Sinolink: What practical constraints does AI penetration face from a business perspective?

Sinolink released a research report stating that the speed of AI penetration is still a common variable for bubble risk and employment risk. A faster penetration rate indicates that AI is entering core workflows, cost savings, and revenue realization more clearly. Concerns about the return on AI investments will further decrease, but the cost is the accumulation of pressure on employment and income distribution. A slower penetration rate means that the labor market has more cushioning time, but the return cycle on capital expenditure is lengthened, and concerns about bubbles are easy to resurface. In the short term, the AI industry chain will maintain a state of "micro-optimism, macro-prudence": the key turning point of technological revolution often lies not in the moment when technology is stronger, but when society and enterprises find new organizational forms that match the new technology. Sinolink's main points are as follows: As the third report in the AI flood series, it focuses on the realistic constraints faced by AI penetration from an enterprise perspective. Although large language models are increasingly capable, this cannot be directly equated with "increased enterprise profits, organizational efficiency, and productivity." The decisive factor is no longer the model's capabilities themselves, but whether enterprises can overcome the gap created by non-standardized data, old system processes, and outdated incentive mechanisms. In fact, almost all major technological revolutions have gone through the process of "technology available" - "organization available" - "economically available." Currently, AI still integrates new technology into old processes using AI to generate documents, scripts, code, and other materials. Efficiency improvement at a single point is rapid, but true productivity growth requires redesigning data flows, approval processes, and job responsibilities around AI technology itself. Therefore, the speed of AI penetration remains a common variable for bubble risk and employment risk. A fast penetration rate indicates that AI's entry into core workflows, cost savings, and revenue realization is clearer, and concerns about the return on AI investments will further decrease. The cost is the accumulation of pressures on employment and income distribution. A slow penetration rate means that the labor market has more time for cushioning, but it may lead to a reemergence of concerns about return on investment. In the short term, the AI industry chain will maintain a state of "micro-optimism, macro-prudence": the key turning point of a technological revolution often lies not in the moment when technology is stronger, but when society and enterprises find new organizational forms that match the new technology. Key to productivity enhancement: not human replacement, but enterprise restructuring The key to productivity enhancement is not whether AI can replace humans, but whether enterprises are restructured by AI; Invention does not equate to productivity, only when AI changes the workflow, data flow, and decision flow of enterprises, can the model's capabilities potentially sustainable increases in enterprise ROI. This means that the impact of AI on the economy should not only focus on model scores, daily active users, subscription revenue, or employee utilization. What should truly be observed is whether AI has entered the core workflows, changed organizational processes, formed measurable revenue or profit contributions, and ultimately raised macroeconomic productivity. In this framework, the employment impact faced by enterprises actually becomes a constraint on the path of AI penetration. The faster AI enters the production system, the more it can alleviate capital market concerns about bubbles, but it will also intensify the impact on employment structure; The slower AI enters the production system, the more it can ease short-term unemployment pressures, but it may easily reignite questions about return on investment. Especially when the affected population consists mainly of young, highly skilled individuals, both enterprises and governments will weigh between efficiency, stability, and distribution if the "new Luddite movement" intensifies, the diffusion of AI may face even greater backlash. Historically, almost all technological revolutions have gone through a period of "technology available" - "organization available" - "economically available." The power revolution provides a reference that is closer to the concept of a "technological gap." Traditional steam-powered factories relied on central spindles, belts, and gears, and machine layouts had to be centered around the Beijing Dynamic Power. When electric motors first entered factories in the early 20th century, many enterprises simply saw them as cleaner and more stable steam engines; however, the original factory structure, division of tasks, and process sequence did not change costs were reduced in some areas, but no new production order was created. It was not until the appearance of distributed motors that allowed machines to be rearranged, assembly lines to be reconstructed, management tiers and operational processes to be remade, that the efficiency dividends of electrification were finally concentrated. The early days of the internet were more focused on emails, web pages, and information retrieval tools, until enterprises migrated their supply chains, channels, payments, advertisements, and organizational coordination online, truly creating new business models. Currently, using AI to create presentations, documents, emails, and code snippets essentially involves completing old tasks within an old organization, with only higher efficiency at specific points. If in the future, tasks are completed through collaboration between agents, these "intermediate files" may not be necessary, and decision-making and execution will be directly linked between systems. The real efficiency release comes from redesigning task breakdowns, reporting methods, approval authority, and responsibility boundaries around AI. The key turning point of a technological revolution often lies not in the moment when technology is stronger, but when society and enterprises finally find new organizational forms that match the new technology, and the restructuring of the AI era has not yet begun. Enterprise organizational structures and workflows are three AI gaps being reconstructed The first obstacle that enterprises face in applying AI is the large amount of non-standardized data that is not efficient from an AI perspective. Large volumes of industry data, process data, and proprietary data are difficult to standardize, call upon, and share compliantly. Public internet texts can be centrally extracted and trained, but the valuable data within enterprises is scattered across ERP, CRM, supply