How AI screening works

Every application that arrives with a CV is screened on its own: the CV is read, turned into a structured profile and scored against the job's requirements. Nobody has to press a button, and every verdict comes with the evidence behind it, so you can check it in seconds.

What happens to each application

1. 1

   Step 1: Queued

   Screening starts as soon as an application arrives, whether it came through an apply page, by email, from a bulk import or from the talent pool. A background check runs every couple of minutes and picks up anything that was missed, so no application waits for someone to remember it.

2. 2

   Step 2: Read the CV

   The text is pulled out of the PDF or Word file by document readers, not by an AI model, so what the AI later reads is exactly what the document says. The file type is worked out from the file's contents rather than its name. If no readable text comes out, for example from a scanned image, the application is not scored at all rather than scored on nothing.

3. 3

   Step 3: Build a profile

   An AI parser turns the CV into the same structured profile every time: roles and dates, skills, projects and education. That makes candidates comparable however their CV was laid out. The parser is told to record only what the CV says and never to judge, rank or score.

4. 4

   Step 4: Score

   A second model checks the candidate against each of the job's requirements and the numbers are worked out from its verdicts. When it finishes, the scoreboard appears on the candidate's page and your team gets a notification.

Each step saves its result as it goes. If something is interrupted part way, screening picks up again instead of losing the application.

Every verdict comes with evidence

For each must-have, preferred and nice-to-have requirement, the AI decides whether the candidate has Matched, Partial or Missing it. For every Matched or Partial verdict it must copy the sentence from the candidate's own CV that supports it, word for word. If it cannot quote anything, the requirement is Missing.

You see those quotes under the scoreboard, next to each requirement. See Reading the scorecard.

Quotes are checked against the source

Every quote is matched back to the original text automatically. A quote that cannot be found is not trusted: the requirement drops from Matched to Partial, its confidence is capped low, and the scorecard shows the AI's own sentence ending in (evidence unverified) instead of a quote. This catches a model that misremembers a CV, and a CV that tries to talk the AI into a better score.

The numbers are arithmetic

No model picks a score. The verdicts are turned into numbers by fixed rules:

• A requirement counts in full when Matched, half when Partial and not at all when Missing.
• The ATS score from 0 to 100 blends four parts: the must-have score, the preferred score, experience and technical depth. How much each part counts is set in your screening profile. See Tune your screening profile.
• The recommendation follows from the ATS score: Strong match at 80 or more with no must-have missing, Proceed with caution at or above the job's minimum ATS (55 unless you set your own), and Low match below it.

The same verdicts always produce the same score, and changing a weight changes every score in the same, predictable way.

What the AI does not use

The scorer never sees who the candidate is. Before scoring, the candidate's name, email, phone number, links and home address are replaced with neutral placeholders such as [candidate] and [email]. That happens in the CV text itself, not only in the structured profile, and also in the screening answers, the email thread and any shared pages. A gender title in front of the name, such as Mr or Ms, goes with it. Employers, schools, skills and the places in the candidate's job history stay, because they are evidence about the work.

The scorer is told that these details were removed and never to guess them, and that a name, a location or a photo is never a signal. Its quotes are checked against the text it read, then mapped back to the original lines of the CV, so the evidence you open always shows the candidate's own words, never a placeholder.

This applies to scoring only. When you ask ari about a candidate, it reads the full CV, contact details included.

Fairness rules are written into the instructions both models receive:

• The parser never extracts or infers age, date of birth, gender, family status, nationality, race, religion or disability. It records a right to work only when the CV states it.
• The scorer never lets protected characteristics influence a judgement. If interview feedback mentions age, gender, nationality, ethnicity, family status or health, that sentence is ignored.

Links the candidate shares

If a candidate gives a GitHub profile, a portfolio or a personal website, on the apply form or in their CV, screening reads the public page once. Its career-relevant facts appear on the candidate's page under Shared links, and a page can evidence a requirement the CV does not mention. A quote from a page only counts when it is found on that page.

LinkedIn and social media are never read, whatever field they were pasted into, and personal details on a page are ignored. The apply form tells candidates this happens. If a page cannot be read, screening simply carries on from the CV.

Emails with the candidate

Every scoring run also reads the email thread for the application, with each message labelled by who wrote it. Whenever the candidate writes, whether they answer an Opener email, reply to one of your messages or write on their own, they are re-scored automatically with what they said. A colleague forwarding the candidate's own email to the thread counts the same way. Your team's messages, copies of them sent to your BCC mailbox, and automatic replies such as out-of-office messages, when the candidate's mail system marks them as automatic, never start a re-score.

A re-score after a reply waits at least a minute, and at least ten minutes after the previous score, so several replies sent close together are read in one go. A candidate is re-scored this way at most six times a day; anything they write after that is read the next time they are scored. The candidate's activity shows Score updated after the candidate's reply. Your team is not notified again, since the reply itself already was. A candidate who has already been hired, rejected or has withdrawn is never re-scored this way, even if that happened while the re-score was waiting, and neither is an application that was not screened, for example one without a CV, or whose screening failed. See Email threads with candidates and Opener: the first contact.

Interview feedback

When interviewers submit their scorecards, the feedback is added to the assessment as first-hand evidence from your own team. It can confirm, upgrade or downgrade a requirement even when the CV says nothing about it, and where the CV and an interviewer disagree, the interviewer's direct observation wins. Such a line shows From Interview feedback with the interviewer's name.

The candidate is re-scored once for each new set of feedback. A candidate who has already been hired, rejected or has withdrawn is never re-scored this way. See Interviewer scorecards.

When screening cannot finish

• No CV to screen: the application has no CV, or no readable text could be pulled from it. Open the CV yourself to review it.
• AI screening failed: something went wrong, and the reason is shown. Screening is retried automatically up to three times. If the last attempt fails, your team gets a notification, and Retry screening on the scoreboard starts it again. See Rescoring.

If your workspace works in a language other than English, the scorecard's summary and assessments are written in that language. Quotes always stay in the language of the CV.

The AI only ever recommends. A person makes every decision. See The AI never rejects anyone.