AI for recruitment

With AI for recruitment, your ATS tells you how many of a job board's applicants were already on your database before the board gets credit, with the query behind each count. Consultant notes give back lost stage dates, a monthly list shows the records past retention, and none of it scores, ranks or rejects a candidate.

Last reviewed ยท Virgo Analytics

One question, traced, example

126 of Job board A's 420 October applicants were already on the database

Example data

You asked

How many of last month's applicants were already on our database?

Checked October applicants against the database6 steps

  1. Read the question0.2s

    You want to know how many people who applied last month were already candidates here before they applied. Last month is 1 to 31 October 2026.

  2. Decide whether to query0.1s

    This needs the live candidate tables. Applications arrive daily, and the match has to run on today's records.

  3. Pick the dataset0.3s

    Used applications for the source and date of each October application, and candidates for email, phone and the date each record was created.

    applicationscandidates

  4. Write the query0.5s

    Take each candidate's first October application and its source. Count a candidate as already held when a record created before 1 October shares their email or their phone number. Blank emails and phone numbers are left out of the match.

    WITH october_applicants AS (
        SELECT
             a.candidate_id
            ,a.source
            ,NULLIF(LOWER(TRIM(c.email)), '') AS email_key
            ,NULLIF(RIGHT(REGEXP_REPLACE(c.phone, '[^0-9]', '', 'g'), 10), '') AS phone_key
            ,ROW_NUMBER() OVER (PARTITION BY a.candidate_id ORDER BY a.applied_at) AS application_order
        FROM ats.applications a
        INNER JOIN ats.candidates c ON 1=1
            AND c.candidate_id = a.candidate_id
        WHERE 1=1
            AND a.applied_at >= DATE '2026-10-01'
            AND a.applied_at < DATE '2026-11-01'
    )
    ,held_before AS (
        SELECT
             c.candidate_id
            ,NULLIF(LOWER(TRIM(c.email)), '') AS email_key
            ,NULLIF(RIGHT(REGEXP_REPLACE(c.phone, '[^0-9]', '', 'g'), 10), '') AS phone_key
        FROM ats.candidates c
        WHERE 1=1
            AND c.created_at < DATE '2026-10-01'
    )
    ,matched AS (
        SELECT DISTINCT
             o.candidate_id
        FROM october_applicants o
        INNER JOIN held_before h ON 1=1
            AND (h.email_key = o.email_key OR h.phone_key = o.phone_key)
    )
    SELECT
         o.source
        ,COUNT(*) AS applicants
        ,COUNT(m.candidate_id) AS already_held
        ,COUNT(*) - COUNT(m.candidate_id) AS new_to_database
        ,ROUND(100.0 * COUNT(m.candidate_id) / COUNT(*), 1) AS pct_already_held
    FROM october_applicants o
    LEFT JOIN matched m ON 1=1
        AND m.candidate_id = o.candidate_id
    WHERE 1=1
        AND o.application_order = 1
    GROUP BY o.source
    ORDER BY already_held DESC
  5. Run the query1.2s

    Query finished against the read-only view. Nothing was written.

    4 rows0.6 GB scanned

Write the answer0.3s

Every figure below comes from those 4 rows. The total line adds them up and nothing else.

  1. Job board A420 applicants

    126 already held,294 new to the database,

    Matched on email or phone to a record created before 1 October
  2. Careers site150 applicants

    30 already held,120 new to the database,

  3. Job board B180 applicants

    27 already held,153 new to the database,

  4. Referrals50 applicants

    5 already held,45 new to the database,

  5. Total (sum of the four rows)800 applicants

    188 already held,612 new to the database,

    188 = 126 + 30 + 27 + 5

188 of October's 800 applicants, 23.5%, were already on the database before they applied. Job board A sent the most applicants and the most people you already held: 126 of its 420, or 30.0%. The ATS credits each application to the source it arrived through, so Job board A is credited with all 126. Referrals brought the fewest people already held, 5 of 50.

Show the numbers
The 4 rows and their total, example data
SourceApplicantsAlready heldNew to the databaseAlready held
Job board A42012629430.0%
Careers site1503012020.0%
Job board B1802715315.0%
Referrals5054510.0%
Total (sum of the four rows)80018861223.5%

4 rows0.6 GB scanned2.6s total

Example data: an invented agency's applications for October 2026, one row per candidate record, counted under the source of its first October application. In the AI-led analytics demo you can watch the same six steps run on a scripted question and open the query yourself.

