Why should an AI hiring score not automatically reject job candidates?
Use an AI hiring score to prioritise review, not as an automatic rejection threshold. A composite score compresses job requirements, missing information, model uncertainty and historical patterns into one number. It does not tell a recruiter whether a person lacks an essential qualification, omitted a detail from the CV, used unfamiliar wording or was evaluated against a questionable proxy. The first release should help people parse, search, compare evidence and ask better questions. An authorised person should own the employment decision, its reason and any reconsideration.
Wavesteam's public AI recruiting system case study illustrates product capabilities such as CV parsing, talent search, screening support and the organisation of role and candidate information. It demonstrates a relevant delivery context; it does not prove that one model score can safely decide who loses an opportunity, and its project metrics are not acceptance targets for another employer.
What a single score hides
Suppose three applicants receive 82, 76 and 69. That ordering may combine very different signals:
- a licence, language or shift requirement explicitly stated for the role;
- inferred experience from job titles, projects and CV wording;
- correlations with schools, employers, location or career gaps;
- fields lost when the document parser misread a layout;
- patterns inherited from previous hiring decisions; and
- a model output produced despite weak or contradictory evidence.
These signals have different reliability, business meaning and correction paths. Adding them together creates false precision. A more useful interface separates verified essentials, evidence-backed capability signals, questions to ask, missing information and parsing failures. Every conclusion should point to the original application, the current job rule or a recorded reviewer decision.
The employer retains legal and product responsibility
China's Personal Information Protection Law requires transparency and fair, impartial outcomes when personal information is used for automated decision-making. Where an automated decision has a significant impact on a person's rights and interests, the person may request an explanation and refuse a decision made solely through automation. The law also requires a prior personal-information protection impact assessment for automated decision-making.
China's Employment Promotion Law requires employers and employment intermediaries to provide equal opportunities and fair conditions and prohibits employment discrimination. A Ministry of Human Resources and Social Security notice on recruitment information further says that criteria should relate to job duties and that the publisher remains responsible for review.
These sources do not mean that every AI recruitment feature is prohibited. They mean that buying a model or SaaS product does not transfer the employer's responsibility. Counsel, HR and the organisation's privacy owner should assess the actual role, data, workflow and applicable rules. This article is a product-scope and acceptance framework, not legal advice.
Convert the role into reviewable criteria first
Before configuring a model, divide requirements into four groups:
| Requirement | Appropriate system role | Human decision |
|---|---|---|
| Mandatory licence or qualification | Locate the claimed certificate, identifier and validity evidence | Confirm that it is genuinely required and verify the evidence |
| Demonstrable capability | Extract relevant projects, duties and outcomes | Decide whether the evidence maps to the work and whether an interview or test is needed |
| Trainable capability | Surface adjacent experience and learning evidence | Decide whether the ramp-up is acceptable; do not reject on a missing keyword |
| Attribute that should not drive selection | Exclude it from scoring or conceal it during the relevant review stage | Confirm whether it is unrelated to the inherent requirements of the role |
The job description, screening rules and interview rubric must use the same definitions. A manager's preference for a particular school, age range or uninterrupted career history must not silently become a model weight. A proposed new hard criterion needs an owner, a job-related rationale, an approver, an effective date and an identified set of affected applications.
Rank evidence; do not automate rejection
A sensible first release can use AI to:
- parse CVs into structured contact, education, employment and project fields;
- expand search with synonyms and adjacent skills so a single keyword does not decide visibility;
- display each role requirement beside the relevant source passage;
- create a review task for missing, contradictory or low-confidence information; and
- draft interview questions from confirmed evidence for a recruiter to select and edit.
A rejection action should require an accountable person to choose an intelligible business reason, such as a verified absence of a legally required qualification, the applicant's confirmed inability to meet an essential work arrangement, or failure to perform a critical task in a structured assessment. “Low model score”, “poor fit” or “other applicants ranked higher” is not enough. Bulk processing should not bypass the evidence requirement.
Human review must be capable of changing the outcome
If the system removes 90% of applicants before a recruiter can see them, a final confirmation click is not meaningful oversight. A workable human-AI configuration should ensure that:
- reviewers can inspect the original material, extracted fields, rule version and uncertainty;
- higher-impact actions require appropriate permission and cannot be bulk-approved by an ordinary operations account;
- audits sample high-ranked, low-ranked and parsing-failure groups rather than only successful applicants;
- candidates can provide or correct information and the organisation can locate every affected recommendation;
- model, prompt or parser upgrades are tested against a fixed regression set before release; and
- overrides are classified as model error, rule error, missing evidence or a legitimate judgement call.
The NIST AI Risk Management Framework calls for defined human-AI roles, documented knowledge limits, validation in the intended context and continuing risk monitoring. NIST is not Chinese employment law, but its governance method is useful: a human reviewer needs the information, authority and time to disagree, rather than merely endorsing an automated result.
Acceptance must examine whom the workflow misses
The evaluation set should cover multiple roles, document formats, career paths, locations and levels of completeness, including genuine positive and negative cases. If historical hiring outcomes are used as labels, first assess whether those outcomes are an appropriate reference. Past preference should not automatically become future policy.
Measure more than time saved:
| Measure | Decision it supports |
|---|---|
| Critical-field extraction accuracy | Whether dates, employers, roles, projects and qualifications were read correctly |
| False-negative rate among qualified applicants | Whether suitable people disappear below the visible review queue |
| Outcome differences across relevant groups | Whether a difference can be explained by a legitimate job requirement |
| Human override rate and reason | Whether the failure is in the model, the job rule or the available information |
| Correction resolution time | Whether amended information updates the related recommendation and decision |
| Cost per valid review | Whether saved screening time outweighs appeals, rework and missed-candidate cost |
The U.S. Equal Employment Opportunity Commission's guidance on employment tests and selection procedures says that selection tools should be validated for the specific job and purpose and updated as requirements change. It applies in the United States and does not replace Chinese requirements. The reusable product lesson is that a supplier's generic accuracy claim does not show that a tool predicts performance for this employer's particular role.
A defensible first release
Start with one role family and one recruiting team. Deliver versioned job criteria, CV parsing, evidence-backed search, a review queue, a structured interview rubric, reason codes, correction handling, permissions, audit logs and a monitoring dashboard. Evaluate de-identified historical samples offline, then run in shadow mode where recommendations do not affect real applicants. Only introduce the tool into the live workflow after the team understands false negatives, outcome differences and the workload of genuine review.
Wavesteam can connect AI to the recruitment workflow, data, permissions and evaluation layer. The client's HR owner must retain the job truth and final employment authority, while legal and privacy owners confirm the processing boundary. The product goal is not to help an employer say “no” faster. It is to make screening better evidenced, reviewable and correctable.
Sources
- Wavesteam: AI recruiting system case study: public evidence of recruitment workflow and application-processing experience, not proof of autonomous rejection or universal results.
- Cyberspace Administration of China: Personal Information Protection Law: automated-decision transparency, explanation, refusal and impact-assessment requirements.
- Ministry of Human Resources and Social Security: Employment Promotion Law: equal employment opportunity and anti-discrimination principles.
- MHRSS notice on regulating recruitment information and promoting equal employment: job-related criteria and review responsibility.
- NIST: AI Risk Management Framework Core: human-AI roles, intended-use validation, knowledge limits and monitoring.
- U.S. EEOC: Employment Tests and Selection Procedures: role-specific validation and continuing review of selection tools; U.S. scope.
Sources accessed and article reviewed on 24 October 2026.