NiCE Cognigy updates its platform to manage autonomous customer service agents

Customer service centers have become a prime testing ground for artificial intelligence tools. While much of the tech industry continues to debate which underlying model offers the sharpest reasoning, enterprise software firm NiCE Ltd. argues that raw intelligence is turning into a commodity. Instead, the firm is betting that the real challenge lies in operating and overseeing a network of AI agents reliably across real customer interactions.

Supervising Autonomous Fleets

NiCE released the next generation of its NiCE Cognigy platform this week. The system is designed to help businesses manage automated agents without letting errors reach the customer. According to NiCE, while recent AI benchmarks show models can operate on their own for roughly 16 hours before a human must step in, that metric reflects a 50 percent success rate. In a production call center, a failure rate that high can lead to lost accounts or regulatory violations.

To fix this, the new release adds step-by-step tracing to track tool calls, skills, and system handoffs during a chat. It also includes an automated testing system that uses one AI model to evaluate simulated conversations across up to 10 setup options, letting managers compare tokenTokenA small piece of text, about three quarters of a word, that an AI model reads or writes. Open in glossary consumption and speed across competing models.

Hybrid Controls and Low-Code Building

The updated software offers a redesigned Agentic Experience Studio and a helper tool called AI Assist, which creates, tests, and deploys bots to channels like WhatsApp. This feature lets non-technical team leaders stand up tools quickly, while still giving engineers access to system prompts and guardrails.

Most enterprise clients are not yet ready to let autonomous bots run unchecked. Because of this, Cognigy supports a hybrid approach. Companies can route traffic through deterministic workflows first, ensuring private data like credit card numbers is removed before any input reaches a third-party model. The system also supports direct voice-to-voice models and agent-to-agent interactions, including links to tools made with Microsoft software.