Parloa builds voice AI agents for large-company support on OpenAI models

Parloa, a Berlin-based developer of a customer service platform, has described how it uses OpenAI models to launch voice AI agents at large companies. The AMP platform is meant to do more than answer calls: it helps design, test and push to production systems that must run reliably in real time.

How AMP works

Parloa's story began with a fairly mundane problem. One of the company's co-founders, Stefan Ostwald, spent a day in an insurance call center and watched employees handle the same requests over and over: password resets, policy questions, routine account changes. The company first built rule-based voice bots, but with the arrival of ChatGPT and new OpenAI models it moved to the AI Agent Management Platform, or AMP. It is no longer about rigidly scripted flows, but about a platform where companies can build, test and launch LLM-based voice service.

The core idea of AMP is that it is not only for developers. A business team or subject-matter experts set the agent's role, instructions, constraints and connected tools in plain language, with no intent tree and no manual description of every step. The system can then be run through a simulation in which one model plays the customer and another plays the configured agent. The team sees how the agent responds, whether it calls APIs correctly and whether it strays outside the scenario, and can quickly fix the configuration before any real calls.

Betting on verification

Parloa leans heavily on an evaluation-first approach. For enterprise customers, polished demos are not enough; predictability in production matters, because moving to a new model always carries cost and risk. So the company does not treat abstract benchmarks as truth. It builds its own test sets that mirror real customer support scenarios. They measure how well a model follows instructions, how reliably it calls tools, what the response latency is and how the system behaves in edge cases.

"Models only matter when they work in production," is how

Parloa explains its approach to real-time voice systems.

A good result on paper is not enough. Only configurations that consistently pass simulations and automated checks go to production. The platform combines LLM-as-a-judge with deterministic rules: some evaluations check answer quality and adherence to instructions, others verify that critical steps actually happen in the right order. The approach is already producing business results: in one deployment, a global travel company cut the number of requests transferred to a live agent by 80%.

Voice without pauses

For Parloa, a voice interface is a separate engineering challenge. Unlike a text chat, every second is felt directly by the user. The whole chain must run with minimal latency: the system first recognizes speech, then the model composes a reply, then voice synthesis kicks in. Even a small pause at the model layer turns into noticeable silence on the line, so Parloa works with OpenAI to optimize not only answer quality but also speed, stability and how reliably the model follows instructions.

• Speech recognition is checked by word error rate, especially on sensitive data such as policy numbers and account identifiers.
• Speech synthesis is evaluated with blind listening tests to see how natural the voice sounds to real people.
• Speech-to-speech models are tested separately for production readiness on latency, accuracy and cost.
• Multilingual benchmarks are run for different markets, because enterprise customers need the same reliability globally, not in just one country.

Today Parloa's agents handle millions of conversations in retail, travel and insurance. The company is looking beyond the classic phone call: a single service scenario can start by phone, continue in chat and include links or interactive elements along the way. In this approach, channels no longer live separately from one another. For the customer it should be one continuous conversation, not a set of disconnected touchpoints, and that is the format the company is building its platform around.

What this means

Parloa's story shows that the enterprise support market is shifting from simple IVR trees to full AI agent management platforms. The winners will not be those with the loudest model, but those who can test it on real scenarios, keep latency low and connect it safely to a business's internal systems.