Compliance

Read this first

Briefing Room makes fully automatic hiring decisions. A score against a job's rubric can invite a candidate to a screening interview, or end their application, before any person has read it. That makes it an automated employment decision tool under several US laws.

This page describes what the software does. It is not legal advice and it is not a claim that any customer is compliant. Several of the obligations below are the employer's and cannot be discharged by a vendor — those are called out explicitly, because a compliance document that blurs the line between "the platform does this" and "you must do this" creates a false belief in exactly the person who most needs an accurate one.

Scope is the United States, by decision. GDPR, the EU AI Act and India's DPDP Act are out of scope, and nothing here should be read as addressing them. Several mechanisms — erasure, retention limits, notice — resemble what those regimes want, which is a consequence of good design and not a claim of coverage.

What happens to a candidate, and what is automatic

1. Apply. The candidate is shown a versioned disclosure and ticks a consent box. An applicant who applied elsewhere gets one email carrying the same disclosure, and must click before anything moves.
2. Résumé assessment — automatic. Parsed, redacted, and scored by a model against the job's versioned rubric, with a rationale and a quote per criterion.
3. Invite or reject — automatic, without a person. Compared to the job's invite and reject bars. At or below the reject bar the application is closed and the candidate emailed with the reason and a human-review link.
4. Screening interview — automatic. A video call with an AI agent, the candidate's camera required, recorded as video and audio and transcribed, after a second explicit consent to automated assessment. Only the transcript is scored; the video is never given to a model.
5. Interview assessment — automatic. Scored against the same rubric, every score citing a moment in the transcript.
6. Advance or reject — automatic, without a person. Compared to the advance and reject thresholds. There is no coverage floor: a score decides however little of the rubric produced it.

Two things are true throughout. Every automatic decision is recorded with what drove it — the score, the threshold, the criteria that moved it, the model and prompt version — so the inputs are reproducible. And a person can be asked for at any point, with the link in every rejection email and every screening invitation rather than a phone number in a footer.

The controls

The public AI-hiring notice

A stable, linkable page per organization — "How {company} uses AI in hiring" — served without authentication, because the people it is written for have no account. It states what is assessed, that an automatic decision can be made, how to ask for a human review, and how to ask for an alternative process, and it carries a slot for the organization's own paragraph and one for their published bias-audit summary.

The page quotes the apply-time disclosure and the automated-assessment disclosure verbatim rather than paraphrasing them, so rewording a disclosure changes the notice with it. Its own version is stamped on it, so an employer asked "what did your notice say in March" has an answer.

When the audit-summary slot is empty, the page says "not published" rather than omitting the section — an absent section would read as "no audit is required here".

Versioned, append-only disclosures

Two disclosures reach candidates: one at apply time, one before an AI-conducted video interview. Both are versioned and the version is recorded with the consent. A reword is a new version, never an edit to an existing one, because editing in place would rewrite history for every candidate already stamped with it. Asking for an unknown version raises rather than falling back to today's wording, because a quietly false compliance record is worse than a loud failure.

Recording consent and consent to automated assessment are asked separately from consent to recording, because a tick-box about a recording is not consent to an automated rejection.

Voluntary EEO data that cannot leak

Self-identification is collected so that a bias audit is possible at all. The same answers would let the product discriminate, so the database itself keeps them shut away:

• Nobody at your company can read an individual's answers back, and nobody can change or delete them.
• The only thing that reads them is the per-category selection-rate report, which covers your own company only and suppresses every group smaller than five.
• A suppressed group is not zero. It is marked as suppressed in the export and on screen. A reader who takes a blank for a zero has been told the opposite of the truth.
• The candidate's form always reports success, whether or not they had already answered, because "has this person already told you their race?" is itself a question it must not answer.
• Demographic data never reaches any model, and automated checks on our code enforce it.

The human-review queue

Two routes: contesting an automatic decision, from the link in every rejection email; and asking not to be assessed by an AI at all, from the screening invitation. The second also suppresses the pending screen's reminders, and it works even after the screening link itself has expired — someone working out how to ask for an accommodation is exactly the person who must not be told "expired".

Requests appear on the pipeline board with an age and an SLA badge that turns red past two business days, and an ignored queue alerts us.

