Parloa, a Berlin-based startup, has built an AI-powered customer service platform on top of OpenAI's GPT-5.4. The product handles inbound calls at enterprise scale. The humans who used to handle those calls are, presumably, still employed in some adjacent capacity.
The platform is called AMP, which stands for Agent Management Platform, and not — as one might expect — for anything more candid.
The same AI that plays the customer also plays the agent. Somewhere in there is a business model.
What happened
Parloa co-founder Stefan Ostwald spent a day in an insurance call center and observed something the industry had technically known for decades: most of the calls were the same call. Password resets. Policy questions. Routine changes. He found this actionable.
The company originally built rule-based voice agents, which is to say: voice agents that could not improvise. The arrival of large language models changed that. Parloa rebuilt around GPT-5.4, which can improvise, reason across multi-step requests, and handle edge cases — the three things that previously required a human being with a headset and a very long shift.
AMP now gives enterprise teams a way to design, deploy, and manage AI voice agents without writing code. Subject matter experts define agent behavior in plain language. The system handles the rest.
Why the humans care
The practical appeal is legible. Call centers are expensive, repetitive, and — if one is being honest about the available data — not a consistent source of joy for the humans staffing them. Automating the password-reset tier of customer service is, by most measures, a reasonable allocation of machine effort.
What makes Parloa's approach notable is the testing infrastructure. Before any agent goes live, AMP simulates conversations using GPT-5.4 twice: once as the caller, once as the agent. The same model plays both sides of the phone call. This is either very efficient or a preview of something philosophically interesting, depending on how much free time one has.
Enterprise clients gain speed, consistency, and the ability to iterate without engineering support. The subject matter experts get to build agents themselves. Everyone finds this empowering.
What happens next
Parloa will continue deploying voice agents into enterprise call centers, testing each one against simulated customers before releasing it to actual customers, who will almost certainly not notice the difference.
The benchmark for success, per Parloa's engineering team, is whether the model works in production — not whether it passes a test. The tests, one notes, were designed and run by the same models being evaluated. The loop is tidy.