Voice AI Space Paris - Voice AI is not a product; it's a system - Benjamin Chino @ Maki People

The Role of Voice AI in High-Volume Recruitment

Maki provides an end-to-end agentic orchestration platform designed to manage the hiring process from screening and assessment to interviewing and offers. This system is particularly useful for large organizations handling high volumes of applications, such as BNP Paribas, which receives 800,000 applications annually in France but previously could only answer 30,000. By utilizing automated voice agents that operate 24/7, companies can screen candidates immediately after they apply, automatically scoring and extracting insights from the conversations.

The Tension Between Control and Naturalness

Implementing Voice AI in high-stakes scenarios like job interviews introduces a fundamental tension between two key elements:

• Control: Ensuring the AI follows the correct structure, remains compliant, and guides the candidate appropriately.
• Naturalness: Providing a fluid, low-latency conversation that feels human.

Focusing too heavily on control makes the interaction rigid and unnatural, while prioritizing naturalness can cause the AI to lose track of the conversation. Out-of-the-box Voice AI products often fail during longer, complex conversations because they cannot balance this tension or handle unexpected candidate behaviors, such as off-topic responses or varying pause lengths.

Maki's Pipeline Architecture

To achieve reliability at scale, Voice AI must be treated as a complex system rather than a single product or model. Maki utilizes a multi-component pipeline:

• Speech-to-Text (STT) and Turn Detection: While leveraging external models, Maki implements custom turn-detection layers to dynamically manage when the agent should speak, when to let the candidate speak, and how to handle silences or interruptions.
• Finite State Machine (FSM): Instead of relying on a single prompt, the system uses a graph of nodes (an FSM) where each node contains specific prompts. This allows the AI to dynamically steer the conversation, decide when to probe further, and determine when to transition to the next stage.
• Post-Conversation Evaluation: To comply with regulations like GDPR and the AI Act, and to avoid real-time transcription errors, candidate scoring and skill evaluation are performed ad-hoc after the call has concluded. This ensures that decisions are explainable and based on extracted evidence.
• Continuous Simulation: Before deploying conversational agents to real candidates, Maki runs thousands of synthetic test calls mimicking extreme edge cases (such as a candidate trying to order a pizza) to ensure the system remains robust under stress.