Is AI CV Screening Accurate? Evidence & Limits

AI CV screening can help teams apply a reviewed screening rubric across a batch of candidates, but there is no single accuracy percentage that applies to every tool or role. Accuracy depends on the criteria, available CV evidence, document extraction, and how the team checks recommendations. Test the tool on your own hiring context and inspect disagreements before relying on its shortlist.

Marxel brings job listings, sourcing, applications, and explainable screening into one hiring workspace. For screening, it turns a job description and briefing notes into a user-reviewable rubric, then shows candidate-level reasoning and match/miss breakdowns. Those features make recommendations inspectable; they do not establish an accuracy rate or guarantee a hiring outcome.

What Does "Accurate" Mean in CV Screening?

Separate these questions when comparing tools:

Extraction: Did the tool read dates, skills, qualifications, and experience correctly?
Evidence: Does each assessment point to something the candidate actually supplied?
Criteria application: Did the tool assess the reviewed requirements rather than invent new ones?
Shortlist quality: Did it overlook someone the team would advance, or advance someone without the required evidence?
Repeatability: Do repeated runs with the same inputs produce materially similar assessments?

Agreement with a recruiter is one signal, but neither the recruiter nor the tool is automatically correct. A shared mistake can produce high agreement. Resolve disagreements against the role criteria and candidate evidence.

How to Test AI CV Screening on One Role

Use the one-role test before a wider rollout.

Choose a representative role and record the job description plus hiring-manager briefing notes.
Review the Screening Rubric before processing CVs. Remove vague requirements and check that each criterion is relevant to the role.
Have a reviewer assess the same candidates independently, using the same rubric and documenting the evidence.
Run the AI screening and compare recommendations. Include candidates outside the suggested shortlist.
Ask another reviewer to resolve disagreements against the evidence. Record unresolved cases as unresolved.
Repeat a small subset with unchanged inputs to check whether important assessments change.

For an initial workflow check, use synthetic CVs with deliberately varied evidence. Include a clear match, a confirmed gap, missing information, an unusual career path, and a difficult document format. Synthetic examples test whether the workflow is understandable; they do not prove performance on real applications.

Metrics Worth Recording

Metric	Calculation or check	What it reveals
Missed-candidate rate	Candidates the resolved review would advance but the AI would not, divided by all candidates the resolved review would advance	Qualified candidates overlooked in this sample
Unsupported-advance rate	AI-advanced candidates the resolved review would not advance, divided by all AI-advanced candidates	Shortlist entries needing closer inspection
Evidence errors	Count invented, misread, or incorrectly attributed facts	Whether explanations can be trusted
Unresolved cases	Count cases where the evidence cannot support a decision	How much follow-up is still needed
Repeatability	Compare important assessments across unchanged reruns	Whether recommendations change without new evidence
Time to reviewed shortlist	Include setup, processing, evidence checking, corrections, and follow-up	Whether the complete workflow saves effort

If a denominator is zero, report the rate as unavailable. Retain the sample size and role context. Ten easy examples cannot establish accuracy across different roles, languages, or CV formats. Agree acceptance criteria before the test; this guide does not prescribe a universal pass percentage.

Where Errors Can Enter the Workflow

Incomplete or Misread CVs

Scanned documents, complex layouts, overlapping dates, and abbreviated skills can change what the tool reads. Check the original CV when extracted evidence appears incomplete. A missing statement is not proof that the candidate lacks the skill.

Poorly Defined Criteria

Requirements such as "top university," "culture fit," or unexplained career-gap penalties can introduce irrelevant assumptions. Review what each criterion measures and whether the CV can provide evidence for it. Use the fair screening criteria guide to structure that review.

Unsupported Inferences

Listing JavaScript experience does not establish expertise in every web framework. A leadership title does not demonstrate a particular management style. Check whether the explanation distinguishes explicit evidence, inference, and missing information.

Unusual Career Paths

Career changers and returning professionals may describe relevant experience differently from the job description. Inspect transferable evidence and unresolved requirements. Do not treat a non-linear career as a reason to reject someone.

Changing Inputs or Models

Changing the rubric, document parser, or model can change results. Record inputs and available configuration so you can distinguish an improvement from a different test. Recheck important cases when the workflow changes.

What to Check in Marxel

Marxel groups screening results into Aligned, Potential, Hold, or Unclear and provides reasoning for team review. Inspect the evidence and match/miss breakdowns for each candidate, including candidates outside the strongest-match group. Use follow-up questions where information is missing, and keep the hiring decision with the team.

It is a useful option to evaluate when your team needs shared criteria and visible reasoning across a batch. Teams hiring very infrequently may find a documented manual scorecard sufficient. Teams seeking fully autonomous hiring decisions should reconsider their process before choosing a screening tool.

Before uploading real candidate information, review the product's current security information, your organisation's data-handling requirements, and pricing and plan limits. Product explanations and general governance guidance do not establish accuracy for your deployment.

Frequently Asked Questions

What accuracy percentage should I expect?

There is no defensible universal percentage. Ask a vendor for its dataset, role mix, definition of correctness, sample size, and evaluation method. For Marxel, this guide does not claim a validated accuracy percentage. Measure role-specific outcomes before expanding use.

Is consistent screening the same as accurate screening?

No. A system can consistently apply a poor criterion or repeatedly misread a CV. Check the criteria and evidence as well as repeatability.

How do I know whether AI is missing good candidates?

Review a sample outside the suggested shortlist and compare it with an independent assessment against the same rubric. Resolve disagreements using CV and follow-up evidence. Reviewing only shortlisted candidates cannot reveal people omitted.

Does an explanation prove the recommendation is right?

No. An explanation makes a recommendation easier to inspect, but it can still contain unsupported inferences or misread facts. Check it against the original evidence.

How should we decide whether to continue?

Continue when the test meets your team's acceptance criteria for evidence quality, missed candidates, review effort, and unresolved cases. Correct the criteria or pause if important explanations cannot be supported. A successful synthetic demonstration is a reason to try a controlled role-specific evaluation, not to skip it.

Evaluation Guidance

The NIST AI Risk Management Framework provides general guidance for incorporating trustworthiness into AI evaluation and use. It is not evidence of Marxel's screening accuracy.