AI inside a recruitment CRM can parse new CVs, suggest field updates, flag possible duplicate candidates and queue follow-ups from recorded events. Let it prepare changes, then have a recruiter confirm identity, consent, employment history and message context before merging or contacting anyone. Keep candidate ranking and hiring decisions outside this administrative workflow.
A CRM is only as useful as the record a recruiter trusts. Duplicate profiles split conversation history; stale titles mislead searches; blank contact preferences create awkward outreach. The practical aim is one defensible history per person and a clear next action, not a bigger database. Start by fixing how records enter the system before cleaning years of accumulated data.
Map the four points where a record changes
A candidate record changes when a CV arrives, when a recruiter adds a conversation note, when an outside source proposes an update and when the candidate responds. Each point needs a different approval rule. A CV parser can suggest a current title, but the recruiter should compare it with the dated document. A candidate's own reply may supersede an old preference. An external profile update is a lead to verify, not permission to overwrite what the candidate told the agency.
Define a minimum record: unique internal ID, name as provided, contact route, source and date, contact permission, current status, latest confirmed CV version, communication owner and next action. Optional fields should have a clear purpose. If a team cannot explain why it stores a field, do not ask AI to populate it. Keep original documents and prior values available so a correction does not erase evidence of how the record changed.
For example, a CV arrives showing a new job title and phone number. The parser puts “Operations Director” in the title field, but the document says “Acting Operations Director”. The recruiter corrects it, adds the CV date and asks the candidate whether the phone number is preferred for contact. A subtle title change matters when a colleague later prepares a client summary. The AI saved retyping, but the original document settled the fact.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowCheck what Bullhorn, Loxo and Manatal actually offer
These three vendors illustrate different administrative features; they are not a ranking. Bullhorn's Automatic Resume Parser help page says an incoming CV with an email address already in the database updates or attaches to the existing candidate according to the organisation's settings, rather than creating a new duplicate. That is an exact-email behaviour, not proof that every duplicate is caught. The same page describes fields it attempts to extract and the settings that affect an existing record.
Loxo's Data Agents page describes a Deduplication Agent that flags likely duplicate candidates, contacts or companies and recommends merges; it says a recruiter must confirm before the merge. Loxo also describes data hygiene and updating agents. Those claims warrant a live demonstration on the agency's own sample records. Check whether an external update is displayed as a proposal, what source it uses and how a recruiter can reject it.
Manatal's candidate sourcing page says it flags possible duplicate profiles from multiple data points and lets users merge them manually or automatically, while keeping several CV versions for one candidate. That makes the merge setting a deliberate governance choice. An agency should begin with manual review, test the audit and recovery options and only then decide whether any narrow class of duplicates is safe to merge automatically.
Do not assume that a feature is in your current plan, configured for your account or integrated with your existing inbox. Ask each vendor to show the candidate record before and after a parse, a duplicate suggestion, a rejected merge and an opt-out. Capture the exact screen and setting names from your account. Product pages change; this tutorial's workflow remains useful even if the vendor packaging does not.
Treat duplicate detection as a review queue
Use multiple identifiers to suggest a duplicate, but do not treat a match as proof. A shared email account, changed surname, reused phone number or common name can create a false match. Conversely, a candidate who changes both email and phone may escape a simple exact-match rule. The review screen should display two records side by side: source, dates, CV versions, contact permissions, applications, notes, assigned owner and recent messages.
An illustrative agency finds two records with the same name and overlapping work history. One record carries a personal email and an application for a finance role; the other uses a work email and records an earlier call about operations roles. The recruiter checks the CV dates and phone conversation, confirms they are the same person through an appropriate contact route, then selects a master record. The merge retains both emails with their status, both CV versions and the full activity trail. The recruiter does not replace a current contact preference with an older one merely because it sits in the older master record.
A second example goes the other way. Two records share a phone number, but the CVs show different people and the number appears to be a household contact. The AI flags a likely duplicate. The recruiter rejects it and marks the reason so the tool does not keep proposing the pair. A false merge here could reveal one person's job search to another. This is why even a high-confidence score must be examined against source documents and communication history.
Write the merge rule as a checklist. Confirm the same person, choose a master record, preserve applications and activity, resolve conflicting fields one by one, retain the original CVs, check contact permission, and verify that no queued automation fires twice afterwards. Record who approved the merge and how to reverse it. If the CRM cannot show what happened or undo a mistake, keep merging in a restricted administrator workflow.
