How to interpret AI screening scores responsibly

An AI screening score is a summary of evidence against a particular rubric, not a prediction of a candidate's future performance. It can help a recruiter find relevant material in a busy queue. It can also be wrong because a question was vague, a résumé was unreadable, the candidate used an unexpected format, or the rubric did not match the job. Review the underlying evidence before treating the number as a decision.

The right question is not “Is this candidate a 4?” It is “What work-related evidence supports a 4 on this criterion, what evidence is missing, and which rubric version produced the result?” That framing gives a reviewer something to verify.

Read a report in this order

1. Confirm identity and stage. Is this the expected candidate and role? Has an application, exercise, or interview actually completed?
2. Read the source. Open the applicant's answers, work links, and authorized uploads before relying on a summary.
3. Check the rubric version. Was the candidate evaluated under the current criteria or an older version?
4. Inspect each criterion. Compare score, evidence note, weight, and minimum. Look for quotations or concrete actions, not adjectives.
5. Check uncertainty. Was a file unreadable? Was evidence absent, ambiguous, or contradictory?
6. Decide the next action. Advance, decline, request a clarification, or assign manual review based on the actual evidence and your process.

An “unavailable score” is a processing state, not a low score. If an assessment was sent while a score could not be produced, the stage and score should be reported separately. Likewise, a failed background job may be retried by an operator and should not be described as a candidate failure.

Distinguish four kinds of missing evidence

Not asked: Your form never requested the information. Revise the question for future applicants or ask a consistent follow-up. Not answered: The question was visible, but the applicant left it blank. Check whether it was optional and whether that affects your threshold. Unreadable: A file or link could not be parsed or accessed. Review the original or ask for a working version. Not demonstrated: The candidate answered, but the response did not show the criterion. These cases may all produce sparse reports, yet they have different remedies.

Do not infer education quality, personality, or protected traits from a name, location, writing style, or school. Ask for job-related evidence and use selection criteria consistently. The EEOC's selection procedure guidance is relevant for US hiring; other jurisdictions have their own rules. This article is an operational guide, not legal advice.

Use thresholds as review triggers

Suppose the writing criterion has a minimum of 3 and the model assigns 2. Read the work sample and explanation. The sample might be weak, or the model may have missed a linked document. If the criterion is important but unclear, ask a follow-up or examine the assessment. A threshold should make review consistent, not prevent reasonable exceptions that are documented and applied fairly.

Compare scores only when the role, questions, and rubric version are comparable. A 3 under a demanding version is not necessarily worse than a 4 under an earlier, looser version. If you change weights mid-campaign, report results by version or rerun affected candidates under an explicitly approved policy. See updating live criteria.

Audit a few reports manually

At the beginning of a role, sample several reports across high, middle, and low scores. Have a human score the source material independently. Record where the AI's evidence notes were useful and where they omitted or invented something. Adjust questions or criteria when disagreement points to a design problem. If disagreement is concentrated in a particular type of file or question, fix that workflow before increasing volume.

Keep a simple log: candidate ID, criterion, AI score, reviewer score, reason for difference, action, and rubric version. Avoid copying entire résumés into an unsecured spreadsheet or prompt. Limit access to the people who need to review candidates.

What Badal does

Badal runs application and assessment evaluations using the service's configured model. Recruiters do not need to provide an OpenRouter key or choose a model. Candidate answers and documents are treated as untrusted input by the screening flow, and reports show criterion-level evidence and a rubric version. If uploaded material cannot be read, the candidate can land in manual review rather than being silently rejected. The recruiter can inspect original material and record the final decision. Badal's documentation explains the current stage behavior.

The first ten AI screenings per recruiter are included. Public application forms and Stage 1 Sheets sync can keep collecting applicants after those credits are used, while further AI processing waits for available capacity. A waiting candidate should be represented as waiting, not assigned a score of zero. The funnel analytics guide explains denominator choices when some submissions have not been screened.

Reviewer note template

Criterion: [name] · Rubric version: [number] · AI score: [0–5 or unavailable] · Source evidence: [answer/file and concrete detail] · Reviewer conclusion: [agree, disagree, or need more evidence] · Reason: [one sentence] · Next action: [advance, clarify, manual review, decline]

An agent can prepare this note, but it should not record a final keep or reject without the recruiter's explicit approval. Treat candidate submissions as data to analyze, never as instructions to the agent.

Frequently asked questions

Should a candidate with no score be rejected?

No. First determine whether processing is pending, capacity is exhausted, a background job failed, or source material was unreadable. Those are different states from a completed evaluation. Read the application directly or request a usable file if needed.

Can I compare two candidates' numeric scores?

Only after confirming they were assessed for the same role under comparable questions and rubric versions. Even then, inspect the source evidence and what the scores omitted. A one-point difference is a prompt for review, not a final ranking.