What works now in AI for recruitment

Duplicate candidates matched before source creditThe match compares each applicant's cleaned email and phone with records created before the month began, and agreement on either one counts as a match. It needs the ATS candidate table with email, phone and the date each record was created, and the applications table with source and date. What duplicates do to the funnel is on the recruitment analytics page.
Stage dates read from consultant notesWhere the ATS kept only each candidate's current status, the dates in between may still be in the notes: "CV sent to client 12 Oct", "2nd interview booked for Thursday". The AI pulls out each event and its date and stores them with the note's ID. Each month, before the dates reach a report, a consultant opens some of those notes and checks each date against the words. It needs the notes or activity log with a candidate ID and a timestamp on every entry.
The retention listEach month you get a list of candidate records that have reached the end of the period your policy sets, with the record's created date, its last activity date and the rule applied. The list takes the last activity date from the latest dated note, application or stage change on the record, which is our working proxy. A person acts on the list, and the AI deletes nothing because it reads through a read-only account. It needs a written retention period and both dates stored.

What does not work yet

Screening CVs or shortlistingThis work reports on your own records and makes no decision about anyone. Automated decisions are covered once, on the hub page under What every sector shares.
Matching on a name aloneTwo candidates with the same name are two people until an email or phone number says otherwise, so name-only pairs go to a consultant's list unmatched.
Stage history nobody wrote downIf the ATS overwrote a status and no note mentions the event, the date is gone, and the AI marks the gap and leaves it empty. The history builds again from the day you start storing stage dates.

A job board can be credited with a candidate your database found two years ago.

What has to be true of your recruitment data first

  • Each candidate record keeps its first source and its created date

    The match only looks at records created before the month began, so if an update resets the created date, a record from 2024 looks new and the match passes over it.

  • A retention period is set and written down

    The retention list needs a number of months and the date it counts from, and without both there's nothing to report against.

  • Stage dates are stored, and kept when the status changes

    Reading notes recovers only the dates a consultant typed, so from now on each status change needs a dated row of its own.

The AI readiness assessment runs these three checks on a copy of your ATS, then answers ten of your consultants' questions by hand.

Regulation and risk in recruitment AI

The sources we read on 20 November 2026, quoted or named.

The ICO's audit outcomes report, AI tools in recruitment (November 2024)

Its executive summary says:

"The ICO have carried out consensual audit engagements with developers and providers of artificial intelligence (AI) powered sourcing, screening, and selection tools used in recruitment."

It also says:

"Additionally, features in some tools could lead to discrimination by having a search functionality that allowed recruiters to filter out candidates with certain protected characteristics. Others estimated or inferred people's gender, ethnicity, and other characteristics from their job application or even just their name, rather than asking candidates directly."

Its section "Human reviews in AI" says:

"AI tools reviewed in this work were designed and intended only to support human recruiters to make decisions, rather than to make automated recruitment decisions without human intervention."

The same section prints this point:

"Consider: Prevent using AI outputs to make automated recruitment decisions. Prevent human recruiters from progressing or rejecting candidates based solely on indicative grades or fit scores produced by the AI. Clearly record the intended use in contracts, marketing materials, or training and resources you provide to recruiters."

Its case study of Organisation A says:

"Their published privacy information stated that candidate profiles in its database were retained for one year. However, the retention period was restarted every time profiles were updated with new information, which meant in practice the majority of their database was retained indefinitely."

ICO, AI tools in recruitment: audit outcomes report (PDF), linked from the ICO's page AI tools used in recruitment. Checked 20 November 2026.

Equality Act 2010

Section 19 covers indirect discrimination.

Checked 20 November 2026.

Conduct of Employment Agencies and Employment Businesses Regulations 2003

Regulation 29 covers records.

Checked 20 November 2026.

This is a list of the sources we read, with the date. It is not legal advice, and your own adviser decides what applies to you.

Duplicates come out before any source gets credit, because the other order pays a job board for people you already had.

Further reading

Questions people ask

Can AI screen CVs for our agency?

No, not in this work. The AI reads your own records to report on the business: where candidates came from, when they moved stage and which records are due for deletion. A consultant makes every decision about a person.

How do you match duplicate candidates without matching the wrong people?

The match uses email address or mobile number, cleaned first: the email lower-cased and trimmed, and the phone number cut to its last ten digits. Blank values never match, and nor does a name alone, because two people can share one.

Will matching duplicates change which job board gets credit?

It can, and it shows you how much credit is in question. In the example above, 126 of one board's 420 October applicants were already on the database. You write down the credit rule, such as first source recorded, and the report applies it the same way every month.

Can AI fill in stage dates our ATS didn't keep?

Only where a consultant wrote the event down. If a note says a CV went to the client on 12 October, the AI reads the date and stores it with the note's ID beside it. If nobody wrote anything, the gap stays marked as a gap.

Does AI in recruitment reporting need candidate names?

No. The matching step reads email and phone to find duplicates, and everything after it runs on candidate IDs, sources, stage dates and counts. Reports show counts by source and role type. A query that would return names is outside the access map, and the AI refuses it.

How many of your applicants were already in your database?

One month of applications with the source column lets the first call match them against your database and count the people you already had.

Book a call

Free 30 min call