The past two years have seen capital markets chase artificial intelligence, primarily focused on one question: whose large model is more powerful. Parameter scale, reasoning capabilities, context length, and multimodal performance became the most prominent labels in this round of AI narrative. However, as the hype truly reached enterprise settings, the core question quickly shifted. Enterprises do not lack a model that can chat, write summaries, or generate content; what is truly scarce is a system capable of integrating data, entering workflows, calling tools, adhering to rules, and completing task cycles. In other words, what determines whether AI can deliver commercial value is no longer just the model itself, but the "harness system" built around it. This is the context in which the "Harness" concept is frequently discussed. (Harness is one of the hottest topics in the AI engineering circle in 2026.)
If a large model is a horse, then Harness is more like the tack - the reins, bridle, saddle, and harness. Without this equipment, even the strongest horse merely possesses potential power; equipped with it, the horse can truly enter the production system, be guided, controlled, and coordinated, ultimately undertaking transport, combat, or labor. The logic is the same in the AI era: the model provides intelligence, while the Harness determines whether this intelligence can be organized, deployed by enterprises, and genuinely used by industry. Re-examining HAIZHI TECH GP (02706) through this lens suggests that the capital market's understanding of it should also evolve.
A shift is underway: The AI competition is no longer just about models. Since the explosion of large models, the most visible aspect for the market has been the leap in model capabilities. OpenAI, Anthropic, Google, Meta, and a host of domestic model vendors have continuously pushed the boundaries, creating a visible competition around "who is smarter." But the enterprise side quickly demonstrated that a strong model does not guarantee successful business integration. McKinsey's "The State of AI in 2025" research noted that while 88% of surveyed enterprises use AI in at least one business function, only 33% have achieved scaled application; merely 39% of respondents stated that AI has delivered improvements at the EBIT level. The implication is clear: AI adoption is not low, but the proportion truly translating into operational results remains limited. The problem is not that enterprises lack models, but that models often fail to integrate into the organizational operating system. Many enterprises can already use AI for copywriting, meeting minutes, report generation, and customer support, but when entering more complex business areas—such as risk control approvals, industry analysis, equipment diagnostics, relationship identification, compliance governance, and cross-system task coordination—models often reveal limitations: they don't understand the enterprise's unique data structure, are unaware of business rules, lack stable tool-calling capabilities, and suffer from insufficient state management and result verification mechanisms. Consequently, the competitive logic of the AI industry is quietly changing. The previous stage was about who had the stronger model; the next stage will likely be about who can integrate the model into workflows, knowledge systems, permission frameworks, and real-world tasks, ultimately converting it into productivity that enterprises are willing to pay for long-term. This shift is essentially a move from "model competition" to "system competition." The rise of Harness responds directly to this phase's most practical challenge.
What is Harness? Putting the 'Tack' on Large Models. "Harness" is not a new model name nor just industry jargon. It is more of a framework for understanding the second half of enterprise AI. Using a simple analogy: the large model is the horse, Harness is the tack. The reins determine direction, the bridle enables control, the saddle allows for bearing loads, and the harness determines if it can truly pull a cart. Without this整套 system, the horse's strength remains just potential; with it, strength becomes schedulable, organizable, and utilizable productivity. Tencent's Senior Executive Vice President and CEO of the Cloud and Smart Industries Group, Dowson Tong, used a similar analogy when discussing the "Harness Era": the truly important next step for the AI industry is not just training a powerful model, but building a complete operational mechanism around the model that allows it to be harnessed, called upon, and integrated into industrial scenarios. The model is like the horse, and Harness is the equipment that enables the horse to enter the production system; using an automotive analogy, the model is the engine, Harness is the wiring harness and control system, and the user is the driver. Tong pointed out that global attention over the past three years focused on the "engine," but today there is a growing realization that real value creation comes from the "wiring harness" that controls the engine. The ceiling of model capability lies not only within the model itself but also outside it; constraints are not about suppressing intelligence but guiding it; AI cannot reliably evaluate itself, so the significance of Harness is to make models safer and more controllable. In enterprise AI, Harness encompasses at least five layers of capability.
Layer One: Context and Enterprise Knowledge Integration. Models know vast public knowledge but are unaware of company policies, contracts, supply chains, customers, equipment, organizational structures, and historical relationship chains. The first task of Harness is to connect the model to the enterprise's own knowledge sources, enabling AI to understand questions within the enterprise context, not just answer generically.
Layer Two: Tool Calling and System Connectivity. Enterprise work isn't done solely through language. Querying databases, calling APIs, triggering approvals, reading reports, writing to systems, and executing rules all require AI to have "hands and feet." Without tool-calling capabilities, AI remains merely an advisor, rarely an executor.
Layer Three: State, Memory, and Continuous Task Management. Real work involves continuous processes, not one-off Q&A. What was done the previous day, the judgment from the last step, and what to call next all require state and memory. Without this layer, AI operates with persistent "amnesia."
Layer Four: Rules, Permissions, and Governance Boundaries. Enterprises are not open playgrounds. Finance has compliance boundaries, government services have permission boundaries, and industry has safety boundaries. A key task of Harness is to ensure AI operates within rules, not freely.
