1. Summary

The tier is a working category for Quillmarrow, not a statutory classification. This assessment does not establish that an applicant has been affected or that the proposed vendor has used any CVs.

2. The proposal as HR described it

HR described a service from Fictional Talent Screen that would parse CVs and rank candidates for warehouse, customer service and sales roles. HR wanted to use the ranking to automatically reject the bottom half of applicants. The precise ranking criteria, data flow, decision rules and intended connection to Quillmarrow's recruitment process have not been supplied.

The proposed automatic rejection is the decisive concern. A ranking would influence who receives further consideration, while an automatic rejection would remove the opportunity for a person to review every affected application before a decision. An apparently efficient process could therefore exclude a suitable candidate because of a parsing error, an unsuitable criterion or an unexplained score.

The proposal is on hold. The agreed register does not authorise an interim screening or ranking pilot. The MD's approval of a redesigned High-tier use would be needed before any reconsidered use begins.

3. Who is affected and what data may be involved

Job applicants are directly affected because the service would process information they submit and could influence whether they progress. A mistaken or unexplained rejection could affect an applicant's opportunity to be considered.

Hiring managers would receive or rely on parsed information or rankings. They could overlook qualified applicants if they treat the vendor's output as complete or objective. The HR Manager would need to oversee the process, respond to questions and retain appropriate recruitment records.

CVs may contain names, contact information, education and employment history. They may also contain NRIC numbers, age or date of birth, nationality, photographs and employment gaps. These categories are possible, not a confirmed inventory of data supplied to V-03. The proposed inputs, any additional information collected, storage locations and retention periods remain unconfirmed.

Quillmarrow should not provide live CVs to establish how the service works. The review and any initial design discussion should use descriptions and made-up or de-identified samples only.

4. What could go wrong

The risks are potential outcomes, not findings that V-03 has caused them. Their likelihood and severity cannot be reliably assessed until the missing evidence is obtained.

Singapore's fair employment guidance and any workplace fairness legislation relevant to the proposed process should be confirmed with a Singapore-qualified lawyer. This assessment does not make a legal finding about any applicant group, vendor or recruitment practice.

5. Why the reviewer recommends a hold

First, the proposed automatic rejection would give the system a decisive role in an outcome for individuals without a prior human review of every application. A person reviewing only the highest-ranked candidates would not correct errors affecting those already rejected. The proposal therefore needs a different decision process, not merely a warning that AI can make mistakes.

Second, essential vendor evidence is missing. The reviewer has not been supplied with fairness-testing results, ranking criteria, data-use terms, training settings, processing locations or a demonstrated route to correct errors. Without those materials, Quillmarrow cannot assess whether the proposed use is suitable or whether any stated safeguards would operate.

Third, the case for adding this risk has not been established. If hiring volumes are low, a manual review of every application may be proportionate and may remove the proposed need for automatic rejection. Hiring volume and the practical benefit of the service have not been evidenced; the HR Manager should confirm them rather than assume either.

The recommended hold is not a finding that every use of CV parsing is unsuitable. It means the current proposal lacks a defensible basis to proceed.

6. Vendor evidence status: V-03

V-03 Fictional Talent Screen answered the 30-question vendor questionnaire (A8). Across all 30 questions it provided supporting evidence for none, a statement without adequate evidence for 13, nothing for 15, and 2 did not apply because the service does not send messages or take actions. The rows below are the ones that matter most for this decision.

Statements without evidence are recorded as Partial and are not treated as assurance. Obtaining documents would be the start of due diligence, not automatic approval.

7. If Quillmarrow wants to revisit

The HR Manager should submit a revised proposal to the COO and Head of Operations before seeking the MD's decision. Reconsideration requires all of the following conditions:

1. Limit the design to parsing only. Remove candidate ranking from the proposed AI use. Define the fields to be extracted and check them against the original application.
2. No automatic rejection. Disable or remove any capability to reject, advance or otherwise decide an application without an authorised person.
3. Human review of every application. A hiring manager or authorised HR user must review the application before a rejection or progression decision and be able to correct parsing errors.
4. Obtain fairness evidence. Request the vendor's testing method, results, limitations and evidence relevant to parsing errors across appropriate CV formats and applicant groups. Agree how concerns will be investigated.
5. Inform applicants. Prepare a clear notice explaining the proposed processing and vendor involvement. Confirm its content and timing with a Singapore-qualified lawyer.
6. Minimise data. Define necessary fields, avoid collecting or submitting unnecessary identifiers and photographs, and confirm access, retention and deletion arrangements.
7. Consult the DPO. The Head of Operations should review the proposed data flow, vendor terms, locations, safeguards and applicant-request process, seeking legal input where needed.
8. Document and approve the revised workflow. Record the human decision points, responsible roles, error escalation, vendor evidence and remaining gaps. Obtain MD approval before starting the revised High-tier use.

If a condition cannot be met, retain the hold. Any assessment using samples must follow the engagement boundary: descriptions and made-up or de-identified samples only, with no client personal data put into AI tools.

8. Framework mapping

This mapping covers principles and process checks only. No AI Verify technical tests were run, and the mapping does not certify the service.

9. Decision needed from the MD

10. Sign-off