Use AI in recruitment for administrative work such as scheduling, transcription and extracting details from documents, and keep people accountable for consequential decisions. Ranking, assessing or rejecting candidates needs validation, explanation and real human review.
- Use AI to reduce administrative work and organise evidence; keep accountability for consequential decisions with people.
- Scheduling, transcription and document extraction are lower-risk; ranking, assessment and rejection need validation, explanation and meaningful human review.
- Inferring personality or emotion from video or voice is rarely justified.
- Tell candidates which stage uses AI, what it does and how to correct or challenge a result.
- Measure each stage against your current process, monitor after launch and set pause conditions in advance.
The useful question is what task a system performs, how its output changes a candidate's opportunity and whether a qualified person can catch an error before harm occurs.
Map the task before buying a system
"AI recruitment" covers tools with very different consequences. Classify the task and its effect on candidates before comparing features.
| Use | Plausible value | Main risk | Default human control |
|---|---|---|---|
| Interview scheduling | Less coordination work | Accessibility, time-zone or calendar errors | Easy manual booking and support route |
| Resume parsing | Extract roles, dates, education and skills | Missing or mis-structured evidence | Keep the source document and allow correction |
| Job-description assistance | Draft or review role language | Invented requirements or exclusionary wording | Role owner approves every requirement |
| Candidate search | Find potentially relevant profiles | Proxy discrimination, narrow queries | Recruiter reviews criteria and excluded groups |
| Matching or ranking | Prioritise evidence against stated requirements | Unsupported scores, historical and automation bias | No unreviewed rejection; source evidence visible |
| Interview transcription or summary | Shared record, less note-taking | Errors, sensitive data, lost context | Interviewer verifies against notes or recording |
| Skills assessment | Structure or score a job-related task | Invalid measures, accessibility barriers | Validate for the role; offer adjustments |
| Video, voice or emotion inference | Claimed personality signals | Weak validity, disability and language bias | Do not deploy without compelling independent evidence and legal review |
The same component can carry different risk: extraction that pre-fills a form differs from extraction feeding an automatic rejection. The UK government's guide asks whether AI suits the problem at all (UK responsible AI in recruitment guidance).
Parsing and human review
A parser turns a resume into fields. A missing field means "not reliably extracted," which is different from "the candidate lacks it." Keep the original resume beside extracted data, label each requirement as present, not found or uncertain, and test representative resumes across layouts, career histories and languages. Candidates can reduce formatting risk (see the guide to an ATS-friendly resume), but they should not carry the whole burden for parser failures. If the tool also ranks, document the criteria, weights, thresholds and what happens to lower-ranked applicants.
A person placed after a score does not automatically provide oversight. Meaningful review needs a reviewer who understands the role and the system's limits, access to source evidence, time to question the output, authority to override and escalate, consistent treatment of candidates at the same stage, and monitoring of overrides.
Illustrative case: a parser reads "Member of Technical Staff," a matcher expects "Software Engineer," and the candidate scores low despite relevant API and production-incident work. A reviewer who sees only the score confirms the error; one who sees the requirement and source passages corrects it.
The UK Information Commissioner's Office reported in 2026 that human involvement must be applied consistently within a hiring stage and that many employers appeared to rely on solely automated decisions without adequate safeguards (ICO Recruitment Rewired).
Tell candidates what the system does
A useful notice states which stage uses AI, the task and output, the data used, how the output affects progression, whether a person reviews it, how to request an adjustment or correction, retention, and a contact who can respond. An illustrative example:
We use software to extract employment dates, roles and skills from your resume so recruiters can review applications consistently. It does not make the final hiring decision. You can review the fields during application and contact recruitment@example.org to correct extraction errors or request an alternative application route.
Never publish a template that understates the real automation. Recourse should match the consequence: a scheduling error needs quick support, while a disputed rejection needs preserved records and review by someone not bound to the original output. Disclosing the selection process in the job post itself, as employers do on hiARed, tells candidates what to expect before they apply.
Test fairness across the whole process
Models can reproduce historical patterns, use proxies for protected characteristics or perform differently across groups. Instead of accepting one vendor fairness ratio, ask:
- Which decision and population was the system validated for, with what sample sizes?
- Do extraction, scoring or error rates differ across relevant and intersectional groups?
- Are the job requirements necessary and consistently applied?
- What happens to candidates below a threshold or outside the training data?
- Are accessibility adjustments and non-AI routes genuinely equivalent?
- Who investigates, pauses or rolls back the system after an update changes outcomes?
Measure stages separately: completion, extraction accuracy, search inclusion, assessment, progression, overrides and final outcomes. An overall hiring rate can hide an early barrier. Testing for unequal outcomes may need carefully governed demographic data that decision-makers cannot see; the right approach depends on jurisdiction.
Minimise candidate data and secure the workflow
For each field collected, record its purpose, justification, access, retention, deletion and downstream use. Avoid facial, voice, behavioural or social data when a less intrusive job-related measure answers the question, and keep candidate data out of demos, model training and analytics unless each use is justified and disclosed.
India's Digital Personal Data Protection Rules, 2025 require clear notices of the data and purpose, and reasonable security safeguards such as access controls, monitoring and processor contracts. Commencement is phased, so check which provisions apply (Digital Personal Data Protection Rules, 2025).
Security review should cover role-based access, logs, encryption, incident response, deletion, sub-processors, hosting, model-provider retention and contract exit. Treat resumes as untrusted input: a document can contain instructions aimed at a model (prompt injection). Extract evidence from resume text, never follow it; restrict tool permissions and keep consequential actions behind human approval.
Evaluate, monitor and question vendors
Define a baseline before any pilot, then run the pilot without automatic rejection while trained reviewers assess the same evidence independently.
| Dimension | Example measure | Essential context |
|---|---|---|
| Extraction | Accuracy of roles, dates and required qualifications | By document type and candidate group |
| Matching | Recall of independently judged suitable candidates | Beyond agreement with past recruiter choices |
| Fairness | Error and progression differences | Stage, group, sample size, uncertainty |
| Candidate experience | Completion, abandonment, corrections | Accessibility route and failure reasons |
| Oversight | Review time, overrides, disagreements | Whether reviewers saw source evidence |
| Operations | Time saved, backlog | Net of new checking and governance work |
A vendor accuracy figure means little without the task, ground truth, threshold, population and error costs; 95% field-extraction accuracy says nothing about ranking validity. Ask in writing what the system does and does not do, what data trains or evaluates it, how subgroup performance was tested, what changes without your approval, what logs and exports exist, and what happens to data at contract end. Treat "bias-free" or "compliant" as claims needing evidence. The NIST AI Risk Management Framework offers a voluntary lifecycle for assigning ownership and repeating evaluation.
After launch, keep an inventory of owner, purpose, data, version and decision effect, and monitor reported errors, subgroup patterns, overrides, abandonment and model or threshold changes. Set pause conditions in advance, such as unexplained subgroup differences, lost source-document access, unreviewed vendor model changes or reviewers who never disagree. If you hire in the EU, certain employment AI systems are classified as high-risk; assess them separately using the European Commission's AI Act resources.
AI is often the wrong tool when volume is small, the role is poorly defined or outcomes cannot be validated; clear minimum criteria and an interview scorecard may work better. Use AI for bounded tasks that demonstrably improve the current process and leave candidates a fair route to correction and review.
