AI CV matching compares candidate information with a vacancy’s requirements and helps recruiters identify evidence worth reviewing. A useful system brings relevant experience, missing information and follow-up questions into one place. The recruiter remains responsible for deciding who progresses.
For a recruitment agency, the business question is whether this produces a credible shortlist with less total effort. A fast ranking that takes hours to check can make the process worse.
This guide explains Binarify’s proposed approach to CV and resume matching, recruiter review, ATS integration and evaluation. The examples are illustrative, not measured client results. See our AI consultancy for recruitment agencies for the wider implementation approach.
What matching should do in your recruitment workflow
Three activities often get grouped together:
- Parsing: extracting information from CVs and attaching it to the correct candidate record.
- Matching: comparing that evidence with the agreed role criteria and finding potentially relevant candidates.
- Shortlisting: reviewing suitability, resolving important gaps and deciding who should move forward.
AI can assist with the first two and prepare material for the third. A similarity score alone does not confirm a qualification, current availability or willingness to take a role. Do not present an unvalidated score as the probability of a successful hire.
This guide focuses on evaluating candidates for a specific vacancy. If your main problem is finding people hidden in old records, start with our candidate rediscovery guide.
Check your existing ATS before commissioning a new tool
Your applicant tracking system may already provide parsing, candidate search or matching. Test what your agency can actually use before paying to reproduce it.
For example, LinkedIn’s applicant-review documentation describes qualification summaries and recruiter review, with applicant evaluations available inside the ATS for users with Apply Connect. Access depends on the product, permissions and rollout.
Bullhorn’s Matching Engine documentation describes matching within automation and Search & Match, including recognising variations in job titles and skills. Its documentation specifies edition and add-on requirements.
These are examples of available capabilities, not evidence that either product will meet your agency’s requirements. Compare three options: configure what you own, connect an existing capability to your workflow, or build the missing part. Ask a vendor to demonstrate your difficult cases as well as straightforward matches.
A practical CV matching and shortlisting workflow
1. Agree the vacancy criteria before processing CVs
Have the consultant and hiring manager distinguish essential requirements, preferences and points that need clarification. Record what evidence would satisfy each criterion.
For an accounts role, preparing month-end reconciliations may be essential while experience with a particular accounting package is preferred. A matching system should not turn that software preference into an automatic exclusion.
Keep the approved criteria with a version and date. If the client changes the role, identify which candidates need another review. Otherwise two consultants can assess the same person against different requirements without realising it.
2. Read the right documents and flag extraction problems
Connect each CV to a stable candidate identifier and its application or vacancy. Check duplicate records, document dates and whether a newer CV has replaced an older version.
Scanned documents, tables and unusual layouts can produce incomplete text. Route unreadable or uncertain documents for review instead of treating a parsing failure as a lack of experience. Preserve the original document so the consultant can check extracted statements.
Treat CV content as evidence to analyse. Instructions embedded in a document must not change the matching rules or trigger actions in connected systems.
3. Build an evidence view for each requirement
Show what supports a suggested match, where that information came from and what is unresolved. Use clear states such as evidence found, needs verification and not stated.
Illustrative example: a candidate’s CV describes preparing reconciliations and supporting a monthly close. It mentions one accounting package but says nothing about the client’s preferred package or current availability.
| Requirement | Evidence available | Recruiter’s next step |
|---|---|---|
| Month-end reconciliation work | A passage describes preparing reconciliations | Check ownership, complexity and recency |
| Preferred accounting software | A different package is named | Establish whether transferable experience is sufficient |
| Availability | Not stated | Ask the candidate before making a commitment |
Do not infer that an unstated skill is absent. Equally, mentioning a tool once is not enough to claim proficiency. The explanation should stay within the source evidence.
4. Give the recruiter a real review step
The reviewer needs access to the original CV, the approved criteria and the reason each candidate was surfaced. They should be able to correct extraction errors, ask questions, change the recommendation and record a reason.
Review candidates outside the highest-ranked group as part of evaluation. If people below a score threshold are never examined, relevant candidates can disappear from the process while the visible shortlist looks convincing.
Our proposed starting scope does not automatically reject applicants or make final hiring decisions. Human review must include the time and authority to disagree with the system.
