Leadership Spotlight
How Do We Think AI with Diglio Simoni
The most important question in AI isn’t what the technology can do. It’s what people need to decide. The best systems don’t replace human judgment, they give it better information, better context, and better options.
A Conversation with Diglio Simoni, Vice President and Distinguished AI Scientist
Few careers span as many disciplines as that of Diglio. He has spent more than three decades exploring some of the biggest questions in science, technology, and human knowledge. His journey has taken him from high-performance computing at NASA and research in neuroscience and genetics to enterprise AI, knowledge systems, and agentic technologies.
Yet despite the diverse career path, a common theme runs through his work: understanding how knowledge is created, structured, and transformed into action. Today, as organizations race to operationalize AI, Diglio brings a perspective shaped by both scientific inquiry and real-world business execution. His focus is not on building smarter models for their own sake, but on helping organizations create systems that are explainable, accountable, and capable of supporting better decisions.
Your career has taken you across physics, computer science, neuroscience, epidemiology, and AI. Was there a common thread connecting all these disciplines?
Yes. When I was young, my father told me to choose a career by asking the biggest question I could imagine. Mine was: how do we think? Everything followed from that.
Understanding billions of neurons meant learning to make many processors work together, which is how I ended up at JPL on the Caltech/JPL Mark III Hypercube inventing parallel methods for processing SAR image data from the Magellan mission to Venus. Understanding perception led me to single cell recordings on awake macaques and visual psychophysics in humans using infrared eye trackers, and then to NASA Ames Research Center, where I ran the experimental setup for the NAS Virtual Wind Tunnel where we studied computational fluid dynamics simulation results. Early on I studied genetics, because the way life stores information continues to fascinate me. That paid off years later at RTI International. There, as part of a bioinformatics program, I led an informatics center that helped scientists in two cancer research consortia find and reuse each other’s work, and I helped move a national epidemic simulation onto the TeraGrid.
The fields look different on the surface, but they share one thread: how knowledge is created, represented and turned into decisions. Thirty-five years on, I’m still working on a version of that first question. When does a signal acquire meaning? How do we know what we know?
You have worked in both research institutions and enterprise environments. What have you learned from moving between those worlds?
In research, being wrong is a result. In a bank, being wrong costs money, and often draws a regulator. Working at Citibank, Credit Suisse, T-Mobile and Apple taught me that “why did the system do that?” matters as much as “is it right?” Research taught me to ask questions nobody else was asking. Industry taught me that an answer nobody can adopt isn’t really an answer.
You joined Evalueserve at a time when enterprise AI is moving rapidly from experimentation to deployment. What excites you most about this moment?
The models are rarely the bottleneck anymore. Knowledge usually is. The technology has advanced remarkably, but the real challenge now is operationalization. How do you combine AI with workflows, expertise, governance, and business goals?
Most organizations hold what they know in documents, spreadsheets and people’s heads, in forms an agent can’t reliably use. Evalueserve is a knowledge company, so that gap is exactly where we sit. Much of my work now goes into structuring that knowledge, and into laying a path forward so that our own engineers can build agentic systems that hold up in production.
Your recent work has focused on Agentic AI and Knowledge Graphs. Why is that combination important?
I never try to sell a knowledge graph for its own sake. When I sit down with a client and list the questions they actually want answered, almost all of them join the same few things: products, customers, suppliers, markets. The graph is simply that structure made explicit. An agent without it is guessing across disconnected data. An agent with it can reason over real relationships, and it leaves a trail you can check, and possibly even reuse later. An agent is only as good as what it can look up, and what it can be held accountable to.
You've spent decades working with advanced technologies. What misconceptions do organizations still have about AI?
In 2007 I used genetic programming to evolve decision rules for where people look when they search a scene. The rules were readable. You could argue with them. I’ve wanted that property in every system since.
The biggest misconception in the field of AI is that fluency means understanding. A model that writes like an expert isn’t an expert. The most valuable thing people bring is the thing a machine wouldn’t think to do: the reframing, the judgment call, the question nobody asked. The best systems take the routine work off people’s plates so they have more time for that.
Looking back over four decades of technological change, what has surprised you most?
The pace of change is remarkable, but what’s perhaps more surprising is how many of the fundamental questions remain the same. How do people learn? How do we represent knowledge? How do humans collaborate effectively with machines? The tools keep evolving, but those underlying questions continue to shape the field.
Decades ago I worked on hypercube machines, trying to get many processors to cooperate on one problem. Today every large language model runs on a descendant of that idea. The hardware changed almost beyond recognition. The questions didn’t: how knowledge is represented, how learning happens, and how people and machines should share the work.
What role do you see scientists and AI leaders playing in the years ahead?
I think the responsibility is larger than building models. We need to help organizations understand what technology can realistically achieve, where it should be applied, and how it can be implemented responsibly.
Part of the job is saying “not yet” or “not like this.” Trust comes from being clear about limits. Every system that acts in the world also needs someone who is answerable for what it does. That’s an engineering requirement, not a public relations one.
Your background spans everything from fundamental science to enterprise applications. Does one perspective influence the other?
Science taught me humility, because nature is always more complicated than your model. Enterprise taught me to start from the decision someone has to make, not from the technology. My pilots are small, gated, and measured against that decision.
The most successful innovations respect both perspectives simultaneously.
Finally, what would you like colleagues and clients to know about your approach to AI?
I’m optimistic about AI, but I’m even more optimistic about people. The goal isn’t to build smarter machines for their own sake. It’s to help individuals and organizations achieve outcomes that weren’t previously possible. In practice that means starting from what a person needs to decide, and building the system around that.
After four decades spent exploring the frontiers of science, computing, and artificial intelligence, Diglio remains focused on a surprisingly simple idea: technology is only valuable when it helps people make better decisions. He believes success will depend less on model sophistication and more on how effectively knowledge, context, and human judgment are woven into intelligent systems.
Latest Posts
Beyond Traditional Lending: The Rise of Specialized Credit Solutions in Modern Banking
Explore how specialized lending is reshaping modern banking, from private credit and asset-based lending to evolving credit risk assessment.
Building a Repeatable IP Licensing Business: Insights from IP Stars Korea
Many companies have closed a licensing deal, but few have built a function that keeps producing them. Four questions from IP Stars Korea show how to make the shift.