Layer Five: Verification, Feedback, and Error Correction Mechanisms. Enterprises are less concerned with AI being slow than with it being confidently wrong. AWS Labs technical documentation on enterprise generative AI highlights that hallucination management and contextual grounding are crucial for trustworthy enterprise AI deployment, especially in high-accuracy scenarios like finance, healthcare, and law. Therefore, Harness addresses not "is the model smart enough?" but "can the model be integrated into an institutionalized, process-driven, governable work system?" Harness, a top global DevOps vendor founded by serial Silicon Valley entrepreneur Jyoti Bansal and former Apple DevOps lead Rishi Singh, states on its website that customers need not a model that can demo, but "reliable, trustworthy, governed systems at scale." Ultimately, what enterprises truly want to buy is not a "chatty AI" but an "AI that can finally go to work."
Why HAIZHI TECH GP Increasingly Resembles a Representative Company of the Harness Era. Applying the Harness logic to enterprise AI companies changes the evaluation criteria. The noteworthy companies are not necessarily those with the strongest models, but those integrating models into real industrial processes. They solve problems related to connecting models with knowledge, relationships, rules, data, tools, and systems, going beyond simple Q&A or chat interfaces. In this sense, HAIZHI TECH GP's significance lies not in "being another AI company," but in addressing a critical segment of Harness. According to its IPO prospectus, HAIZHI TECH GP is a leading Chinese industrial-grade AI enterprise. Its core technology is graph-model fusion, primarily used to develop industrial-grade agents and provide industrial AI solutions. By 2024 revenue, it ranked fifth among Chinese industrial AI agent providers and first among graph-centric AI agent providers with a 53.3% market share. Public materials show two main product lines: the Atlas Graph Solution (including the DMC data intelligence platform, Atlas knowledge graph platform, and AtlasGraph graph database) and Atlas Agents (developed using graph-model fusion for scenarios like anti-fraud, risk identification, data governance, and smart manufacturing). By end-2024, it served over 300 clients across 100+ application scenarios, primarily in complex industries like finance, public services, energy, and telecom. The key point is not that "Haizhi also makes agents," but *how* it makes them.
HAIZHI TECH GP's Distinctive Feature: Equipping Large Models with 'Maps, Tracks, and Guardrails'. HAIZHI TECH GP's most distinct technical label is "graph-model fusion." While technical, its industrial meaning is straightforward. Large models excel at natural language understanding, generalization, and open-ended reasoning but struggle in complex enterprise environments,容易脱离具体业务语境 and prone to providing plausible but inaccurate answers, especially in high-stakes scenarios like risk control, governance, manufacturing, and industrial decision-making, where model hallucinations are critical flaws. Knowledge graphs, conversely, structurally represent objects, relationships, rules, processes, risk points, and constraints within the enterprise world. In other words, the graph is the map, the model is the driver. Without the map, even a smart driver gets lost; with the map, the model's reasoning is grounded, following relationship networks, business rules, and knowledge structures. HAIZHI TECH GP's prospectus emphasizes it is China's first AI enterprise to effectively reduce large model hallucinations using knowledge graphs. This is significant not for novelty but for targeting the core pain point of enterprise AI: enterprises need not a more creative model, but a less likely to fail model. This aligns closely with the essence of Harness: not granting infinite freedom to the model, but equipping it with tracks and guardrails to operate within industrial knowledge and business relationship networks. Thus, HAIZHI TECH GP differs from many agent companies that simply wrap a workflow around a model. Many agent companies focus on the "action layer" and "interaction layer"; HAIZHI TECH GP focuses more on the "knowledge layer" and "constraint layer." It doesn't just help AI take an extra step; it attempts to solve *why* AI can operate reliably in complex industries. As enterprise AI moves into deeper waters, this capability is often more difficult and slower to develop but creates stronger barriers.
HAIZHI TECH GP Occupies the Most Challenging and Valuable Segment of Enterprise AI. Markets sometimes view new companies through old labels. HAIZHI TECH GP has long been seen as a knowledge graph company, which is correct but incomplete. Sticking to this label underestimates its current position. In today's AI industry structure, its value resembles an overlay of three capability layers. Layer One is the Data and Knowledge Foundation, including data governance, entity recognition, relationship modeling, knowledge organization, and graph database management. This layer answers: What does the enterprise actually know? Layer Two is Reasoning and Constraint Capability. Through graph-model fusion, it combines the large model's language understanding with the graph's precise relationship expression, rule constraints, and structured reasoning. This layer answers: How does AI understand this knowledge without misunderstanding it? Layer Three is Scenario-specific Agents and Industry Solutions, packaging the first two layers into deployable business systems like anti-fraud, risk identification, data governance, and smart manufacturing. This layer answers: How does AI start working and continuously create value within enterprise processes? If traditional knowledge graph companies mainly reside in Layer One, and many agent companies in Layer Three, then HAIZHI TECH GP's significance lies in connecting all three layers. This is the hallmark of a Harness-type company—it sells not a component, but a "working system." Financially, HAIZHI TECH GP's revenue grew from RMB 313 million in 2022 to RMB 621 million in 2025. Revenue from Atlas Agents grew rapidly from RMB 8.9 million in 2023 to RMB 146 million in 2025. The company is beyond the conceptual stage, converting graph-model fusion and agent capabilities into real revenue, with market acceptance extending from underlying platform capabilities to higher-level agent capabilities.
