The public debate around artificial intelligence has spent years focused on models: which lab produced the largest one, which benchmark it topped, which company owns it. That framing made sense when capable models were scarce and access was the constraint. It is becoming a less useful lens. Organizations put eleven times more AI models into production in 2024 than in the prior year, and the enterprise AI market has expanded from $24 billion in 2024 toward a projected $150 to $200 billion by 2030. As models proliferate and commoditize, the competitive question is no longer which model a company can access. It is which systems they have built around it, and whether those systems integrate meaningfully into the workflows, interactions, and learning environments where people actually spend their time.
Xiaosi Yang is a Partner at Xstar Capital and a technology executive with 14 years of experience spanning global TMT investment, product strategy, and cross-border capital deployment across markets including China, the United States, the Middle East, and Southeast Asia. A Senior Member of IEEE, Yang has worked across semiconductor platforms, mobile ecosystems, and global carrier relationships before turning his focus to AI system development as an investor and strategic advisor. Through his investment work, he has focused on platforms that move beyond standalone AI models and toward integrated systems that embed intelligence into how people learn, work, and communicate, spanning conversational AI designed for specific communities, AI-native approaches to higher education, and workplace agent tools that reshape how individuals manage information and tasks.
The Model Era’s Inflection Point
Enterprise generative AI spending reached approximately $36 billion in 2025, with the infrastructure layer alone, covering foundation model APIs, training compute, and data orchestration, capturing $18 billion. Meanwhile, 71% of firms reported using generative AI in at least one business function in 2025, up from 55% the prior year. Those numbers tell a story of rapid normalization. Foundation models are no longer exotic assets that confer advantage simply by being deployed. They have become what cloud compute became a decade ago: necessary infrastructure, widely available, increasingly undifferentiated at the commodity tier. The organizations generating durable value from AI are not the ones with access to the best model. They are the ones that have built coherent systems on top of it, systems with interfaces that fit how people naturally communicate, pipelines that connect AI outputs to real decisions, and workflows that make the technology usable without requiring users to understand the model at all.
The distinction matters because it changes what investors, builders, and enterprises should be optimizing for. A model without a system is a capability without a use case. An end-to-end system, one that wraps the model in the right interface, trains the users who interact with it, and integrates it into an existing workflow, is a product. That is where value accretes, and that is where the next phase of AI investment and development is being concentrated.
“The model is not the innovation,” Yang says. “What you build around it, the interface, the context, the workflow integration, that is where the difference gets made. We are at the point where access to intelligence is not the hard problem. Building systems that actually use it is.”
Conversation as the New Interface
Natural language processing now accounts for 50% of specialized AI library usage in enterprise deployments, with 75% year-over-year growth, according to Databricks’ analysis of production AI systems across more than 10,000 global customers. The implication is direct: the dominant mode of human-AI interaction has shifted from structured inputs and dashboards to natural language, and in many cases to voice. Gartner has noted that a fundamental shift is underway, away from traditional keyboard-centric interfaces and toward AI assistants embedded directly into applications. By 2028, 68% of customer-facing interactions are projected to be handled at least in part by agentic AI systems. The interface layer, the point where a person encounters an AI system, is now one of the primary design and differentiation challenges in the field.
This shift is visible across domains, but it carries particular weight in contexts where the user population is not technically sophisticated and where the interface must earn trust before it earns adoption. Conversational AI systems designed for specific communities, whether defined by profession, faith, geography, or communication style, represent a different category of product than general-purpose chat interfaces. They require the model to be embedded in a voice or conversational layer that reflects the norms and expectations of a particular user community. Getting that right is a system design problem, not a model problem.
“Conversational AI that actually gets used by real people is not just a model with a microphone in front of it,” Yang notes. “It requires understanding who is speaking, what they already trust, and how they prefer to receive information. The interface is the product for most people. The model is invisible.”