5. Confirm material gaps before client submission
Prepare focused questions about unresolved requirements, interest and availability. After a recruiter checks the answers, create a draft submission summary using confirmed information.
Separate candidate statements from independently verified facts. The consultant approves the final client-facing submission, including any qualification or experience claims. This helps prevent a polished AI summary from overstating what the candidate has actually demonstrated.
How ATS integration should work
Keep the ATS as the operational record. A separate matching dashboard is useful only if its results return to the vacancy, candidate and consultant responsible for the next action.
Before implementation, establish:
- Read access: which vacancy fields, applications, CV attachments and notes the integration can access under the agency’s permissions.
- Write access: where proposed matches, evidence, review decisions and follow-up tasks belong. Avoid overwriting original CVs or recruiter notes.
- Record identity: how candidate and vacancy identifiers prevent duplicates or results being attached to the wrong person.
- Updates: what happens when a CV changes, a job closes or criteria are revised. Mark old results as stale until reviewed again.
- Failures: who sees a failed document read or write-back, how it is retried and how repeated attempts avoid creating duplicate records.
- Access and deletion: how brand restrictions, retention rules and corrections carry through to any copied documents or search index.
Start with read-only access and a reviewable output where practical. Enable limited write-back after testing the mapping and recovery process. The available API, subscription, vendor permissions and data quality determine what can be integrated; a connection should not be promised before those checks.
Measure shortlist quality as well as speed
Define the evaluation before choosing a score threshold. Use a representative sample of vacancies and authorised candidate records, including different CV formats, related job titles, incomplete information and plausible but unsuitable matches.
Have experienced recruiters independently assess the sample against the same criteria, resolve disagreements and preserve their reasons. Historical placements can provide useful examples, but they do not identify every suitable candidate who was overlooked.
Compare the current workflow with the assisted workflow on:
| Measure | What to record |
|---|---|
| Review precision | Candidates judged relevant divided by candidates suggested for review |
| Relevant candidates found | Independently identified relevant candidates retrieved, divided by all relevant candidates in the evaluated sample |
| Unsupported claims | Factual claims in summaries that lack supporting source evidence |
| Total review effort | Time spent reading, checking, correcting and preparing the approved shortlist |
| Shortlist turnaround | Elapsed time from agreed vacancy criteria to an approved shortlist |
| Submission-to-interview conversion | Interview invitations divided by client submissions, with a consistent time window |
The second measure estimates recall within the evaluated sample. It does not establish how many suitable people exist across an entire database or the labour market.
For example, suppose an independently reviewed sample contains 12 relevant candidates. A system suggests 10, of whom eight are relevant. Its review precision is 8/10, or 80%, and its recall in that sample is 8/12, or about 67%. The four missed candidates matter even though most suggestions look reasonable. These numbers illustrate the calculation; they are not performance targets or Binarify results.
Keep a separate evaluation sample for checking changes. Compare results by role type and document format, and investigate unequal error patterns with appropriately governed data. Do not infer sensitive characteristics from names or photographs to create those comparisons.
Track downstream outcomes over time, but account for differences in vacancies, client responsiveness and market demand. Faster shortlisting alone does not prove more placements.
Candidate information and accountable decisions
The UK ICO’s 2024 recruitment-AI audits found problems involving fairness, excessive collection and transparency. Those findings make data handling and review part of implementation, rather than an afterthought.
Establish the applicable requirements for the countries and roles involved. Document the purpose of processing, information needed, access, retention, candidate notices and routes for correction. This guide is an implementation approach, not a claim of compliance across jurisdictions.
Choose a first project with a measurable business case
Start with one recruitment desk or a repeatable role family. Record its current workload and total shortlist preparation effort, then agree quality and operating-cost limits for a pilot.
Estimate capacity released using observed time savings after review and correction. Include software, integration, support and ongoing evaluation costs. Recovered time has business value when the agency can use it for candidate conversations, client service or additional work; it is not automatically a cash saving.
Binarify’s AI Impact Diagnostic assesses the workflow, existing software and baseline before recommending an implementation. Explore our recruitment AI consultancy or book a conversation to discuss where matching and shortlisting are slowing your agency down.