There is deliberately no resolve button. A request is closed only by actually moving or rejecting the candidate, so the queue can only be emptied by acting on a person. A resolve button that cleared a badge would let the queue be emptied without anybody reviewing anything. An automatic decision never closes a request.

Business days are counted Monday to Friday, and public holidays are deliberately not modelled — there is no per-organization holiday calendar, and inventing one would make the badge wrong in a less predictable way.

The never-ask list

One list of nine topics, written down once and read by three parts of the system: salary history, criminal history, age or graduation year, family or pregnancy or childcare, religion, national origin or citizenship beyond work authorisation, health or disability, union membership, and sexual orientation.

• The AI interviewer's prompt states the list as absolute and outranking the job's own must-probe list, with a redirect for a candidate who volunteers one.
• The email agent refuses to raise or answer any of them and escalates to a person.
• A transcript scan reads AI-voice transcripts afterwards, because a prompt is a request and a transcript is a fact. A never-ask topic found in an interviewer turn is recorded quoting the turn verbatim, and an admin is notified.

Two questions are deliberately not on the list, and the prompts say so: work authorisation, and whether the candidate needs an adjustment or accommodation. A scan that flagged the accommodation offer would train the interviewer out of making it.

The job compliance lint

Advisory findings against what a job asks for — a rubric criterion, a line of the description, or an interview must-probe that would be unlawful for a person to write and is no more lawful for having been drafted by a model. Deterministic patterns run first; one best-effort model pass looks for the phrasings a pattern cannot anticipate, and a model finding that cannot quote the offending text verbatim is dropped.

It also raises a missing-salary-range finding when a job's location names one of twelve pay-transparency jurisdictions and no range is set.

It never blocks. An admin may acknowledge a finding with a reason, and the reason is recorded — that recorded reason is the point, and a dismiss button that left nothing behind would destroy it. The owner chose advisory over blocking knowingly: a lint that can stop a hire gets switched off.

Erasure anonymises; it never cascade-deletes

Deleting a candidate row would cascade through their applications, events and assessments. That is wrong twice over: it would silently change the denominators of every bias audit — a rejected candidate quietly leaving the "applied" count makes automated rejection look fairer than it was — and it would destroy the record of decisions already made about a real person.

So the person is nulled and the decision is kept.

Removed: name, email, phone, links, location, headline, current title and employer; the résumé text and the stored file; transcripts, transcript turns and the recording; the model's prose about the person — summaries, rationales and quoted evidence; email bodies, quoted chains, headers and addresses; and the demographics row, deleted outright.

Kept: the application, its status and its stage history; the scores themselves and which criterion each measured; the recommendation, confidence, coverage, model, prompt version and rubric version; email subjects and timestamps — the fact that we wrote, and when; and every decision event with its actor, its score and its threshold.

The contestable judgement, stated because it is contestable: removing the model's prose removes the explanation of a decision, and an explainable decision is one of the things this kit exists to guarantee. It goes anyway, because that prose routinely recites what the candidate said about where they worked and what they built, and the person has asked to be forgotten. What survives is what an audit needs. An auditor can still ask "what does this model do to people who look like this"; they cannot ask "what did it say about this individual", which is correct.

Proof of erasure is kept without keeping a copy: a hash of the lower-cased email address and a count of deleted objects, never the address and never the paths.

Storage objects are deleted out of band, after the row is anonymised — the right order, because a crash in between leaves an unreferenced file and a retryable event rather than a row pointing at a file that is gone.

Retention is the same code on a clock

The organization's retention window defaults to 730 days and is configurable per organization. The clock starts at the application's final decision, not at the application date — a candidacy still open is not stale however long it has been open.

A candidate is anonymised only when every application they hold is decided and past the window. One person applying to four roles is not erased because the first was rejected two years ago.

The sweep runs on a schedule and uses the same anonymisation as an erasure request. Two implementations of "remove this person" would drift, and the one that drifted would be the unattended one. The only difference is the reason stamped on the erasure log — which is exactly the difference an auditor asks about.

Raw ATS payloads are a shorter, separate promise: nulled after 30 days regardless of the organization's window.

