Agentic AI Represents an IVR Revolution

The world learned about generative AI just over two years ago, and it has consumed the tech sector since. AI and especially generative AI are unlike other modern inventions. Most technologies are tools that assist us in accomplishing tasks faster, better, and cheaper. Tools like the shovel have evolved over the years to more powerful earth-moving devices, but they are still tools designed to assist us. Generative AI can be a tool and, more significantly, an agent. Rather than simply assisting us with our tasks, agents can complete them independently.

The idea that generative AI can automate more tasks is fueling the excitement. So far, we have mostly used AI and generative AI as tools. AI is a predictive engine, and generative AI is savvy enough with language to summarize and evaluate our conversations. It’s a powerful tool that has offered incremental improvements in existing processes.

Generative AI can do more. It can be trained to act on our behalf. It can make decisions and interface with other systems. That’s the notion of Agentic AI. There’s tremendous excitement (and hype) about using agentic AI to automate tasks more than we have seen before.

“Agentic” is the latest buzzword that’s so new it hasn’t found its way into (Webster’s) dictionary yet, but it’s all over the web. Several CCaaS providers, including Cisco, Five9, and Talkdesk have announced Agentic AI-powered solutions, and others are demonstrating use cases external to the call center.

Agentic AI refers to a system or program that is capable of autonomously performing tasks on behalf of a user or another system. Like other agents, it can act for or in place of another person or system. It seems new, but the use of the term agent is familiar. A travel agent is not an airline employee but can act in place of one by providing assistance and tickets. Agentic agents have “agency” to make decisions, take actions, solve complex problems, and interact with external systems. At least, that’s the theory.

Last year in Enterprise Connect keynotes, we saw that if one was double-booked, they could skip one meeting and use their AI assistant to catch up. It could summarize the content and action items. Assistants are tools, not agents. The assistant could cover a meeting, assuming one doesn’t need or want to contribute to it.

Agentic AI is the latest AI breakthrough, but realistically, it will have a limited impact because most of us are not ready yet to give it agency to represent us. We all know modern chatbots can’t be fully trusted. Agentic AI is built on generative AI, a poorly understood technology. The architects of the generative AI models cannot explain how or why these models do what they do. As a result, we need to build “guardrails” to limit the technology. This is primarily accomplished today by restricting them to internal data for knowledge and training, but even this approach is not foolproof. Most businesses will hesitate to give these bots agency to negotiate prices, accept refunds, or essentially any action that requires judgment. 

That’s why the net result of agentic AI, in the near term, will be similar to an interactive voice response (IVR) system. IVRs became popular in the late 1980s. They didn’t understand speech, so we had to communicate with them in simple touch-tone code. You can ask directions, store hours, or any other question as long as it exists as a menu option. If a menu branch matched your question, there was a good chance the inquiry could be resolved without a human agent. You could even do complex data entry, such as touch-tone course registration at a university.

Technically speaking, agentic AI is nothing like an IVR, but the result or impact to the customer will be very similar. I’ve seen several Agentic AI examples, and it’s impressive just how far the IVR has evolved in just 40 years. We can interface with agentic IVRs using natural language (spoken or typed) instead of touch tones. The whole conversational experience is significantly improved, and modern intent detection eliminates the need for static menus.

IVRs were complex to set up. They are rules-based systems that require explicit code for every possible interaction and response. In contrast, agentic AI solutions leverage machine learning to identify patterns and relationships within the data. These agents can respond to a wider range of inputs without explicit training. The breakthrough with agentic AI is less about what it does and more about how it’s setup.

Because we can’t fully trust these agents, they will be severely restricted in what they can do. They will be effectively limited to well-documented processes that don’t require judgment. A potential example is automating a return policy that involves issuing an RMA number. If the criteria, such as the number of days, are met, it will issue an RMA number. Essentially, we will limit these bots to if-then types of actions. As a result, they won’t do anything more than what an IVR did 40 years ago, but without the touch tones.

For 2025, the big gain that Agentic AI will deliver is faster programming. All that’s required is to upload the document(s) with the rules. I saw one demo that uploaded a separate document of allowed exceptions. That’s key; there are always exceptions to the rules, but they usually require human interpretation and judgment.

Why am I returning it after the two-week return window? Because of holidays. Because I was injured. Because my cat ate the return policy. Because I already returned it once, and the replacement broke. Because I only received it yesterday due to a storm. It’s very difficult to document every exception. Even with internal data (no Internet), we know that generative AI will make compelling untruths when it doesn’t have a direct answer within its training data.

There is a persistent hope or fantasy that AI will do everything we can imagine. Keep dreaming. This is why we are continuously disappointed with AI. Today’s generative AI is incredibly powerful and useful but in very limited ways. I keep poking and prodding, and it does amaze and impress, but only about half the time. I guess the same can be said for IVRs. We don’t think about them when they work seamlessly. We know intuitively when an iVR can’t get the job done. We haven’t reached that phase yet with generative AI.

Generative AI has limited memory, no ability to reason or plan, and its empathy is programmed. Agentic AI solutions will be unable to respond to complex, emotional situations. These systems will be vulnerable to attacks and data breaches, and will be difficult to troubleshoot. Agentic AI will not have human-equivalent skills in problem-solving or negotiation. Implementing agentic AI could worsen overall satisfaction scores if skilled human agents leave.

Perhaps I am too critical. The demos have indeed been impressive, but oddly few of these products have actually been released yet. I do believe Agentic AI has a lot of potential, and expect it to rapidly evolve. For now, viewing them at IVRs is recommended. Marvelous, advanced IVRs. Take the win, and keep implementations limited to simple rules-based scenarios.

Dave Michels is a contributing editor and Analyst at TalkingPointz.