Resolve a proposed merge field by field
The most damaging merge errors often happen after identity is confirmed. A system chooses a master profile and quietly treats its old fields as authoritative. Instead, make the reviewer decide where each field should come from. The following illustrative record shows how a recruiter might handle two profiles that are genuinely the same person. The actual rules should reflect the agency's data policy and the candidate's latest instructions.
| Field | Record A | Record B | Review decision |
|---|---|---|---|
| Personal address, confirmed last month | Work address from an older CV | Use personal address for contact; retain older address as history | |
| Title | Manager from last year's application | Acting director on recent CV | Use exact recent wording and record the CV date |
| Contact preference | Email only, candidate stated | Blank | Preserve email-only instruction |
| Applications | One closed role | One active role | Preserve both with distinct statuses |
After the merge, search the resulting profile for both application histories and the latest permission note. Then check whether an acknowledgement or reminder is scheduled twice. A reviewer should also inspect links to client submissions: an old submission must remain attached to its original role and date. Do not turn a duplicate cleanup into a silent rewriting of the candidate's history. If the vendor's merge screen hides these consequences, export a backup and test on a dummy record first.
Update fields with provenance, not silent overwrites
AI can extract employer, title, skills, certifications and contact details, but every field has a date and source. “Current company” from a CV uploaded two years ago is not current merely because the parser placed it in a current-company field today. Add a last-confirmed date and distinguish candidate-stated information from inferred or externally sourced information. If the system has only one field for a fact, keep the source note nearby.
An illustrative record has “available immediately” from an application submitted last year. A new CV arrives, but it says nothing about availability. The AI should not carry the old value forward as current. It should queue a question for the recruiter or mark the field unconfirmed. The same principle applies to work authorisation, salary expectations and contact preferences: absence of a new statement does not prove an old one still holds.
Run a small sample after every parsing change. Select 20 incoming CVs, compare each extracted field with the original and record corrections by type. If titles are mostly right but dates are routinely wrong, keep parsing for convenience while requiring date review. If contact details are copied from an old page footer, stop automatic updates to that field. This is a more useful evaluation than a vendor demo with one tidy CV. Evaluating AI features in recruitment software covers the wider buying decision.
Queue follow-up from real events and promises
The CRM can flag a candidate who applied but has no acknowledgement, a recruiter who promised an update by a date, or a conversation that needs a reply. These are administrative triggers. Do not send a generic sequence simply because a candidate has been inactive for a fixed number of days. A candidate may be waiting for a client's decision, have declined contact or be in a sensitive discussion. The queue should show the triggering event and the last human note.
For example, a recruiter notes “will update after client feedback on Thursday”. On Friday morning the CRM creates a task. The AI drafts: “Hi [first name], the client has decided not to proceed.” That is an invented outcome; the note contains only a promised update. The recruiter changes it to: “Hi [first name], I promised an update after client feedback. I am still waiting for a confirmed decision and will let you know when I have one.” The candidate gets an honest message, and the CRM records the next check date.
Another illustrative case is a duplicate follow-up. The same candidate applies through two job boards and the CRM creates two acknowledgement tasks. The merge reviewer should resolve the duplicate before either sequence runs, then keep the correct application references in one record. If the agency cannot pause automation while a duplicate is under review, that is a reason to pilot draft-only follow-up first. Candidate follow-up automation covers sequencing in more depth.
Every message rule needs a stop condition: candidate replies, opts out, asks for a person, changes contact preference, withdraws an application or raises a complaint. The reviewer checks the role name, previous promise, status and recipient. Do not let AI turn “no update yet” into a false rejection or “we have your CV” into an implied interview invitation. A personal message is useful only when it is anchored to the actual record.
A useful message queue has columns for reason, last verified event, draft, reviewer and disposition. For example: “Reason: promised feedback; event: recruiter call note dated Tuesday; draft: no decision yet; reviewer: account owner; disposition: edit and send.” Another row might read “Reason: no reply after application; event: candidate opted out yesterday; disposition: cancel”. These filled rows show why a queue is safer than a timer that sends every draft. The AI can prepare the text, but the event and permission determine whether contact is appropriate at all.
When a candidate replies, the original queued task should close or change owner immediately. Test this with a reply that arrives minutes before a scheduled send. If the system does not cancel the draft, use a manual review window rather than timed automatic delivery. The same rule applies when a recruiter phones the candidate: the call note must update the queue so the person does not receive a redundant “we have not spoken” message. Contact quality is partly a synchronisation problem, not a writing problem.
A four-week pilot with measurable record quality
Imagine an illustrative five-recruiter agency processing 240 incoming CVs in four weeks. It chooses two tasks: parse new records into a review queue and suggest duplicate pairs. Before switching on the feature, a recruiter manually samples 30 older imports and finds six field corrections and four duplicate pairs. This baseline is a local measurement, not an industry average. The agency then processes the first 40 new CVs with every AI-suggested field reviewed against the original.
In the pilot, the parser produces 40 records. Reviewers correct three titles, two employment dates and one phone number. The deduplication queue proposes nine pairs; recruiters approve five, reject three and leave one unresolved pending candidate confirmation. That means the team should not report “nine duplicates removed”. The unresolved pair remains visible and its automated messages are paused. The agency also finds that two approved merges would have lost an older application note without a manual preservation step, so it changes the checklist.
