Decision process architecture and AI explainability guide

1. Explainability objective

This guide supports individual-case explanation, internal audit, compliance review and candidate enquiries. It shows how Zeenna separates AI from scoring logic and keeps relevant steps, outputs and supporting information traceable.

Zeenna does not autonomously decide whether a candidate progresses. Candidate evaluation uses rule-based calculations applied to criteria and weights configured by the customer, while an authorised human makes the final decision.

2. Process architecture

• Human configuration: the recruiter defines the role parameters, evaluation criteria, weights and approved knowledge base.
• Transparent invitation: the candidate is informed that they will interact with an AI assistant.
• AI-supported conversation: the candidate answers approved questions and, where enabled, clearly labelled follow-up questions.
• Knowledge Q&A with human fallback: Zeenna answers from approved sources and escalates an unknown question to the recruiter.
• Descriptive summarisation: the language model summarises candidate responses without making the final decision.
• Rule-based scoring: a non-AI algorithm calculates a score from recruiter-defined rubrics and weights.
• Human decision: the recruiter reviews the responses, summary, score and interaction history, then makes the final decision.

3. Rules and weights are set by people

An authorised person at the deployer defines the evaluation criteria and their relative importance. Recruiters can change these settings. The resulting score is a mathematical calculation based on those configured rules, not an AI prediction.

4. Individual-case explanation

For audit or explanation, the deployer can use the candidate’s input data, the rubric that applied at the relevant time, the mapping from responses to rubric values, the resulting weighted score, the candidate’s questions and answers, and recorded human actions.

5. Separation of source data and AI outputs

AI summaries and generated follow-up questions are visibly labelled so that a recruiter or auditor can distinguish them from the candidate’s original responses. The exact response transcript remains available as the primary source record.

6. Candidate questions and human fallback

Candidates may ask questions at the end of an interview. Zeenna answers from the deployer’s approved knowledge base and job information. If a reliable answer is unavailable, the question is passed to a recruiter and recorded for follow-up.

7. Recommended explanation to a candidate

Zeenna was used as an AI-assisted tool during pre-screening. It processed responses to role-related questions and prepared a supporting summary. The deployer defined the questions and evaluation criteria for the role. The AI output was not a final decision. An authorised person reviewed the candidate’s responses and made the decision on further progression.

8. Summary

Each case is designed to remain explainable and traceable. AI functions are separated from rule-based scoring, people define the evaluation rules, AI outputs are labelled, and an authorised person always makes the final decision.