Rethinking Education for the AI Era
The global AI in education market was valued at $5.88 billion in 2024 and is projected to reach $32.27 billion by 2030, growing at a compound annual rate of 31.2%. Corporate training and skill development is the fastest-growing segment within that market, expanding at 44.8% annually, driven by enterprise demand for platforms that can reskill workers as AI reshapes job functions faster than traditional curricula can track. A 2024 Educause survey found that 72% of educators prioritized platforms offering personalized content over static curricula. The deeper challenge, one that market size figures do not fully capture, is not simply whether students have access to AI tools. It is whether education systems are redesigning the cognitive skills they develop, given that many of the skills optimized for in traditional curricula are being automated.
The most consequential AI-driven education platforms are not those that use AI to deliver existing curricula more efficiently. They are those redesigning what students are asked to learn in the first place: how to reason through ambiguity, how to evaluate AI-generated outputs critically, how to communicate with precision when AI can produce fluent text on demand. This is a systems problem, not a tools problem. As a member of the SXSW Eco Application Review Board, Yang has evaluated emerging technology platforms across multiple domains, including edtech, and brings that evaluative lens to his view of where AI-driven education is going. The platforms likely to define this space are those built around a theory of learning in the AI era, not just a delivery mechanism for existing content.
“The question for education platforms right now is not whether to use AI,” Yang reflects. “It is whether the platform has a genuine point of view about what human thinking looks like when AI handles the parts that AI can handle. That question determines whether you are building something that matters or just adding features to something that already exists.”
The Rise of the Agent Layer
The global agentic AI market reached $5.25 billion in 2024 and is growing at a compound annual rate of 43.84%, with projections placing it at $199 billion by 2034. Gartner predicts that 40% of enterprise applications will embed task-specific AI agents by 2026, up from less than 5% in 2025. McKinsey’s research shows 23% of organizations are actively scaling agentic AI systems, with an additional 39% in experimental phases. These agents are not chatbots that answer questions. They are systems that take actions, manage information flows, draft communications, route decisions, and operate across software environments on behalf of the people they serve. The practical implication for workplace productivity is significant: organizations deploying AI agents report up to 70% cost reduction in automated workflow categories and measurable gains in employee time spent on non-routine work.
The agent layer represents the furthest extension of the shift from model to system. A workplace agent that manages an individual’s information flow, filters what requires attention, surfaces relevant context, and handles routine communication, does not succeed because its underlying model scores well on benchmarks. It succeeds because it has been integrated into the real communication and workflow patterns of that individual’s working day. That integration is the system. As a judge for the Globee Awards for Technology, Yang evaluates technology innovation against criteria that include real-world adoption and business impact, not just technical sophistication, a standard that maps directly onto how the agent layer will be measured by the organizations deploying it.
“Agents that actually change how people work are not autonomous in the abstract sense,” Yang explains. “They are deeply embedded in the specific rhythms of how a person or a team operates. Building that integration is harder than building the agent. It is also the part that competitors cannot easily replicate.”
Capital in the Age of Systems
23% of organizations are actively scaling agentic AI systems today. Another 39% are in active experimentation. That combined 62% engagement rate, drawn from McKinsey’s global research, represents a market in transition, one where the early production deployments are becoming visible and the patterns of what works are beginning to solidify. The enterprises pulling ahead are not those with access to the largest models. They are those that have invested in building coherent systems: interfaces that fit their user populations, education and change management programs that bring their workforces along, and agent deployments that integrate into real operational workflows rather than sitting alongside them.
Capital that flows toward model development alone, without regard for system architecture, user interface design, or the human-learning component that determines whether adoption actually follows deployment, is misallocated against where the value is being created. Yang’s view, grounded in 14 years of experience building and backing technology platforms across global markets, is that the next several years of AI investment will be defined by the builders who understand this. The platforms demonstrating early traction across conversational AI, AI-native education, and workplace agent systems share a common characteristic: they were designed as systems from the outset, with the model in a supporting role.
“The future of AI will be defined not by models alone, but by the systems that integrate AI into everyday life,” Yang says. “That is where the lasting value gets built. And that is where the work of this generation of builders is going to matter most.”
