1970s: “That’s not real AI – that’s just heuristics. Real AI is an expert system.”

1990s: “That’s not real AI – that’s just an expert system. Real AI is machine learning.”

2010s: “That’s not real AI – that’s just machine learning. Real AI is neural networks.”

2025: “That’s not real AI – that’s just neural networks. Real AI is LLM models with RAG and MCP tools.”

I’m sure we’ve all heard or even said this kind of thing. It’s not easy to separate the goal of artificial intelligence from the latest implementation. We get swept up by the hopeful promise of each new technology wave, chasing the hype instead of the outcome. And because we struggle to crisply define human intelligence, we don’t maintain clear sight of the goal of artificial intelligence. It’s easier to say: “I’ll know it when I see it.”

That’s why the latest advancements with generative AI have so many industries excited. Tools like ChatGPT make AI approachable to the general populace in a way that no previous AI technology has. We know it because we see it. However, even though we think we’re seeing AI in action, the fact remains that 95% of projects using generative AI have yet to produce a return.

Don’t get me wrong – generative AI, and especially agentic AI, have me very excited about the future. But at the rate things have been moving lately, I’m sure next month there will be some new technology standing on the shoulders of these giants, reaching even greater heights. 

How can you keep pace with the relentless evolution of AI and still create value?

Real artificial intelligence doesn’t specify a technology; it describes humankind’s aspiration for innovation. We want to take human roles, human tasks, human processes and have systems that perform them. You could say that any innovation throughout history has had this goal. Real AI must embody the hallmarks of human intelligence, regardless of the underlying technologies. 

As we consider various human processes, we can boil down human intelligence into some key characteristics: whether that involves driving a car, assembling fire extinguishers in a factory, developing front-end GUI software or troubleshooting a subscriber’s internet. These processes require human expertise and collaboration to be performed accurately and safely. The proper action requires first sound and informed reasoning. However, only experts who have learned their craft have the basis for reasoning through related tasks. Experts explain the output of their work to other participants in the process, as well as teach novices to perform their role in the future. REAL human intelligence bears these four interrelated characteristics: Reasoning, Explaining, Acting and Learning.

REAL artificial intelligence must also bear these same hallmark characteristics of intelligence. REAL AI Reasons like a human, Explains its reasoning in human terms, Acts on that reasoning and Learns through this process to improve its reasoning.

In the past, AI technologies performed limited reasoning, explaining and acting with moderate success. But they’ve struggled mightily with scaling learning. Expert systems represented human expertise remarkably well, but harvesting that expertise met with insurmountable challenges. Machine learning and neural networks found more success modeling physical systems, but they scale with difficulty due to the massive number of samples required to train them.

What if, alongside the reasoning, explaining and acting provided by AI, we could also have it natively learn from experts?

Industries like the automotive industry are figuring this out – and they are establishing strong patterns the rest of us can follow. I still remember the days when I drove, completely unassisted (yep, dating myself a bit there). But then cars introduced blind spot and lane departure warnings, making me a better driver. These capabilities assisted me, but I still drove completely on my own. Some car makers used these new capabilities to collect data about traffic conditions and my behavior, which in turn enabled them to provide capabilities like lane assist and adaptive cruise control. Rather than a mere assistant, the car became a copilot and continued learning how to take on more of my tasks.

And now we’re seeing the fruit of this progressive learning – cars are driving on their own. Automated agents weren’t introduced into the driving process on day one, but artificial intelligence was first employed as an assistant to the expert, then as a copilot to the expert, and then finally as the expert driver. This gradual evolution was about developing trust and meeting drivers where they were comfortable. Each phase allowed humans to gain confidence while the artificial intelligence learned how to reason, explain and act better, ultimately taking on a human’s role in a human’s process.

So, let’s bring this home to telecom and focus on a process where a phased introduction of REAL AI produces operational efficiency gains while also delighting subscribers: customer support.

Subscribers, frontline technical support, tier 2 technical support, network engineers and technicians each play a unique role within the customer support chain. Yet, much like the days when I drove my car with no AI, many service providers today operate customer support with no REAL AI – stuck in a reactive support mode. A subscriber calls in with a problem, the tech support answers and scrambles to pull data from many systems, tries a few fixes, and then either escalates or rolls a truck. Even with their expertise and ability to get the right information, they often make the wrong decision, costing the company money and frustrating subscribers. 

Adtran is changing that. Our AI-powered assistants (like Mosaic One Care, Mosaic Subscriber Insights and Intellifi®) initially just pulled data from disparate systems into a single, holistic view. But then they started explaining what might be degrading the subscriber’s experience and directing actions to remediate the issue. The result: responsive support – support that still reacts to the subscriber’s call but directed towards making the best decision. 

Here’s the exciting part: just like blind-spot warning helped the car company train its AI to do more advanced things, we use this data to improve our troubleshooting reasoning. This has led to products like Mosaic One Operate and now, Mosaic One Clarity, which guide a proactive approach to enhancing the subscriber experience. They identify subscribers who have a degraded experience before they call and pinpoint the root cause and remedy.

As we look forward to a customer support world characterized by autonomous optimization, I can’t help but take a detour into the latest fad in AI technology: agentic AI. If you’ve used ChatGPT to any extent, you’re half the way to understanding the concept. Simply put, an AI agent utilizes a combination of an LLM (ChatGPT uses a large language model) for reasoning, specialized system prompting and a set of tools to interact with specialized knowledge sources and external actions. Effectively, an AI agent provides an LLM with data beyond its training set and the ability to interact with the outside world, allowing it to specialize in a given human task. Not only do agents perform the reasoning, explaining and acting of human intelligence, but they also hold the promise of solving the problem of learning from experts by utilizing them. It’s really this latest aspect that has the industry in such a tizzy – the applications are countless. 

But this technology comes with a wealth of cautionary tales – tales of hallucinations, costly mistakes and even seemingly willful deception. Today, unleashing a legion of agents to automate every workflow in a telecom provider’s operations spells nothing but disaster – and high costs. Adtran’s approach follows that of the car industry: a phased introduction of REAL AI, using a coalition of the best technologies for each job. So, today, our core troubleshooting root cause analysis engine powering our Mosaic One and Intellifi products is built on consistent, cost-effective and understandable algorithms. We use agentic AI to assist our domain experts in constantly improving our root cause AI. We are beginning to use agentic AI to support our tech support team in more easily troubleshooting issues and answering questions. And we recognize that Adtran won’t provide all the agents you want to build, so we’re introducing our REAL AI Factory – a platform purpose-built for agentic development that opens up all of our data and intelligence, whether from Adtran or third-party equipment.

I don’t know what the next generation of real AI technology will be tomorrow or next year, but I can guarantee you that it will be more transformative than the technology we have today. But if you want to improve your operational efficiency and delight your subscribers today, you can join us in phased transformation now with REAL AI. Built to reason. Designed to explain. Ready to act. Always learning.