Alon Eirew, CTO (AI) of MCE Systems – Interview Series

Alon Eirew, CTO (AI) of MCE Systems, leads the company’s AI strategy and end-to-end technology stack, spanning architecture, model development, infrastructure, product integration, and the broader adoption of AI across internal operations. He previously spent more than nine years at Intel (INTC ), progressing from senior software engineer and software architect to Lead Natural Language Processing (NLP) Research Scientist, where he directed knowledge-extraction research, secured funding for multi-year projects, published papers at major NLP conferences, and mentored startups through Intel Ignite. Earlier in his career, Eirew helped develop personal-assistant technology at Telmap that later became part of Intel, and he has since advised early-stage AI companies while teaching computer science, object-oriented programming, data mining, and information retrieval at several Israeli universities.

MCE Systems develops an AI-native experience layer for telecommunications providers, connecting legacy systems, real-time data, and intelligent workflows to support customer interactions across multiple touchpoints. Founded in 2005, the company built its business around digital device lifecycle management, with technology covering device diagnostics and care, trade-in programs, remote device grading, and returns. MCE is increasingly applying generative and agentic AI to these workflows, helping operators detect device problems, guide customers through potential resolutions, and coordinate more complex service journeys across digital and assisted channels.

You spent nearly a decade at Intel working on NLP, knowledge extraction, information retrieval, and even early personal assistant technologies that originated at Telmap before becoming part of Intel. Looking back, which ideas from those early AI systems turned out to be ahead of their time, and which assumptions did the industry get completely wrong?

One idea clearly ahead of its time was the personal assistant we worked on before the LLM era. These were systems that understood intent, connected context across tasks and helped users accomplish goals naturally. These had the same ambitions of today’s agents. The difference was the underlying approach: Rather than language models, we relied on a combination of machine learning and algorithms to infer user intent. This method, though, proved difficult when applied at scale considering the different ways users express the same request.

Imagine employing a service rep to handle thousands of customers’ inquiries, but they can only handle inquiries that are framed with specific sentence structure. As you can imagine, it would only cover a rather small range of use cases and it would be hard to handle many customer issues. LLMs completely changed everything. Today, understanding the user intent has been largely solved. So in our analogy, the rep is now flexible and can speak to millions of customers.

While I wouldn’t claim the industry got something completely wrong, I do believe many underestimated just how central language models would become. The major leap came when the first language models like ELMo, BERT and RoBERTa were introduced. At the time, few truly understood the gravity of the transformation that was on the way. These were the backbones models of the modern LLMs we use every day that we can hardly imagine not having. 

Most companies saw these initial models as useful components in narrow applications, not as the foundation for an entirely new computing paradigm. In hindsight, many of them missed the opportunity to position themselves as major AI players.

During your research career, you focused on multi-document knowledge extraction and event understanding. As enterprises increasingly rely on large language models, why do you believe structured knowledge extraction remains such a critical challenge?

LLMs are powerful tools to address an array of topics, but enterprise applications and end users need expertise to solve a specific issue, which requires access to knowledge. This contextual proprietary knowledge isn’t typically a model’s training data, so deploying an LLM in an enterprise means gathering understanding across potentially millions of internal data points. Today’s LLMs largely struggle with this, missing substantial context due to technical and access constraints, and that gap limits their impact.

At MCE, for example, we provide a customer care AI agent solution. One of its goals is to allow telecom carriers to help their customers solve technical mobile device issues without needing human intervention. Without feeding the AI agent contextual technical information about the device, all the AI agent would do is provide generic results that don’t solve the issue or are irrelevant and then force the customer to seek human help. This is the kind of knowledge extraction issue we had to overcome in order to deploy AI with our customers and it’s where organizations often fail in getting the results they want.

My own research dove into knowledge extraction and information retrieval from large document collections, exploring the relationship between “events,” a term in the discipline of AI used to define the underlying causality and timeframe of any kind of representation of information. We apply this at MCE within our AI solutions for human interaction by building a structural understanding that can help LLMs better summarize, represent and search knowledge to solve a customer’s problem. 

You’ve published research at top NLP conferences including ACL, EMNLP, and NAACL. What breakthroughs over the past five years have had the biggest real-world impact, and which highly publicized advances do you think are currently overhyped?

There are a few breakthroughs that stand out to me and they happen in a progression as LLMs developed.

The first was around instruction tuning and prompt engineering. This had to be solved first before reaching much wider adoption. We’ve all been there where you ask an LLM for something specific and it goes in circles, getting close but not getting you the response you need. So figuring out how models could understand the prompt and deliver accurate results was a first big step.   

By extension, the next breakthrough was for agents, which needed LLMs to go beyond just one-off prompted tasks and actually do a few things: create plans, interact with an end user, use tools, use memory and access knowledge. This one was essential to where we are today.

