Deciding

The problem it solves

A score of 92 built on two criteria out of eight looks exactly like a 92 built on all eight. One is a judgement about a candidate; the other is a judgement about two criteria. So every score is shown with its coverage: the share of the rubric's weight the interviews actually evidenced.

Coverage is shown, not enforced. No decision waits for a minimum coverage, so a score is acted on however little of the rubric produced it. That is why the coverage beside it matters to whoever reads the decision.

The two numbers

A job decides with two numbers:

• Advance: at or above it, the candidate goes forward. At the résumé stage that means an invitation to the AI screen; after a round, the next round.
• Reject: strictly below it, the candidate is rejected and emailed, with a link to ask for a human review.
• Between them, a person decides.

The same pair is read at the résumé score and after every scored round. New jobs start at 80 and 65, and you can change them per job. A job cannot be published without both (see Publishing a role).

An advance starts the next round: a one-use link for the AI screen, or a link to pick a time for a round a person leads.

A résumé score and an interview score are produced under opposite rules about silence and are never averaged together. The thresholds are one pair, but each moment reads the score that belongs to it.

Automatic decisions are off until you turn them on

Settings → Hiring → Automatic decisions is off for every new company. While it is off, each decision the numbers would make waits for a person in Inbox → Needs you, with the scoring model's recommendation and its reasons, ready to accept in one click.

An admin can turn it on, and then switch off any of the four decisions it covers:

• reject after the résumé;
• invite to the AI screen after the résumé;
• advance after a round;
• reject after a round.

Anything switched off keeps waiting in Inbox → Needs you.

When a decision waits for a person anyway

Even with automatic decisions on, these hold a decision for a person, who is told why in Inbox → Needs you:

• The candidate never applied. Someone your team uploaded is never rejected or advanced automatically. Once your AI notice period has passed, one who clears the advance number at the résumé stage is still invited to the AI screen like anyone else.
• New York City. For a job in New York City, or a candidate who says they are there, nothing is decided automatically until ten business days after your AI-hiring notice went live.
• An interviewer disagrees. A yes from someone on the round blocks an automatic reject; a no blocks an automatic advance.
• A deal-breaker has no score. An automatic advance out of a round waits if any of the rubric's deal-breakers has not been scored.
• The work was started by an API token or an assistant. Only a person can start something that reaches a candidate.

Two facts before anything is decided at the résumé stage

No automatic decision happens without a final score and a recorded consent. The two arrive in either order: an applicant on your careers page consents first and is scored afterwards, while one from elsewhere is scored first and consents when they click the link in the disclosure email. The decision is taken when both are in.

An application whose résumé could not be read never gets a final score, so nothing automatic can end that candidacy.

The rule after a round, in order

1. No score yet: a person looks.
2. At or above the advance number: advance. Equality advances, because 80 typed as the bar means "80 is good enough".
3. Strictly below the reject number: reject. Equality does not reject: 65 typed as the bar means "65 is still worth a look". The résumé stage reads the same way, so one score cannot be rejected at one moment and reviewed at another.
4. Otherwise: a person looks.

The reject number can never be above the advance number, and a score sitting exactly on both advances, because advancing on a tie is the easier mistake to undo.

What an automatic decision records

Every automatic move is written to the candidacy's timeline, marked as taken by the system rather than a person, with the score and the threshold it was measured against. A rejection email says the decision was made automatically and carries a link to ask for a human review.

Offers

An offer belongs to an application. Sending it moves the candidate to the job's offer stage. The candidate signs, someone at the company countersigns, and the candidate moves to the job's hired stage. A sent offer is never edited; a change is a revised offer that replaces it. See Settings → Offers for the letter your offers are written on.

There is also an offer-preparation assistant that coaches you on the offer conversation and negotiation. It decides nothing.

What you do

• Set the two numbers per job, or keep the 80 and 65 a new job starts with.
• Decide whether to turn on automatic decisions, and which ones.
• Work Inbox → Needs you. Everything waiting on a person is there.
• Read the reasoning on any automatic move on the candidacy's timeline.

Limits and edge cases

• A fallback résumé score decides exactly as a model score does. It is compared against the same thresholds. It is visibly marked as a fallback.
• There is no coverage requirement at any stage. Coverage is shown beside the score, not enforced.
• Nothing blocks a threshold that is producing a disparate impact. The product can show that automatic rejection rates differ by category; it cannot tell you whether the difference is justified, and it does not stop you.
• Rejecting a candidate stops what was already scheduled for them. See The pipeline.
• Automatic decisions never close a human-review request. Only a person acting on the candidate does.
• A hire is a stage move. The candidate's place in the job's hired stage is the record.