Deterministic AI Sets the Roadmap for Safer Communications, ICA AI Podcast. Rather than sending every word of every conversation into a large language model, Christensen describes a model where much of the decision-making is based on known patterns, trusted relationships, keywords, context, policy, and call behavior. In sensitive verticals such as financial services, healthcare, legal services, and government, that can be especially important because communications may involve private data, personally identifiable information, account details, medical information, or other sensitive content
By Doug Green
“As AI gets more powerful, the question is not simply whether it can answer a prompt. The question is whether it can be trusted in the communications path,” says Gerry Christensen, associate founder of ICA AI. “For high-security communications, deterministic AI is not just different. In many cases, it is necessary.”
In this Technology Reseller News podcast, Gerry Christensen of ICA AI joins Doug Green to define an important distinction that is becoming central to the future of AI-powered communications: probabilistic AI versus deterministic AI.
The conversation is less about a single product announcement and more about setting out a roadmap. Christensen explains why most people experience AI through probabilistic systems, including large language models that generate answers based on patterns, probabilities and prompts. Those tools can be powerful, but they can also hallucinate, miss context, or create outputs that sound confident while being wrong.
For communications providers, MSPs, UCaaS providers, MVNOs and telecom resellers, Christensen argues that this distinction matters because voice networks are entering an era where AI will be used on both sides of the call. Legitimate businesses will use AI in contact centers. Bad actors will use AI to scale fraud, spoofing, robocalls and deepfake-style attacks. Consumers and enterprises will increasingly need AI to help determine which calls should get through, which calls should be challenged, and which calls should be blocked.
ICA AI, short for Intelligent Communications Assistant, is built around that problem. Christensen describes the platform as an AI-based assistant that can support outbound calling and, perhaps more importantly, inbound call handling. The goal is to allow trusted calls from colleagues, friends, family and legitimate businesses to pass through, while filtering unwanted or suspicious calls.
The core idea is determinism.
Rather than sending every word of every conversation into a large language model, Christensen describes a model where much of the decision-making is based on known patterns, trusted relationships, keywords, context, policy and call behavior. In sensitive verticals such as financial services, healthcare, legal services and government, that can be especially important because communications may involve private data, personally identifiable information, account details, medical information or other sensitive content.
Christensen gives the example of a financial services call. A probabilistic AI system might need to listen broadly and process the conversation through an LLM to determine intent. A deterministic system, by contrast, can look for specific markers of trust or risk: whether the caller is known, whether the call matches expected behavior, whether suspicious phrases appear, or whether the interaction moves toward unusual requests such as gift cards, new account instructions or other red flags.
That approach, Christensen says, also has implications for cost, latency and scale. If most decisions can be made deterministically, the system does not need to rely on a distant AI data center for every interaction. That can reduce exposure of sensitive data, lower dependency on token-heavy AI processing, and support faster call-handling decisions.
Christensen says ICA AI’s approach relies on deterministic AI for roughly 85% to 95% of transactions. He connects that idea to Zipf’s Law, the linguistic principle that a relatively small portion of language often carries much of the meaning. In communications, that means many call-handling decisions may not require open-ended AI interpretation. They may require the right data, the right rules, and the right deterministic understanding of what matters in the moment.
The roadmap Christensen lays out is not anti-LLM and not anti-probabilistic AI. Instead, it is a layered model. Probabilistic AI can still be used when needed, especially when a conversation falls outside known patterns or requires deeper interpretation. But for high-security, high-volume communications, Christensen argues that deterministic AI should carry more of the load.
For MSPs, channel partners and telecom providers, the message is direct: AI call management may become a new category of value-added service. As agentic AI increases the volume and sophistication of automated calls, enterprises and consumers will need tools that can help them determine whether a call is authentic, legitimate and safe.
Christensen compares the coming environment to an arms race. AI will make fraud more scalable, but AI can also make communications more defensible. The providers that begin testing, integrating and understanding these capabilities early may be better positioned to offer customers a practical answer to a growing trust problem in voice communications.
“Everybody is going to need to have an AI-based solution for consumers to handle inbound calls,” Christensen says. “In the world of agentic AI, it is conceivable that networks could be plastered with AI-generated calls.”
Learn more: ICA AI: https://icai.ai/