It is 2 a.m. and your flight leaves in six hours, but the airline has just moved it to a different airport. You call the number on your booking, half asleep, and an AI voice answers. It listens, checks your ticket, and rebooks you in under a minute.
No hold music. No repeating your confirmation code three times. That is the promise of the AI agent in the contact centre. The problem is that the promise keeps breaking in the real world, and the failures are expensive.
NiCE, the company behind one of the world’s biggest contact-centre platforms, thinks it knows why. This week it launched a rebuilt version of its Cognigy platform, and the message behind it is blunt. Building an AI agent is no longer the hard part. Running one reliably, day after day, is.
According to SiliconANGLE, which covered the launch after an analyst briefing with NiCE, the release is aimed squarely at that gap. The new Cognigy brings a redesigned Agentic Experience Studio, an AI Assist companion, reusable “skills”, turn-by-turn tracing and native voice-to-voice support.
The New NiCE Cognigy: Running Agents Is the Real Product
The contact centre has been the obvious first target for agentic AI since the technology arrived. Ask any business leader where they want AI agents to work first, and the answer is almost always the same.
It is where customers actually talk to a company, and where a bad experience turns into a lost customer within minutes.
Yet for two years the industry has mostly argued about which model is smartest. NiCE’s case is that this is the wrong debate. Models are improving fast and getting cheaper at the same time. Raw intelligence is becoming the least distinctive part of the stack.
That framing matters because of who NiCE is. It is the market leader in contact centre as a service, and Cognigy was already a leader in conversational AI before NiCE acquired it.
The company is not trying to build a frontier model. It wants to be the layer that makes other people’s models dependable, through testing, tracing, guardrails and deterministic fallbacks.
Capable Is Not the Same as Dependable
At the briefing, Nhu Ho, senior product marketing manager at NiCE Cognigy, opened with a chart that summed up the argument.
It showed how long an AI can work on its own before a human has to step in. In 2020, the answer was seconds. In 2024, it was minutes. This year, it is roughly 16 hours.
The footnote mattered as much as the headline. That 16-hour figure was calculated at a 50 per cent success rate.
In a research lab, a coin flip is a milestone. In a contact centre, it is a churn event or a compliance violation. “Capable and dependable do not always mean the same thing,” Ho said. “That is exactly where the platform matters.”
A Simpler Studio, With the Power Tools Still There
Cognigy earned its reputation handling complex production deployments for large brands such as Lufthansa, Mercedes and DHL. That flexibility came with a cost. The sheer number of options when configuring an agent created friction, and getting started could be slow.
The rebuilt studio is NiCE’s answer. A builder can now stand up an agent in a few clicks, with models already provisioned. The power-user tools have not gone away, though. Full access to the agent’s system prompt and the deeper configuration options still live in advanced sections, for the engineers who need them.
Clement Boutier, the director of product management who leads the team behind the release, described three kinds of builders the new studio must serve. There are smaller businesses building simple agents, contact-centre administrators and team leads, and engineers at technically sophisticated companies.
AI Assist, the companion that helps build, test and improve agents, is aimed squarely at the first two groups. In the demo shown to analysts, it built an agent, tested it, found the problems and deployed it to WhatsApp.
NiCE wants Cognigy to spread across its CXone installed base, not just among expert teams at the largest enterprises. The challenge is reaching thousands of users without creating a sprawl of agents nobody governs.
Hybrid Is Where Most Customers Actually Live
Despite all the talk of fully autonomous agents, most enterprises are not ready to hand over entire customer journeys to a machine. Boutier outlined three modes: traditional natural language understanding, hybrid, and fully agentic. Hybrid, he said, “is actually where most of our customers are”.
An agent can act as the orchestration layer, calling existing deterministic flows as tools. Flows can also sit in front of agents, scrubbing credit card numbers and other personal data before anything reaches a model. Ho described the release as “an evolution, not a replacement”, with existing deployments, APIs and integrations running unchanged.
That continuity is a genuine advantage over AI-native challengers whose pitch effectively requires a rebuild. Contact centres move fast, and deterministic workflows are not legacy baggage. They are how a business enforces the things a model should never decide alone, like taking a payment or confirming an identity.
Tracing Every Turn: Observability Takes Centre Stage
If there is one feature working engineers will care about most, it is probably the tracing. “You can have the best builder experience, but if you don’t have any feedback loop on what you’re building, you’re never going to build a great agent,” Boutier said.
The new tracing shows each turn of a conversation: which skill loaded, which tool was called and with what arguments, and when the agent handed off to a flow. That is the difference between knowing an agent failed and knowing why it failed.
The Simulator has been upgraded too. It now supports multivariate testing of up to 10 variants, using a large language model as a judge to run synthetic customers against an agent.
In the demo, Boutier ran the same agent on two different models and found that one was faster and consumed fewer tokens. Model selection stops being a matter of preference and becomes a data-backed economic decision.
The pricing model has been simplified as well. Default speech-to-text, text-to-speech and large language models are now included in the per-conversation price. Premium models from Deepgram, ElevenLabs and Anthropic are available in a few clicks, billed by usage.
Voice-to-Voice and the Coming Machine Customers
Voice-to-voice support may be the first feature customers actually deploy. Buyers are open to using a speech-to-text and text-to-speech cascade, Boutier noted, “but what they realize is that voice-to-voice models sound just better”.
Cognigy now works as both a Model Context Protocol client and a server, and it supports agent-to-agent connections. One customer has already linked Cognigy to agents built with Microsoft’s tooling. NiCE is also co-developing the Personal Agent Protocol with Meta and Sierra, though that is not live yet.
Boutier described Cognigy as “the front door” of the contact centre, the place where a consumer’s personal agent must authenticate before it can proceed. It is notable that NiCE is collaborating with Sierra on this, a direct competitor, because a shared standard for agent identity helps everyone.
The timing is interesting. United Airlines banned AI agents from booking flights this week, while NiCE is betting on enabling personal agents instead. SiliconANGLE’s Zeus Kerravala argued banning them is a mistake: if something is good for a consumer, it will happen.
What Happens Next for NiCE Cognigy
For IT leaders, the advice from the launch is refreshingly practical. Map the hybrid line, and decide which processes stay deterministic, such as payments, identity verification and regulated disclosures. Make tracing a go-live requirement, so no agent reaches production without someone able to explain why it did what it did.
Compliance, NiCE is clear, remains a shared responsibility. “We offer the tools and the platform, but we cannot 100% control what they put in it,” Boutier said.
Regional rules on AI disclosure and human escalation have to be built in by the customer. Guardrails should be treated as a budget line item too, since each one adds tokens and latency.
The bigger test for NiCE is execution. The company has to get an easier Cognigy into the hands of thousands of CXone customers without losing the governance that made it trustworthy. If it manages that, the acquisition moves from an add-on to a core part of the platform.
Like Databricks’ always-on AI coworker, it is part of a wider shift from AI that demos well to AI that works unsupervised.
The contact centre is where AI agents meet reality, and reality is not impressed by benchmarks. It is impressed by agents that work, every time, for every customer, at two in the morning.