But the next breakthrough, which is only partially solved but in progress, is completely solving the hallucination issue. So far, there’s been plenty of progress here to the point where we can rely on LLMs much more than before. But there’s still work to do and another reason why human supervision, and guardrails remains essential.

What’s overhyped is the degree of autonomy people are claiming agents should have. Fully autonomous systems operating across complex business environments without human supervision are still not reliable enough in my view. What I find more productive is designing AI to perform high-value tasks, escalate uncertainty and explain its reasoning, without ceding human control of the overall flow. That’s where agentic AI will create the most real-world value in the short term.

You have worked as a researcher, software architect, engineering leader, startup advisor, lecturer, and now Chief AI Officer. How has your definition of successful AI changed throughout those different stages of your career?

Throughout my career I learned that the definition of successful AI continues to evolve and the different core values are across disciplines. Applied research and architecture taught me that AI must work under real-world constraints. Academic research taught me the importance of rigor and evaluation. Engineering leadership showed me that execution and operational discipline matter as much as the model itself. Advising startups taught me that AI creates the greatest impact when it solves a real problem and has a clear path to adoption.

Today, as AI chief at MCE, I define success as AI that creates measurable value while earning trust of the end user. It needs to improve outcomes for customers, employees and the business, be reliable enough for production, transparent enough to be governed and useful enough that people actually adopt it. In that sense, my definition has evolved from “Does the technology work?” to “Does it create value responsibly, repeatedly and at scale?”

This is also the transition we made at MCE, one that started two years ago and led to here now. The latter value is how we think about AI now for internal use cases and for our customers in telecommunications.

MCE Systems is applying AI for telecoms customer experience, customer care and marketing. What are the most compelling AI use cases currently emerging within the telecommunications industry?

Communication service providers (CSPs), or “carriers” as they’re usually called, are in a tough spot, where they’re facing a commoditization problem. Price is really the attracting force to bring new customers or retain them, but it isn’t sustainable. If you look at other giants who sell consumer products or services, like Apple (AAPL ), Amazon (AMZN ) or Netflix (NFLX ), they built a great product and an experience around it. This played out clearly in finance, for example, where banks were forced by new agile players to move from holding peoples’ savings and offering investment products to also offering a full digital experience around it. Same with telecoms.

So our vision is to help bring additional value to CSPs’ customers with a better experience on top of the great connectivity product they already provide. This means being more proactive and communicative with customers about their overall network and device experience. And for us, AI is what finally allows them to do it well.

In the industry, though, it’s become mostly a defensive tool to cut call center labor costs or make employees more productive. But for us, AI is also an offensive tool to be proactive with customer care and personalize service and what you offer customers, all on the customer’s mobile device. 

For example, catching a device or connectivity issue or battery degradation problem before the customer notices it and helping them fix it. Those kinds of experiences build genuine loyalty. Knowing a customer’s situation well enough to offer the right broadband bundle, data add-on or lifestyle product at the right moment drives higher conversion. And pulling together signals across network assets, CRM and device data to accurately predict churn, then acting on it early, is perhaps the most powerful use case of all of them. CSPs that get this right will have a meaningful edge.

There’s plenty of hype around becoming “AI-native” or “AI ready.” How is MCE approaching this paradigm and what do you believe are the most important points organizations miss in this transformation?

What many miss is that AI transformation isn’t just a technology deployment, but also a modus operandi change. The best opportunities are where adoption feels natural: the workflow already exists, the pain is measurable and AI can clearly improve speed, quality or decision-making. MCE is approaching this paradigm with a few ideas.

We started by asking about AI tool maturity and adopting AI in ways that multiply employee productivity. We asked how we could educate our teams, define clear usage guidelines and connect internal systems to AI where it actually could make a difference day-to-day. 

Once we assessed our base of tools, we looked at workflow automation and identified where agents could reduce manual work. Naturally, we found we could drastically cut down time to development. In one example, a highly functional, prototype app took a couple of days to create when it otherwise would have taken two or three weeks.

What’s often underappreciated for internal productivity is data democratization. We worked on and continue to work on making better use of our internal data so both humans and AI systems make better decisions. This includes understanding customer needs, improving operations and building products that learn from past experience. It’s still a work in progress, since it’s an iterative and analytical process to evaluate what works and what doesn’t.

Many organizations are racing to deploy AI agents. What practical lessons have you learned about where autonomous AI can genuinely create value versus where human oversight remains essential?

It’s easy to get caught up in the automation hype. But I believe in approaching agents methodically. I’d advise to never start with a target level of automation; instead start by identifying workflows where agents can create value, break them into smaller pieces and ask which of the pieces can be fully automated, which semi-automated and where human judgment is needed.

In practice, we find that agents create the most value when they’re used in highly-defined tasks, which ranges from gathering information to analyzing inputs to generating recommendations to preparing outputs for the next step. Fine-tuned automation where the task is simple enough works far better than trying to automate an entire complex process.

