NiCE launches next-generation Cognigy, betting on running agents rather than building them

NiCE has released the next generation of its Cognigy platform, and the pitch is deliberately unglamorous: the hard part of enterprise agents is not building one, it is running a fleet of them reliably.

The release centers on a redesigned Agentic Experience Studio that lets a builder stand up an agent in a few clicks with models already provisioned, plus an AI Assist companion that spans building, testing and improving agents. Cognigy's historic strength was deep configurability for complex deployments, but the sheer number of options also created friction for newcomers; this version is aimed at shrinking that gap.

Two additions carry much of the governance story. Reusable skills package instructions and tools into units that can be called again and again, while turn-by-turn tracing shows operators what an agent decided at each step of a conversation. For companies that want frontline business teams, not just central engineering, to maintain agents, those changes matter more than one extra model option.

On models, default speech-to-text, text-to-speech and large language models are now included in the per-conversation price, with premium third-party models available with a few clicks and billed by usage. Native voice-to-voice model support is meant to cut the latency and tonal loss that comes from stitching separate speech components together.

NiCE draws a line between three deployment modes: traditional natural language understanding, hybrid, and fully agentic. Its product management describes hybrid as where most customers actually sit today, with fully autonomous deployments still near the beginning of the adoption curve. Agents can therefore serve as an orchestration layer calling existing flows as tools, while existing flows can also sit in front of an agent to handle steps that must stay deterministic, such as scrubbing card numbers.

Lowering the barrier cuts both ways: more agents will be built by less experienced people, which turns testing, tracing, guardrails and deterministic fallbacks from nice-to-haves into requirements.

The framing rests on a striking number shown to analysts, tracking how long AI can work before a human has to intervene: seconds in 2020, minutes in 2024, roughly 16 hours this year, but at a 50% success rate. In a contact center, 50% is not a research milestone, it is a churn event or a compliance violation. The question now is whether the platform can spread across NiCE's existing customer base rather than staying with expert teams at the largest brands.