Compliance is a process, not a checkbox. The pages above describe the surface AI Recruiter provides — the audit log, the consent flow, the human-in-the-loop default. Whether that surface, plus how you use it, satisfies your specific regulator in your specific jurisdiction is a question for counsel familiar with your hiring practice. We will share what we know with your lawyer and your auditor on request.
FAQ
The compliance questions we actually get.
Do you provide the LL144 bias audit?+
We log per-session scoring evidence today and are building the exportable summary an independent auditor needs to produce the LL144 bias audit. We are not ourselves the independent auditor — that has to be a third party by law. No independent audit has been completed yet; when one is, the summary will be published at /legal/aedt-bias-audit.
What about GDPR?+
GDPR applies to anyone processing EU resident data. We minimize what is collected (audio destroyed on hangup, transcript retained per your policy), support data subject access requests against the audit log, and the EU AI Act work above sits on top of GDPR — not instead of it.
Do you analyze tone of voice, facial expressions, or emotion?+
No. The scoring input is the interview transcript text plus your rubric — nothing else. There is no emotion-recognition, facial-analysis, or voice-tone (prosody) model anywhere in the stack, so there is nothing to disable or configure away. Audio is processed for speech-to-text and destroyed on hangup; we never retain raw audio or video. This also matters legally: the EU AI Act prohibits emotion-inference systems in the workplace (Article 5(1)(f), applicable since February 2025).
I am not in NYC. Does LL144 still matter to me?+
LL144 binds employers using AEDT for NYC-based candidates or NYC-based roles. If you do not screen NYC candidates today but might tomorrow, the LL144 surface costs nothing to keep on — we generate the audit log regardless of geography.
How does the consent flow work?+
Before the agent joins the call, the candidate sees a one-screen consent step: who is running the interview, that AI is conducting it, what is recorded, what is destroyed, who keeps the transcript. They click accept or decline. Decline ends the call with no scoring and no follow-up. The consent record goes in the audit log.
Do you have a vendor risk questionnaire response?+
Yes — happy to share on request. Email security@vettika.com. We answer real questionnaires from real procurement teams, not just SOC2 checkbox prompts.
Are you SOC2 certified?+
Not yet. SOC2 Type I is on the 2026 roadmap. We will not pretend otherwise to win a deal. If SOC2 is a hard procurement gate today, talk to us — we can usually answer the underlying controls questions without the cert and give you a written timeline.
Compliance that ships with the product, not a six-figure add-on.
Run your first three first-rounds free. The audit log is on from interview one.