Where humans remain essential is in oversight, exception handling, customer-sensitive decisions and, generally, situations where accountability matters. The goal is never to replace human decision-making, but to let agents handle the parts where they can reliably create value while letting humans handle the decisions that require accountability.

MCE has spoken about using AI to deliver more personalized device experiences and improve operational efficiency for mobile operators. How do you balance personalization with growing concerns around privacy, transparency, and user trust?

Because AI feels so human, users naturally interact with it in ways that expose more than they sometimes should. That puts real responsibility on organizations to be vigilant and transparent about what data is collected. A simple example: when a customer submits a screenshot as part of a troubleshooting process, that image may unintentionally capture personal information on their screen. It’s imperative that organizations apply safeguards to crop the image to remove excess visible content, keeping the focus on exactly what the interaction is about and reduce unnecessary data exposure.

To me, any data we collect should clearly serve the customer’s interests and needs. It should be collected only when it supports a specific purpose, such as resolving the customer’s inquiry or providing an appropriate degree of personalization. At the same time, the data should be handled in a privacy-preserving manner, avoiding any unnecessary linkage to the customer’s real-world identity.

In other words, understand the customer profile’s specific needs as relates to your business. In our case, as a privacy compliant organization, it’s about knowing the mobile device’s condition and user behavior as it relates very specifically to the provider, not about the customer’s identity and other kinds of behavior.

What also matters is in the interaction with the customer itself. Guardrails are critical here to stay within purpose, brand voice and really avoid drifting beyond what organizations are ready to govern. In practice, guardrails help avoid legal or regulatory violations and negative customer experience. Sometimes, these can be grueling processes, but they’re essential.

You have taught subjects ranging from computer science fundamentals to information retrieval and data mining. If you were designing an AI curriculum for students entering the workforce today, what topics would you prioritize and what outdated concepts would you remove?

The curriculum needs to change in two ways. Topics once limited to graduate programs, in the likes of neural networks, transformers, large language models, retrieval-augmented generation, evaluation, alignment and responsible AI need to become core undergraduate subjects. Students entering the workforce need to understand how these systems work and how to evaluate them critically.

Students also need to learn how to use AI responsibly as a tool for learning and research. That means questioning outputs, verifying sources, understanding limitations and treating AI as a thinking partner rather than a shortcut to do homework.

What I wouldn’t cut is mathematics and algorithmic fundamentals. I’d argue they’re as important as ever. What I would reduce are assignments that reward mechanical implementation without demonstrating critical thinking. The focus should shift toward problem formulation, evaluation, system design and responsible use.

Having advised numerous startups on AI strategy, what are the most common mistakes founders make when building AI products, and what separates companies that successfully move from prototype to production?

The most common mistake I see is falling in love with the technology before really understanding the problem, your customer or the path to adoption. A close second is not knowing AI’s limitations well enough, so forcing it into use cases where it just isn’t mature enough yet.

What separates founders who get to successful production is deep domain expertise and a research attitude. The best AI products come from founders who understand a specific vertical well enough to identify a real technology gap. They’re not starting from “what can we do with AI?” but from “here is a specific problem, here is why existing solutions on the market fall short and here is why AI can now solve it sustainably.”

Moreover, successful founders constantly learn from customers, competitors, advisors and failed attempts. They study how decisions are made, how budgets move and what creates trust for their target audience. At the end of the day, they’re not just building a product, they’re building something that needs to survive in a real environment and not for its own sake.

Looking ahead five years, do you believe the biggest impact of AI will come from more powerful foundation models, agentic systems, multimodal AI, on-device intelligence, or a completely different trend that the industry is currently underappreciating?

The biggest impact won’t come from foundation model improvements alone, but from how models are embedded into systems that can observe and act within real workflows while also improving. That shift from powerful model to capable system is where I expect the most meaningful near-term progress.

This will become especially strong when systems go full multimodal, reason over text, voice, images, video and other contextual real-world signals can closely emulate how humans understand the world. This will open up new possibilities in customer support, healthcare, manufacturing, etc.

Another emerging technology that I believe will become disruptive  are agents that interact with the physical world. Companies like Boston Dynamics are already showing the art of possible and the potential of when perception, reasoning, planning and physical action converge. Whether in humanoid form or specialized robotic systems, agents with bodies may prove to be one of the most tangible and revolutionary expressions of AI we see in the next five years. For MCE, this creates opportunities in logistics and operational workflows, where many repetitive, low-complexity tasks still require physical human interaction. Examples include picking up a device, pressing buttons, connecting cables, or moving devices between stations. These are ideal candidates for automation through robotics and AI that allow systems to be more efficient, consistent and handle much more volume.

For telecommunications, being able to predict customer churn and act on it with AI models is where the money will be in the next few years. If you can build a reliable model that incorporates data from all the different customer points, can accurately evaluate churn probability and then act on it, you will have solved a stagnating issue for providers.

Thank you for the great interview, readers who wish to learn more should visit MCE Systems.