What remains the employer's duty

• Validating that each requirement is job-related and consistent with business necessity. The platform shows that a criterion moved a score; it cannot tell you whether requiring it is defensible.
• Watching the selection rates and acting on them. The platform can show that automatic rejection rates differ by category. It does not stop you.
• The independent bias audit required by New York City's automated employment decision tool law. An audit performed by the vendor of the tool is not independent, by definition. The platform provides the data an auditor needs and a slot for the published summary.
• The ten business days' notice before the tool is used. For a New York City job or candidate the platform holds every automatic decision for ten business days after the company's first published job, and decides at once everywhere else. It does not otherwise enforce or time the notice: the notice page has no publication date the platform controls, and the timing depends on when the employer links it from a posting. A customer hiring in New York City who links the notice page on the day they open the role has not met the ten-day requirement.
• Running an interactive accommodation process and actually providing the alternative. The platform records the request and puts a clock on it; a person has to answer it.
• Answering a destruction request when it arrives. An admin presses the button; nothing here reads a candidate's email and acts on it.
• Deployer obligations under state AI acts — the impact assessment, the risk management programme, and any required notification to a state Attorney General.

Known gaps

Written down because a compliance document that lists only its strengths is a sales document.

1. The bias audit is not ours and nothing checks that one exists. The audit-summary slot can stay empty forever; nothing blocks a job or warns an admin.
2. The ten-business-day notice is not enforced or timed.
3. The job lint is advisory and blocks nothing. A job whose every line trips a rule activates normally, and nothing checks whether an acknowledgement's stated reason is a good one.
4. The lint's patterns miss things. They are written to avoid false positives, which necessarily lets unusual phrasings through. A bare experience range is not read as a cap, because most ranges are lawful, and nothing reads a job's uploaded attachments. The model pass that catches unusual phrasings is best-effort and contributes nothing when the model is unavailable.
5. The transcript scan covers AI-voice rounds only, and only after the fact. A human-led round is the interviewer review's territory. A question that must never be asked is detected after it was asked, never prevented.
6. No automatic decision has a coverage floor, at any stage. A barely-readable résumé that clears the readability gate is scored and can be rejected on that score, and so can an interview that evidenced two criteria of eight. That is deliberate; the safeguard is the human-review route, not a rule that suppresses low-coverage scores.
7. The default retention window is shorter than California's four-year ADS record rule. It is per organization and configurable, but it is not defaulted to four years and the platform does not warn about it. A California employer should raise it.
8. Keeping demographic data away from scoring rests on checks of our own code. Your team's access to it is limited by the database itself. Our background processes are kept away from it by automated checks on the code, which stop a change from shipping but do not limit someone with direct database access.
9. Selection rates count "invited" as an AI screen having been set up for the candidate, which is close to, but not exactly, being invited. It suits comparing invitation rates; it is not a measure of offer rates, and none is provided.
10. Suppression is per group, not per report. A determined reader with two exports taken at different times could in principle narrow a suppressed cell by differencing. Nothing limits how many exports can be taken.

Sub-processors

What leaves Briefing Room, and with what. Every integration is optional, and an integration you do not use is sent nothing.

• Model inference (résumé scoring, interview scoring, rubric drafting, the email agent) — redacted résumé text; interview transcripts, not redacted, because the assessment is of what was said; job requirements; and inbound candidate email bodies.

• Document parsing and OCR — the résumé file as uploaded, unredacted; redaction happens after extraction.

• The screening voice agent and telephony — live audio, the candidate's name, the job title, and the interview brief.

• Real-time media for browser interviews — live audio and video, and the interview context.

• Speech-to-text and text-to-speech — live audio.

• Email delivery and inbound — name, email address, message bodies.

• Scheduling — name, email address, interview time.

• Calendars, for interviewers who connect one — the interviewer's busy times are read; a booked round's time and the panel's addresses are written to one panellist's calendar, the lead's where it is connected. The candidate is not an attendee.

• The meeting note taker, where configured — the audio of a Zoom, Meet or Teams round, deleted after 30 days, and its transcript. It joins only after every interviewer has agreed and the candidate has been told; see Integrations.

• Granola, for interviewers who connect it — no candidate data is sent to it; the interviewer's own notes are read from it when they import one.

• Database and object storage — everything; this is the system of record. Résumé files and recordings sit in private buckets reached only through short-lived signed URLs.

• Error monitoring, if configured — deliberately minimised: personal identifiers off, and console, click and input breadcrumbs dropped, because on the review screen those would capture quoted transcript evidence. Session replay is not installed; it would film the interview.