Parloa builds service agents customers actually want to talk to

Berlin-based Parloa built an AI Agent Management Platform (AMP) that uses OpenAI models to simulate, evaluate, and operate voice-based customer service systems.

It started out as a rule-based voice agent, but now, built on the latest models such as GPT-5.4, it lets business users define roles, instructions, tools, and boundaries in natural language without code, and quickly iterate and validate them.

AMP runs conversation simulations and evaluations before deployment. One model plays the customer role while another runs the agent, and the results are verified with deterministic checks and LLM-as-a-judge to confirm instruction adherence, tool calls, and task completion.

In production, the key is a low-latency pipeline connecting speech-to-text, model inference, and text-to-speech. Parloa separately evaluates the word error rate of speech recognition, the naturalness of speech synthesis, and the latency, accuracy, and cost of speech-to-speech models, deploying only the models suited for large-scale operation in a globally multilingual environment.

As agents grew more complex, the side effects of a single massive prompt grew as well, so Parloa split tasks like authentication, reservation changes, and account updates into separate sub-agents, and fixed critical steps with structured API chains and event-driven logic to improve stability. In one case with a global travel agency, requests for human agents dropped by 80%.