Yu Su, Co-founder and CEO of NeoCognition – Interview Series

Yu Su, Co-founder and CEO of NeoCognition, is a longtime artificial intelligence researcher whose career spans academia, enterprise AI, and advanced agent systems. In addition to leading NeoCognition, he serves as an Associate Professor and Innovation Scholar at The Ohio State University, where his work has focused on AI agents, reasoning systems, and machine learning. His background also includes more than six years as a Senior Researcher at Microsoft, where he worked alongside leading researchers including Percy Liang on conversational AI systems for Outlook using semantic parsing techniques. Across roles at Microsoft, academia, and research institutions such as IBM T.J. Watson Research Center, Yu Su has built a reputation for advancing AI systems that can reason, learn, and interact with complex digital environments, helping bridge the gap between cutting-edge research and real-world applications.

NeoCognition is an AI agent research company focused on developing what it describes as “specialized intelligence” — AI systems that continuously learn and improve through experience. Founded by leading AI researchers, the company is pursuing a vision that moves beyond static large language models toward agents capable of developing deep expertise in specific domains. Its research focuses on areas such as continual learning, reasoning, planning, tool use, and multi-agent collaboration, with the goal of creating AI systems that become more capable and reliable over time. By combining advances in machine learning with structured reasoning and adaptive learning techniques, NeoCognition aims to help shape the next generation of AI agents capable of tackling increasingly complex real-world tasks.

Many companies rushed to release generic AI copilots, but we are increasingly hearing concerns about reliability once these systems enter real production environments. Why do you believe so many current AI agents struggle outside controlled demos?

Most current AI agents struggle in production because they are still fundamentally generalists operating without a durable understanding of the environment they are working in. They can often complete a task once in a demo, but that is very different from developing repeatable judgment inside a real operational system.

Today’s models are impressive pattern matchers, but they still lack the mechanisms humans use to become reliable experts. Humans do not become dependable by memorizing isolated tasks. We specialize by learning the structure of a particular world: the workflows, constraints, relationships, tools, priorities, and consequences that define a profession or organization. Over time, we build an internal working model of that environment, and that model is what allows good decisions to become consistent and repeatable.

Most AI agents today do not build that kind of operational understanding. They rely heavily on prompting, retrieval, or orchestration layers that help them complete isolated actions, but they are still largely improvising each time they encounter a new situation. That is why performance often breaks down once the environment changes and more so when it becomes messy, dynamic, or high-stakes.

The missing piece is specialization. Humans thrive because we can continually adapt to changing environments and become experts inside them through continual learning. We believe AI agents need a similar capability: the ability to learn the local structure of a domain deeply enough to operate with real proficiency.

NeoCognition has described its vision as building agents that can continuously learn and adapt more like humans. What does that actually look like technically compared to the static fine-tuning or retrieval-based approaches many enterprises rely on today?

Most enterprise AI systems today improve performance either by fine-tuning a model once or by retrieving relevant information at inference time. Those approaches can be useful, but they do not fundamentally allow an agent to develop evolving expertise inside a domain.

Fine-tuning is typically static after training. Retrieval systems help surface information, but retrieving knowledge is not the same as developing domain expertise or adapting behavior based on experience over time. In many cases, the agent still lacks a persistent model of the environment it operates in.

When a human joins a company, they do not become effective simply because they can search documents. They gradually develop judgment by understanding how the organization actually functions and expertise emerges from continuously refining that internal model.

We believe the next generation of agents needs a similar learning mechanism. At NeoCognition, we are focused on enabling agents to form those kinds of evolving operational models so they can continuously specialize and improve within a domain over time, rather than repeatedly starting from scratch or depending on constant human re-engineering.

There appears to be a growing divide between AI experimentation and operational trust. Enterprises may successfully prototype agents internally, but deploying them at scale is another challenge entirely. What are organizations underestimating about this transition?

Many organizations are underestimating how dynamic real operational environments actually are. An agent performing at 85% accuracy may sound strong in testing, but at enterprise scale that still translates into a constant stream of errors and recovery situations that human teams must manage. The challenge becomes even more significant in multi-step workflows where failures compound across systems and tasks, making oversight, intervention, and accountability far more difficult than many organizations initially expect.