The key difference between enterprise and consumer AI is error tolerance. A consumer assistant giving a wrong answer is often dismissed; but an enterprise system misjudging risk, misreading policy, misallocating resources, or misleading governance carries entirely different consequences, especially in finance, government, energy, and manufacturing, where AI must be explainable, verifiable, accountable, integrable, and sustainably optimizable. Therefore, the scarcest resource in enterprise AI is not point-in-time intelligence but system-level trustworthiness. Viewing this alongside HAIZHI TECH GP's product roadmap clarifies its strategic position: it avoided the crowded general model race and didn't settle for a lightweight agent interface, instead choosing a heavier, more industrial path closer to core enterprise processes—integrating AI into high-complexity scenarios via graphs, graph databases, knowledge structures, and agents. This path may not be the easiest for storytelling but is closer to enterprise budgets and more likely to form long-term barriers. In other words, HAIZHI TECH GP's value lies not in being "another AI company," but in occupying the most challenging part of the enterprise AI journey: integrating models into industrial settings characterized by complex relationships, dense rules, and high error costs. This is precisely the most valuable segment in the Harness Era.
HAIZHI TECH GP's Scarcity Becomes Clearer Within the Context of Hong Kong-Listed AI Companies. Comparing HAIZHI TECH GP to other HK-listed AI firms highlights its differentiation. PHANCY (06682) is closer to an enterprise AI platform and decision intelligence company, excelling in platformization, model engineering, and industry solutions, representing "enterprise AI platform layer" capabilities. SENSETIME-W (00020) leans towards "large model + infrastructure + agent application ecosystem," possessing its own large model (Riyi), platforms like "Wanxiang" for agent development, and various agent products for office, content, and e-commerce, representing model capability spilling over into application ecosystems. KINGDEE INT'L (00268) follows a different logic, with strengths not in the model layer but in enterprise management software and ERP entry points. Its AI agent strategy essentially embeds agents into existing software workflows, making it more of a "software-entry Harness." In contrast, HAIZHI TECH GP's differentiation lies not merely in making agents, but in being one of the few companies that coherently integrates "knowledge graphs, graph databases, structured knowledge bases, complex relationship reasoning, and industrial agents." If PHANCY's strength is enterprise AI platformization, SENSETIME-W's is models, multimodality, and application ecosystems, and KINGDEE INT'L's is enterprise software entry, then HAIZHI TECH GP's strength lies in being closer to the hardest, most defensible part of Harness—integrating models into high-complexity industrial knowledge networks. Therefore, while HAIZHI TECH GP may not be the most broadly defined Harness company in HK, it is likely one of the most recognizable "industrial-grade Harness" companies.
Of course, labeling it the "Pioneer Harness Stock" is meaningless if just a catchy phrase; it requires support from industrial and capital logic. Currently, this assertion rests on three premises. First, HAIZHI TECH GP operates not in a narrow niche but a rapidly expanding new interface. Citing its prospectus, China's industrial AI solutions market is projected to grow from RMB 65.4 billion in 2025 to RMB 286.1 billion in 2029, a 44.6% CAGR. The market for industrial AI agents integrated with knowledge graphs is expected to surge from RMB 200 million in 2024 to RMB 13.2 billion in 2029, a 140.0% CAGR. This places HAIZHI TECH GP at the key interface where enterprise AI moves from pilot projects to systematic deployment. Second, it addresses the most practical and enduring pain points in enterprise AI. The recurring questions are not "is the model big enough?" but "how to reduce hallucinations, integrate enterprise knowledge, enter business processes, operate stably in high-barrier industries, and make AI outputs more explainable and governable." HAIZHI TECH GP's product roadmap directly targets these issues. Third, its story can transition more easily from "technology narrative" to "capital market narrative." Capital markets favor not "advanced technology" per se, but "advanced technology positioned on the main path of value realization." Thus, calling HAIZHI TECH GP the HK "Pioneer Harness Stock" is not just a slogan today but begins to hold explanatory power in both commercial validation and capital market contexts. If future market valuation logic for AI companies indeed shifts from "model capability" to "system implementation capability," then companies like HAIZHI TECH GP that can integrate models into complex industrial processes may be re-evaluated earlier and re-priced according to Harness logic compared to concept-driven AI firms.
This does not imply an absence of challenges. The capital market will ultimately focus on several practical questions: Can the technical advantage of graph-model fusion be consistently translated into more standardized, higher-reusability products? Can the rapid growth of Atlas Agents be sustained? And can HAIZHI TECH GP defend its composite moat of "graph + model + industrial knowledge" as large model vendors, cloud providers, and software companies increasingly target the enterprise market? The answers will determine whether it evolves from "a company illuminated by a new narrative" to "a company truly priced for the long term."
Comments