How AI Scoring Works

The architecture behind how SageScreen conducts interviews, evaluates transcripts, and produces results.

Overview

SageScreen uses a multi-agent architecture. A single screening runs through 20+ specialized agent layers, each responsible for a distinct stage of the process, and this number grows as capabilities expand. From the reviewer’s perspective, two agents are the most visible: the interview agent that conducts the conversation and the evaluation agent that produces the results. But behind the scenes, additional agents handle tasks such as framework generation, transcript analysis, language processing, and quality checks. This article focuses on the components most relevant to understanding results and reports.

The separation between interviewing and evaluating is intentional: the agent that interviews is not the agent that scores. Each operates with a distinct objective, and the evaluation agent works from the full transcript without the biases that can emerge from live interaction dynamics.

The Interview Agent

The interview agent is the Sage itself: the AI screening agent that candidates interact with during the screening. It conducts a structured, adaptive conversation based on the interview framework generated during the sage build.

During the interview, the agent:

• Opens with the intro question defined in the sage’s cultural alignment configuration.
• Asks contextually relevant questions aligned to the role, seniority, and evaluation criteria.
• Adapts follow-up questions in real time based on the candidate’s responses.
• Maintains the conversational tone selected during sage creation (e.g., Professional, Friendly, Direct).
• Manages the interview within the configured duration.

The interview agent’s role is to gather information. It does not score, rank, or evaluate the candidate during the conversation. Its output is a complete transcript of the exchange.

The Evaluation Agent

After the interview concludes, the evaluation agent receives the full transcript and the sage’s evaluation framework (categories, rubric, and scoring criteria). It processes the transcript as a single unit and produces the structured results that appear in the report.

The evaluation agent operates independently from the interview agent. It was not present during the conversation. It evaluates what was said, not how the conversation felt in the moment. This separation reduces the risk of anchoring effects, where early impressions from the interview might disproportionately influence later scoring.

Important

The agent that interviews is never the agent that scores. This architectural separation is deliberate and prevents live interaction dynamics from influencing the evaluation outcome.

How Categories Are Generated

Evaluation categories are not drawn from a fixed template. They are generated dynamically during the sage build, derived from the role context and cultural alignment inputs you provide.

When you create a sage, you supply a job description, role title, seniority level, evaluation guideline, tone preferences, and any additional instructions. During the build phase, the system analyzes these inputs and generates a set of evaluation categories that reflect the dimensions most relevant to the role. A sage built for a senior software engineer will have different categories than one built for a customer support representative.

Examples of categories might include technical competence, communication clarity, problem-solving approach, cultural alignment, or leadership potential, but the specific set varies per sage. The categories are locked at build time and remain fixed for the life of the deployed sage. Every candidate screened by that sage is evaluated against the same categories.

A sage has between 3 and 8 evaluation categories, depending on the role and the depth of the screen. This count covers the core evaluation dimensions only. It does not include secondary signals such as bilingual assessment, English proficiency, or other supplementary metrics, which are tracked and reported separately.

Category Results

Each evaluation category receives an alignment rating, accompanied by two supporting elements:

• Explanation: a written rationale describing what the evaluation agent observed in the candidate’s responses and why it resulted in that rating.
• Quotes: direct excerpts from the candidate’s actual responses that support the assessment. These provide a traceable link between the rating and the candidate’s own words.

The combination of rating, explanation, and quotes gives reviewers a transparent basis for each evaluation dimension. Reviewers can assess whether the rating aligns with their own reading of the quoted material.

Not every category counts equally. The dimensions a role depends on most carry more weight in the overall result than peripheral ones. Always read the recommendation alongside the category breakdowns: consistent performance across every category and a standout area offsetting a weaker one can lead to the same recommendation. The category-level detail is where the actionable signal lives.

The Recommendation

The evaluation agent produces an alignment recommendation of Strong, Moderate, or Weak. The recommendation reflects the candidate’s performance across all categories, measured against the evaluation guideline selected during sage creation. The guideline sets how high the bar is for a Moderate or Strong result:

Guideline	Description
Flexible	Open to a range of approaches; the lowest bar for a passing result.
Relaxed	Moderate expectations with room for growth areas.
Balanced	Solid performance expected across categories.
Strict	High bar; strong performance required in all areas.

Any inclusion or exclusion rules configured in the sage are also applied, and they can override the recommendation the evaluation alone would produce.

Note

The recommendation is a signal, not a verdict. It is designed to inform the human reviewer’s decision, not replace it.

What AI Scoring Does Not Do

Understanding the boundaries of the system is as important as understanding its capabilities.

1. It does not make screening decisions. The platform produces structured data: category ratings, explanations, recommendations, and transcripts. A human reviewer always makes the final determination.
2. It does not compare candidates against each other. Each candidate is evaluated in isolation, against the sage’s fixed rubric. Results are absolute, not relative.
3. It does not learn from results. There is no feedback loop where human decisions influence future scoring. Each sage’s evaluation framework is fixed at build time and does not adapt based on outcomes.
4. It does not train on candidate data. Candidate responses, transcripts, and evaluation data are not used to train AI models.
5. It does not evaluate identifying information. The evaluation agent assesses candidates based on the content of their responses. Candidate names and locations are hidden from the evaluation process.