Enterprises are still treating deployment as an orchestration or prompting problem, when in reality it is also a learning problem. The difficult part is not just getting an agent to execute a task. It is enabling the system to develop lasting competence and judgment inside a dynamic operational environment.

The burden of customization, prompting, supervision, and re-engineering still falls heavily on human teams today. That’s often a sign the system still lacks operational understanding; they are being manually steered through it every time. That is not a path to scale or trust.

A major theme emerging across the AI industry is governance, guardrails, and policy enforcement. Yet NeoCognition argues governance alone is not enough. Why do you believe reliability ultimately requires systems that continuously adapt to their environment instead of simply following static rules?

Governance and guardrails are extremely important, but static rules alone cannot fully solve reliability in complex environments.

Production-level operational systems are constantly changing. Workflows evolve, tools update, policies shift, and unexpected situations emerge that cannot always be anticipated in advance. If an agent only knows how to follow predefined rules without understanding the environment it operates in, it will eventually encounter situations those rules did not account for.

Humans’ reliability comes from judgement, not just rigid adherence to scripts, but because we develop judgment through understanding the structure and constraints of the world around us. We learn when to escalate, when something looks abnormal, and when context changes the correct course of action.

We believe AI agents need a similar capacity for adaptation and environmental understanding. Safer systems will come from making them more competent and specialized within clearly defined operational worlds. This type of system observes its own environment and own outputs and tracks where it fails and updates its behavior.

The AI industry often emphasizes larger models and broader capabilities, but NeoCognition appears focused on specialization and contextual learning. Do you believe the future of enterprise AI agents will look more like highly specialized digital workers rather than universal assistants?

We strongly believe the future will be driven by specialization. The industry has understandably focused on increasingly general models because broad capability is impressive. But in enterprise environments, the real challenge is not whether an agent can do a little bit of everything. It is whether it can perform a specific role reliably and with sound judgment over time. Our strength is not that we are born experts in every environment. It is that we can learn the structure of a particular world deeply enough to operate effectively within it.

At NeoCognition, we believe the future will not be one super agent that does everything. Instead, it will be an abundance of specialized agents, each one learning a particular world deeply enough to operate with expert-level proficiency, reliability, and judgment. Their purpose is not to replace human expertise, but to make it more abundant: to put frontier-grade capability in far more hands and raise the baseline of what any person or organization can do.

One of the biggest concerns surrounding autonomous agents is how they behave when environments change unexpectedly. How important is real-world adaptation and environmental awareness for preventing failures, hallucinations, or unsafe actions?

It is absolutely critical. Without environmental awareness, agents can continue acting confidently even when their understanding of the situation is outdated or incomplete. That is where operational failures often emerge.

We believe reliability requires agents to continuously learn the local structure of the environment they operate in and update their understanding over time. Reliability changes and evolves over time: what seemed reliable in September may not seem the same way in May. The more deeply an agent understands the systems, constraints, workflows, and relationships around it, the more capable it becomes of recognizing when conditions have changed or when uncertainty requires escalation.

How close do you believe the industry is to deploying AI systems that can genuinely improve themselves through ongoing interaction with real operational environments?

We are still early in building the underlying learning mechanisms required for reliable continual learning and self-improvement  but the industry is much closer to this transition than many people realize. We are living in a compressed timeline. The ingredients for the next technological breakthrough are ready. It can happen quite soon.

What matters is not simply self-improvement in the abstract, but structured specialization inside real environments. That means learning workflows, constraints, relationships, and patterns of successful behavior in ways that are stable, governable, and resistant to catastrophic forgetting. That is the problem we are focused on at NeoCognition.

Looking ahead several years, what do you think will ultimately separate trustworthy enterprise AI agents from the wave of experimental systems currently flooding the market?

The systems that succeed will be the ones that feel less like toy scenarios and more like dependable operators.

Raw model capability alone will not be enough. Enterprises will ultimately care about whether agents can operate consistently within their actual environments, understand local workflows and constraints, respect boundaries and permissions, adapt safely over time, and deliver repeatable outcomes.

The future winners in enterprise AI will not simply be the systems that can answer the widest range of questions. They will be the systems that can learn a particular operational world deeply enough to act with real proficiency, judgment, and reliability inside it.

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