AI Candidate Screening: Criteria, Evidence, and Human Review
Use AI to organize application evidence against approved job criteria. Keep people responsible for screening decisions, accessibility, and checking unequal outcomes.
Updated 5 min read
AI candidate screening uses software to help organize application evidence against job requirements. A supervised workflow starts with approved, job-related criteria, checks the underlying evidence, and gives people responsibility for decisions about applicants.
The useful output is a traceable review, not an unexplained score. Keep applicant records in an approved hiring system. Tools like fenbs can separately preserve the evaluation rules and reusable instructions so different assistants do not apply different versions of the process.
How do you set up a screening workflow?
- Have the hiring manager and appropriate HR reviewers approve the actual duties and evidence needed for the role.
- Separate essential requirements from preferences, and write what counts as relevant evidence for each.
- Approve the data, tools, access, and accommodation process before using applicant information.
- Ask AI to extract evidence and flag missing information, with references to the application source.
- Have a reviewer check the source, resolve uncertain cases, and record the reason for each screening decision.
Begin with a limited evaluation using permitted test material. Include different document layouts, equivalent qualifications described in different words, and incomplete applications. The aim is to learn where extraction fails before a missed phrase affects someone’s opportunity.
Candidate discovery is a separate job: the candidate sourcing workflow covers finding potential applicants. This workflow begins when there is an application to assess against an agreed role.
What makes a screening criterion defensible?
Connect each criterion to work the person will actually perform. Replace vague requests for a particular personality or background with observable evidence. For a support role, handling a written customer problem is more useful to discuss than an assistant’s impression that a resume sounds confident.
The EEOC guidance on tests and selection procedures (opens in a new tab) explains that selection methods can violate discrimination laws through intentional discrimination or unjustified disproportionate exclusion. It also says employers remain responsible for appropriate validation even when a vendor supplies supporting documentation.
| Criterion | Evidence to examine | If the evidence is missing |
|---|---|---|
| Explains a troubleshooting sequence | A permitted work sample showing clear steps | Request the same relevant sample through the agreed process |
| Handles a required support channel | Application example or work sample for that channel | Flag for reviewer clarification |
| Meets a genuinely essential credential requirement | The stated credential and the agreed verification process | Ask a reviewer; do not infer from a school or name |
| Communicates the next action clearly | Specific wording in a structured work sample | Record insufficient evidence rather than inventing a score |
Define equivalent evidence before reviewing applications. Avoid changing the standard after seeing a favored applicant. If a requirement changes, decide how to review affected applications consistently and record who approved the change.
In fenbs, Decisions and rules can preserve that approved rubric and the reason for a revision without storing candidate files in the project record.
What should the AI output contain?
Ask for one row per criterion: source location, relevant passage, and whether the evidence is present, absent, or unclear. An empty field should stay empty. A resume that omits a tool does not establish that the applicant has never used it.
Require reviewers to open the source, especially when the assistant proposes that evidence is missing. A document parser may skip a column, misread a date, or lose a heading. Check the extraction step before using a summary as a basis for a decision.
Do not ask the assistant to infer protected characteristics, health, personality, or suitability from names, photographs, voices, or unrelated personal details. Keep any authorized compliance monitoring separate from the screening prompt and ordinary reviewer access. Agree that separation with your HR and legal reviewers.
Keep the reusable extraction instructions with their version and approved purpose. An AI brain is useful here when it preserves process knowledge rather than collecting unnecessary personal information.
How do you check accessibility and unequal outcomes?
Test whether applicants can use the process, not just whether the software returns results. Provide a clear route to request an accommodation and a person who can resolve access problems. Decide how alternative evidence will be reviewed without treating the request itself as a negative signal.
The Justice Department’s guidance on AI and disability discrimination (opens in a new tab) explains that hiring technology can unlawfully screen out qualified people with disabilities, including when a third party supplies it. It discusses evaluating tools and providing reasonable accommodations. A generic claim that a tool is objective does not answer those questions.
Have qualified reviewers examine selection outcomes and error patterns using appropriately controlled data. Hiding demographic fields from a prompt is not proof that the process avoids bias. Test false exclusions, inconsistent treatment of equivalent evidence, and whether a less discriminatory approach can meet the job’s needs.
Which approvals should come before real use?
Ask HR and legal reviewers to determine which federal, state, and local requirements apply to the actual tool and use. For covered uses, New York City’s automated employment decision tool rules (opens in a new tab) require a recent bias audit, public information about that audit, and specified notices. Do not assume every AI feature has the same legal classification.
Document the approved purpose, who may use the tool, how long records are retained, and what triggers another review. A model change, new ranking feature, or revised job criterion may change the process you evaluated. Human review must include time and authority to challenge the output.
Use an AI approval workflow to distinguish proposed process changes from decisions a person has approved. An assistant can help prepare the evidence; it cannot establish legal compliance by declaring the workflow fair.
Keep one approved screening process
Use fenbs for the job-independent rules, the approved rubric, and the questions waiting on HR or legal review. Store reusable extraction instructions in Skills and prompts so connected assistants can read the same version. Keep applicant evidence and hiring decisions about individuals in your approved hiring system.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account and add a project for improving the screening process. It is free for one person, no credit card.
- Record the rubric and review requirements in Decisions and rules, naming the people who approved them. Leave unresolved choices open.
- Add the approved evidence-extraction instructions to Skills and prompts, using fictional examples instead of applicant details.
- Create tasks for extraction checks, accessibility review, and unresolved compliance questions, then record the actual results.
- Connect a supported assistant using Integrations. Have it read the rules before revising the process and save the next step before stopping.
fenbs does not screen resumes or replace an applicant tracking system, and it is not the right fit for storing applicant dossiers. Use it when the problem is maintaining the approved process across people and assistants. Explore other supervised workflows in AI at work.