Suppose the 40 records create 14 follow-up tasks. Recruiters send nine edited messages, close three because a response already arrived and hold two because the underlying client decision is unknown. The dashboard should show those four dispositions, not 14 automatic contacts. Time is measured as import preparation, field review, merge review, message editing and rework. If the review burden exceeds manual entry, the agency can narrow the fields it extracts rather than discarding the whole pilot.
For a filled daily control note, use: “New records 12; parse corrections 2; duplicate pairs proposed 3; approved 1; rejected 1; pending 1; follow-up tasks due 4; messages sent after review 2; opt-outs 1.” The note makes hidden work visible. It also helps a manager spot a sudden rise in duplicate suggestions after changing an import feed.
At the weekly review, sample at least one record from each disposition rather than only successful merges. Inspect a rejected pair to see whether the same false match returns, a pending pair to see whether someone owns the resolution, and a sent message to verify that its trigger was real. Track correction categories separately. A high number of amended job titles may mean the CV parser mishandles a format; repeated missing permissions may mean the application form is incomplete. Each points to a different fix. Do not ask recruiters to compensate forever for a broken import.
Keep administrative AI separate from selection
It is tempting to reuse cleaned CRM data for automatic suitability scores. That is a different project. This workflow may organise a record and remind a recruiter to reply; it does not decide who deserves an interview or which candidate is best. Candidate ranking, screening and selection require their own evidence, fairness checks, client agreement and qualified advice on applicable employment and data rules. A well-maintained database is a foundation, not permission to automate a consequential decision.
Set role-based access so only authorised staff can see sensitive notes or approve merges. Confirm data retention, deletion and candidate correction routes with the CRM vendor and your adviser. When a candidate says a field is wrong, the team should be able to correct it and see which summaries or messages used the old value. When a person opts out, a queued AI draft must not override that preference. The AI hiring bias checklist belongs to any later screening discussion, rather than the record-cleaning pilot.
At the end of the trial, decide separately on parsing, duplicate review and follow-up drafts. One may work while another needs better inputs or stronger permissions. The agency has succeeded when a recruiter can open a record, see the source of each important fact, understand why a message is due and explain why two profiles were or were not merged. That standard is more valuable than a headline about how many records AI touched.
Questions about candidate data housekeeping
Should a CRM merge candidates automatically?
Begin with a suggested-merge queue. Matching names or email addresses can conceal different people, shared mailboxes or old contact details. A recruiter should compare identifiers, consent records, notes, applications and document versions before approving a merge. Test the vendor's restore or audit options before enabling any automatic merge setting.
Can AI write all candidate follow-ups?
It can prepare drafts from verified CRM events and a recruiter-approved message rule. A person should check the role, status, previous promises, contact permission and tone before the message is sent. Replies, opt-outs, complaints and sensitive questions need a human route. Do not use a guessed suitability score to decide who receives a reply.
Does this replace candidate screening?
No. The workflow here is administrative: parsing records, detecting possible duplicates, finding missing fields and queuing promised contact. Assessing suitability, ranking candidates and making selection decisions are separate tasks with different evidence, fairness and legal requirements. Keep the administration pilot away from those decisions.
Further reads
- AI Agents for Recruitment Agencies: Sourcing, Screening and Review Points — Set review points for broader agency automation.
- Resume Formatting AI Tools for Recruiters: Check Nothing Was Altered — Preserve facts when preparing a presentable CV.
- How Recruiters Use AI to Write Candidate Summaries Clients Read — Prepare client summaries from checked records.
- Recruitment Agency KPIs: A Weekly Scorecard From Your CRM Export — Measure outcomes from a clean CRM export.
- AI for Recruitment Agencies: Re-Engaging Your Candidate Database — Plan later contact with dormant records.
- Free AI Tool for Recruitment: What a Small Agency Gets for Nothing — What free AI plans and notetakers cover for a small recruitment agency, where free limits and candidate-data risks bite, and when paying makes sense.
- AI Boolean Search Generator for Recruiters: Prompts and Checks — How to brief an AI for a Boolean search string, what a hiring manager should review, and how to test the string on three sites and log false positives.
- ICO AI Recruitment Report: A Checklist for UK Recruitment Agencies — An agency-facing checklist that turns the ICO recruitment findings into evidence, owners and review points.
- Job Intake Form for Recruiters: Fill It From the Call With AI — A call-to-form method with a filled example, extraction prompt and confirmation email for ambiguous role requirements.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Bullhorn ATS Automatic Resume Parser help page; Loxo Data Agents page; Manatal candidate sourcing feature page (all checked 4 November 2026).