What is the training process for the AI model? Mercu does not train its own AI models. We use OpenAI's foundation models under their enterprise privacy policy (no model training on provided data).

What models does Mercu use? We use OpenAI’s foundation models to power our AI features. We routinely evaluate their performance to ensure consistency, accuracy, and fairness across use cases. To avoid issues like model drift or degraded output quality, we monitor behaviour over time and may upgrade to newer model versions as they become available.

How does Mercu monitor and control the data that you send to OpenAI? We only send a specific set of data to OpenAI and keep comprehensive logs on every request. These logs include the timestamp, input, output, model, and configurations. All data shared with OpenAI falls under their enterprise privacy policy (no model training on provided data).

AI responses are known for "hallucinating". How does Mercu control for this? We limit AI hallucinations by tightly constraining how and where AI is used, and by applying conservative settings like low temperature. For instance, our candidate FAQ assistant uses retrieval-augmented generation (RAG), ensuring answers are grounded in predefined sources like job descriptions. We carefully design prompts, monitor outputs continuously, and never allow AI to make final hiring decisions. All AI responses are shown with full context, enabling recruiters to verify accuracy before taking action.

What tools are available to monitor for bias and ensure Mercu delivers a fair outcome? To monitor for bias and ensure fair outcomes, Mercu uses AI as a supportive - not decision-making - tool. All AI outputs (like reply scores) are shown with full candidate responses for transparency, and scoring is based only on text transcripts, excluding any personal or demographic data. This design helps reduce bias and allows recruiters to make informed, fair decisions.

Is Mercu a deterministic matching engine or generative AI system? Mercu is a hybrid system that combines deterministic logic with generative AI, depending on the use case. For rule-based decisions - like automatically rejecting candidates who state they lack working rights in a specific country - we use deterministic workflows. For more qualitative tasks, such as assessing open-ended responses in chat or video/voice interviews, we use generative AI. GenAI also powers features like the candidate FAQ assistant, transcription of asynchronous interviews, and semantic relevance checks. So while Mercu isn’t solely a deterministic engine or a pure GenAI system, it uses the strengths of both approaches where they make the most sense.