When you set up AI Signals to screen candidates, the quality of your results depends entirely on how clearly you define the description of a custom signal.

This article walks through the most frequent setup mistakes, explains why they cause unreliable results, and shows how to rewrite them for consistent outcomes.

1. Keep every signal description unambiguous

AI Signals can only be as precise as the instructions provided. If the description of a custom signal leaves room for interpretation (such as net vs. gross salary, or missing fallbacks), the model has to guess, leading to inconsistent evaluations.

Example: Ambiguous salary expectations

Signal description:
"Candidate salary expectation must not exceed 4,000."

The problem: The instruction lacks a currency, a time period (monthly vs. annual), and a gross/net specification. A candidate stating 40,000 could be within budget if annual, but far outside if monthly.

Why the AI reached this conclusion:

• Unclear parameters: The model cannot determine whether numeric values in the CV refer to monthly or annual figures.
• Missing data: The model found no salary information in the candidate's profile.

How to fix it:

• Be explicit: State the unit, currency, period, and gross/net explicitly (e.g., "above €4,000 gross per month").
• Capture input at application: Add an application form question for salary expectations, as candidates rarely provide structured salary data in a standard CV.
• Define a fallback: Specify how the AI should behave if no salary data is provided (e.g., "If no salary is stated, mark as Undetermined").

2. Avoid multi-criteria blending in a single signal

Adding multiple unrelated skills or experience requirements into a single signal forces the AI into an arbitrary calculation. A small edit to one criterion can flip the entire outcome, making results hard to trust or explain.

Example: Blending multiple competencies

Signal description:
"Evaluate BD, Finance, Marketing, and Engineering experience alongside teamwork and adaptability."

Why the AI reached this conclusion:

• Causes of the problem: The prompt combined roughly 10 different departments and soft skills into one check, creating conflicting evidence (e.g., strong engineering experience but no marketing history).
• Impact: A single composite result makes it impossible to audit how any individual competency was assessed.

How to fix it:

• One signal per criterion: Split multi-criteria checks into separate signals.
• Use system fields: Use Personio's built-in system fields (such as Education and Experience) for hard, structured criteria instead of free-form AI descriptions.

3. Use Automated Screening for binary or non-CV criteria

AI Signals evaluates evidence found in the CV or application form. It cannot verify criteria that candidates rarely list on a CV, nor should it be used for strict pass/fail rules.

Example: Information rarely found in a CV

• Signal description: "The candidate must hold a regular driver's license."
• The problem: CVs rarely list driver's licenses or work permits, leaving the AI with no evidence to check. Asking the AI to find missing evidence often results in an unnecessary flag or failed assessment.
• How to fix it: Add a dedicated application form question to capture this information directly from the candidate, then point the prompt at that form response instead of the CV.

Example: Strict binary / knock-out criteria

• Signal description: "Candidate must be willing to work on-site at least 3 days a week in the office."
• The problem: CVs are silent on hybrid or on-site preferences, so the AI typically flags this requirement as failed by default. Binary knock-out criteria belong in screening rules, not AI prompts.
• How to fix it: Use Personio's Automated Screening (dropdown or yes/no rules) for binary knock-out criteria. Reserve AI Signals for qualitative, experience-based assessments where nuance adds value.

Quick reference summary