VagaIA

Instructions for use of AI screening

Last updated: 9 November 2026

These instructions explain to companies and agencies using VagaIA what AI screening is for, its limits and what each team must do to use it with human oversight. Candidate screening is a high-risk use under the EU Artificial Intelligence Act.

1. Intended purpose

Helping a person decide whom to follow up with for a job, by checking each CV against the criteria the company set and showing the evidence. The system does not make decisions about candidates and must not be used to assess performance, promotions, dismissals or personal characteristics.

2. How it works

  • The company sets the job's criteria (knockout, important and bonus). The AI suggests them from the job ad; the company adjusts them. Discriminatory criteria are blocked.
  • For each CV, the AI model (Claude, by Anthropic) states, criterion by criterion, whether the CV meets it, does not meet it or does not mention it, with the sentence and page of the CV.
  • Based on that, the system suggests three groups: Meet, To review and Don't meet. A candidate is only suggested for “Don't meet” when the CV contradicts a knockout criterion with a quote; if the CV does not mention it, they go to “To review”.
  • The full analysis (summary, strengths, gaps, interview questions) is generated for those who meet the knockout criteria or on request.

3. Known limitations

  • It may misread scanned CVs, tables or unusual formats. A hard-to-read warning appears in those cases.
  • It only assesses what is written in the CV: it does not know what the candidate left out and does not assess skills not in the document.
  • It works best in Portuguese and English.
  • It may misinterpret an ambiguous criterion. Clear, verifiable criteria (for example “3 years of accounting experience”) give better results.
  • It does not assess age, gender, nationality, marital status or photos, but a CV may contain hints of them; the human decision must disregard them.

4. Accuracy

Accuracy is monitored through the group changes teams make by hand and through disputed quotes. We do not publish a formal metric yet; when one exists, it will be on this page.

5. What the company's team must do (human oversight)

  • Review the evidence before deciding: read the quote and, when in doubt, the original CV, which is always shown alongside.
  • Confirm every rejection: nobody moves to “Not moving forward” or receives a reply without a person's confirmation, which is recorded.
  • Do not treat the suggested group as the final decision, especially under “To review”.
  • Make sure those who decide are sufficiently trained to use the system and know these instructions.
  • Inform candidates that an AI system is used. The application form and the data processing notice sent to manually uploaded candidates already do this.
  • Answer candidates' requests for an explanation of a decision, using the recorded criteria and quotes.

6. Logs

The system records who did what and when (screening, group changes, confirmed rejections, emails, sharing, client decisions). Logs are kept for at least 6 months, even after the candidate's data is anonymised, and administrators can export them.

7. Reporting problems

If the company detects a serious error, a discriminatory result or another risk, it should stop using the system for that job and report it immediately to suport@vagaia.pt.

See also the Terms for Companies. Contact: suport@vagaia.pt