Your hiring AI rejected them. Can anyone explain why?

A hiring team should understand what its screening system evaluates, how the evidence relates to the role and how a person can correct an error. Adding a human at the end does not automatically create meaningful oversight. That person needs information, time and authority to disagree with the recommendation.

This is an operating guide for hiring teams, not a legal opinion about a particular tool. As checked on 20 October 2026, the European Commission's current AI Act page says the rules for high-risk employment use cases apply from 2 December 2027 following the AI Omnibus changes. The official implementation timeline distinguishes other dates and obligations. Do not assume either that everything started in August 2026 or that no present obligation exists before the employment deadline.

Start with the decision the software actually makes

Does the tool organise applications, summarise evidence, rank candidates or automatically reject them? These are different functions. Find out what happens in the configured workflow, rather than relying on the product's general description or a reassuring diagram from procurement.

Map the inputs, outputs and consequences. Which information enters the system? Does a score change who receives human attention? Is a missing field treated as disqualification? Can an interviewer see the original application or only a generated summary? A nominally advisory score can still determine outcomes when nobody looks below the first page.

Assign an owner for each important decision. The supplier may explain the model; the employer still needs to explain the process it chose to use. Responsibility should not disappear into the gap between HR, IT and a contract appendix.

Make each criterion earn its place

Tie screening criteria to actual job requirements. Distinguish essential eligibility from preferences and learnable skills. A keyword is a signal about wording, not direct proof of capability. Someone may have relevant experience under a different title, language or organisational vocabulary.

Review potentially unjustified proxies, such as gaps, school prestige or historical patterns in previous hires. Ask what a criterion is supposed to establish and whether a more direct assessment is available. This requires qualified fairness and legal review where appropriate, not a manager declaring that a score is objective because it is numeric.

The CIPD selection factsheet supports clear, structured assessment. Applied here, the practical discipline is to preserve job-related reasoning from application through interview. The tool should help organise evidence for that reasoning rather than silently replace it.

Test errors, including the people you would otherwise never see

Review a representative set of cases before relying on a screen. Include varied but relevant ways of expressing experience, incomplete data, unfamiliar job titles and applications that require adjustments. Check original evidence against summaries. Test what happens when a field is missing or a document is parsed incorrectly.

Examine false rejections as well as unsuitable candidates who passed. If only shortlisted people receive review, the organisation can learn about false positives while remaining blind to people it screened out incorrectly. Design a proportionate sampling and correction process with the relevant specialists.

Use synthetic or appropriately governed test data where possible. Do not create a new privacy problem by copying real applications into unapproved tools for a fairness experiment. Record the limitations of the test: a few successful examples are not proof of validity for every applicant group or role.

Give reviewers enough context to challenge a recommendation

Show the relevant evidence, the role criteria and the reason for the recommendation at a level the reviewer can understand. Explain uncertainties and known limitations. A red score with no usable explanation encourages either blind acceptance or arbitrary override.

Measure the reviewer on decision quality, not just the speed of clearing the queue. If challenging a recommendation creates extra administrative work while approval takes one click, the workflow has a strong opinion about what “human oversight” should mean.

Greenhouse's responsible AI principles are one supplier's published approach to this problem. Use such statements to ask precise procurement questions, then verify the configuration and commitments that apply to your own use. A principles page does not independently certify your hiring process.

Explain the process to candidates

Tell applicants what technology is used and provide information and routes required by the relevant rules. Make it possible to report an incorrect application record, request an adjustment or contact a responsible person. A support route that nobody monitors is an attractive footer link with limited practical value.

Avoid promising a detailed individual explanation that the team cannot provide. Decide what information can responsibly be shared and who handles requests. Where an automated decision has legal significance, get qualified advice on the applicable rights and obligations, including data protection and employment requirements.

Keep records that support the actual purpose, with appropriate access and retention limits. More retained data does not automatically mean better accountability. A useful record connects the criteria, evidence, decision and relevant human review.

Review changes as well as launch day

A new model version, altered job requirement or changed rejection threshold can affect the system's behaviour. Define which changes trigger a fresh assessment and who approves them. An earlier review should not become a permanent permission slip for a different workflow.

Track corrections, complaints, review outcomes and unexplained changes. Continue investigating signals rather than declaring a tool fair because the complaint count is low. Candidates may never know what happened or where to report a problem.

For a better assessment foundation, see skills-based hiring. For related concerns, read candidate verification and AI accountability. I help connect product decisions and real operating workflows. Bring the hiring step nobody can currently explain.

Questions people ask

Does a human approving an AI score count as meaningful oversight?

Only when the person can understand the relevant evidence, recognise limitations and change or stop the decision where appropriate. Review time, training and incentives matter. A hurried approval step with no usable explanation can become a rubber stamp rather than an effective control.

When do EU AI Act high-risk employment rules apply?

As checked on 20 October 2026, the European Commission’s updated timeline gives 2 December 2027 for high-risk employment use cases following the AI Omnibus changes. Other AI Act provisions and other laws have different dates and scope. Confirm the classification and current obligations for your specific tool with qualified advisers.

What should employers ask an AI screening supplier?

Ask what the system does, which inputs and criteria affect recommendations, how errors are tested, what a reviewer can see and change, and how updates are governed. Clarify candidate information, correction routes, data handling and contractual responsibilities. Verify the configured workflow rather than relying only on general marketing claims.