According to an opinion column by analyst Zeus Kerravala published in SiliconANGLE, NiCE Ltd. launched the next generation of the NiCE Cognigy platform this week. The thesis underpinning the launch is that the primary challenge for enterprises does not lie in building AI agents, but in operating them reliably across complete customer journeys, while maintaining the visibility and control required in large organizations. Kerravala notes that while the industry has spent the past two years focusing on the question of which AI agent is the smartest, model capabilities are improving and becoming cheaper so rapidly that raw reasoning and intelligence are becoming the least differentiated part of the technology stack.
The new release includes a redesigned Agentic Experience Studio, a companion component called AI Assist that helps with building, testing, and improving agents, reusable "skills" components that package instructions and tools, turn-by-turn tracing across every stage of a conversation, and native support for voice-to-voice models. Default speech-to-text, text-to-speech, and large language models are now included in the per-conversation price, while premium models from Deepgram, ElevenLabs, and Anthropic are available in a few clicks and billed according to usage.
The Gap Between Model Capability and Dependability in the Contact Center
At an analyst briefing ahead of the launch, Nhu Ho, senior product marketing manager at NiCE Cognigy, presented a chart illustrating how long an artificial intelligence system can work on its own before human intervention is required. According to the data presented, in 2020 this was seconds, in 2024 it was minutes, and this year it is roughly 16 hours. However, the footnote in the data showed that the calculation was made based on a 50% success rate. Ho noted that a capable system is not necessarily a dependable system, and that this is precisely where the platform's importance is measured.
Kerravala explains that for customer experience leaders, a 50% success rate might represent a research milestone, but in a customer service contact center it could cause customer churn or a compliance violation. As language models become a generic commodity, value shifts to the system wrapping around them: testing, tracing, guardrails, and deterministic fallbacks. NiCE does not build frontier models, and frontier model developers do not run contact centers; therefore, positioning the platform as the layer providing dependability represents a focused move for the company.
Simplifying the Builder Environment and Expanding Access Across CXone
Prior to its acquisition, Cognigy established its position through complex production deployments for large brands such as Lufthansa, Mercedes, and DHL. However, that flexibility and customizability added complexity. Clement Boutier, director of product management at NiCE Cognigy who leads the development team for the release and owns the platform roadmap, pointed out that the large number of options when configuring an agent created friction that slowed down the process of getting started.
The new workspace enables builders to stand up an agent in a few clicks with pre-provisioned models, while power-user tools—including full access to the agent's system prompt—remain available in dedicated sections. NiCE aims to expand the use of Cognigy across the entire installed customer base of CXone, rather than only among expert teams in massive enterprises. Boutier described three builder personas: smaller businesses building simple agents, contact center administrators and team leads, and engineers at technologically sophisticated companies. The AI Assist tool targets the first two personas; during a demonstration, it built an agent, tested it, fixed the issues it found, and deployed it to WhatsApp. Lowering the barrier to entry increases the number of agents built by less experienced individuals, which makes the governance features the real story of this release.
A Hybrid Work Model Between Autonomy and Defined Flows
Despite extensive discussions surrounding autonomy, most enterprises are not ready to hand over complete customer journeys to an autonomous agent. Boutier outlined three operating modes: traditional natural language understanding (NLU), a hybrid model, and a fully agentic model. According to him, the hybrid model reflects where most customers are today, while fully agentic deployments are at the beginning of the adoption curve. This approach allows enterprises to step into the field gradually and advance toward full autonomy at their own pace.
The product was designed to support this approach: an agent can serve as an orchestration layer executing existing workflows as tools, or structured flows can run in front of the agent, for instance to scrub credit card numbers or personal information before data reaches the model. Ho described the move as an evolution rather than a replacement, so that existing production environments, application programming interfaces (APIs), and integrations continue to run without change, which presents a significant advantage over AI-native challengers whose pitch effectively requires a rebuild.
Observability and Multivariate Testing for Performance Evaluation
Boutier emphasized that without a feedback loop on the build process, one cannot develop a high-quality agent. The new tracing capabilities display every turn of the conversation: which skill was loaded, which tool was called and with what arguments, and when the agent handed off to a structured flow. The Simulator component now supports multivariate testing of up to 10 different variants, using a large language model as a judge (LLM-as-judge) running synthetic customers against the agent. In a demonstration, Boutier ran the same agent against two different models and found that one operated faster and consumed fewer tokens, turning model selection into a data-backed economic decision.
Voice-to-Voice Communication and Connecting to Personal Agents
Support for voice-to-voice models may be the first capability that customers deploy in practice. Boutier noted that buyers are open to using a speech-to-text and text-to-speech cascade, but voice-to-voice models simply sound better. In the demonstration conducted, the agent delegated tool calls to a separate model while continuing to speak, giving the conversation a more natural sound.
In addition, Cognigy now serves as both a client and a server for the Model Context Protocol (MCP) and supports agent-to-agent connections, including one client linking Cognigy to agents built using Microsoft development tools. NiCE is also co-developing the Personal Agent Protocol alongside Meta and Sierra, which is not yet live. Boutier described Cognigy as the front door of the contact center, where a consumer's personal agent will need to authenticate before being allowed to proceed in the process with the brand.
Advice for IT Leaders
In his article, Kerravala suggests several steps for information technology leaders:
- Map the hybrid line: Determine which processes will remain deterministic, such as payments, identity verification, and regulatory disclosures, and which processes will be handed over to an agent.
- Establish full tracing as a go-live requirement: Do not allow an agent to reach production without the ability to explain why each action was taken.
- Use simulations to settle model selection: Compare alternatives by cost per resolved conversation rather than by anecdotal impressions.
- Take ownership of regulatory compliance: Boutier emphasized that responsibility is shared, and while the platform provides tools, it cannot fully control what is inputted into it.
- Treat guardrails as a budget line item: Every guardrail adds tokens and latency, so this budget should be spent where risk is highest.
- Prepare for machine-based customers: Instead of banning personal agents, establish policies in advance regarding which personal agents will be permitted and what they will be allowed to do.
In conclusion, Kerravala notes that the current launch focuses on where value will accrue as AI agents mature: while intelligence is becoming cheaper, dependability, governance, and the ability to combine autonomous reasoning with deterministic control remain the central challenge for enterprises.