chain, risk management, customer service interactions, contract texts, equipment logs, and human experience. This data often exhibits three characteristics. First, non-standardized: the criteria, processes, and historical records vary between different enterprises, departments, and even within the same organization. Second, strong context: when data is divorced from business rules, approval logic, customer relationships, and risk rules, models can struggle to interpret it correctly. Third, high compliance costs: industries such as finance, healthcare, and government involve privacy, trade secrets, regulatory requirements, and competitive relationships, making it impossible to share data on a large scale like internet texts. Therefore, as AI transitions from general capabilities to vertical industry-specific capabilities, the challenge lies not only in the model's capabilities but also in whether enterprises can transform non-standard, institutionally constrained data into production materials that AI can understand and call upon consistently. Data governance, permission systems, interface modifications, and responsibility boundaries are the prerequisites for AI to enter core workflows. Embodied intelligence is a typical case: physical AI requires interaction data from the real world, sensor data, and high-quality simulation data. According to VOXEL51's 2026 Visual AI Annual Survey: 72% of training data comes from proprietary data, 50% from public datasets, and 40% use synthetic data. In short, the more verticalized an industry is and the more complex its processes are, the harder it is for data to become an asset that can be directly trained and deployed, leading to higher AI usage costs. The second obstacle that enterprises face in applying AI is technical debt (technological inertia). Real-world enterprises and government systems are not blank slates but the results of information technology development over the past few decades. For the sake of business continuity, systems continuously add patches, interfaces, temporary fields, and local processes, resulting in redundant code, fragmented databases, incompatible interfaces, and business rules that are difficult to interpret. Many systems are not nonfunctional but are too critical, complex, and difficult to replace, so they must continue to operate. The U.S. Social Security Administration is a classic example: it has long relied on legacy code like COBOL. In a 2016 congressional testimony, officials stated that the system still had over 60 million lines of COBOL code; its core systems and architecture have not been substantially updated since the 1980s. The SSA's IT modernization plan also acknowledges that many of its core systems have been around for over 30 years, leading to rising maintenance costs, increased system vulnerability, and a talent pool familiar with the old systems retiring. Under the leadership of Musk's DOGE, efforts were made to quickly transform the historic code of the U.S. Social Security Administration in a matter of months but raised widespread concerns about potential data loss, security vulnerabilities, and interruptions in benefit payments. This illustrates that strong AI capabilities do not necessarily mean rapid penetration into the real world. Old code, interfaces, databases, processes, and organizations themselves pose as hard constraints on AI penetration. The third obstacle that enterprises face is how to improve decision-making processes, incentive mechanisms, and truly evaluate AI utility. Since the expansion of agents, many U.S. technology companies have experienced a shift from "encouraging extensive use to re-evaluating" scenes that are most frequently used still involve AI writing materials, conducting searches, generating code, compiling meeting minutes, or handling customer service. For example, Uber, after a period of large-scale encouragement of using large AI models, saw rapid expansion of token expenditure, even depleting the annual budget ahead of schedule, ultimately having to set monthly spending limits for employees. While individual task efficiency may have increased, this does not equate to an overall increase in enterprise efficiency. If processes still involve layers of approval and key decisions continue to rely on human approval, the organization as a whole has not been transformed by AI. There may even be a degree of "Token false prosperity": AI usage appears to be high, but it includes repetitive generation, low-cost trial and error, format packaging, or shifting work from labor costs to computing power and subscription costs. According to statistics on the financial reports of the S&P 500 companies over the past 14 quarters, in the samples that mention AI-related impacts (claimed outcomes), only 6% explicitly stated an increase in revenue cost reduction rather than revenue growth remains the main feedback from enterprises (which may also be due to the inability to quantify AI impacts). From a technological revolution standpoint, the difference between AI-native enterprises and traditional enterprises lies in whether the organization has been redesigned around AI. AI-native enterprises can embed data structures, tool calls, permission boundaries, and agent collaborations into workflows from the start; traditional enterprises must undergo transformations within existing positions, budgets, assessments, compliance, and responsibility systems. The difficulty for the latter is not whether employees can use AI, but whether the organization is willing to let AI change power, processes, and job positions. Therefore, the real efficiency space of AI usually does not lie in the front-facing displays but in the middle and back-end processes. Intelligent customer service, search queries, and content generation are most noticeable to customers, investors, and the media, but financial account reconciliations, contract reviews, risk compliance, supply chain forecasting, procurement management, inventory scheduling, and legal audits are what directly influence costs, risks, and turnover efficiency. If the "economic accounts" of AI remain unclear, enterprise decisions to embrace AI may become more hesitant. Micro-optimism, Macro-prudence At the micro level, some enterprises do benefit from AI in terms of cost savings and productivity improvements, which is reflected in the higher valuations that capital markets are willing to give these enterprises. Particularly, companies in modeling, cloud computing, chips, data centers, and vertical software have formed a clear AI capital expenditure chain the cost-effectiveness of large models is also increasing. At the macro level, the scale of AI is still small, with global AI actual revenue (excluding China) accounting for only 0.42% of the U.S. GDP (as of Q1 2025, it was 0.13%, and in Q1 2024 it was 0.04%), and only 3% of total U.S. enterprise profits. However, this does not mean that AI is without value. Much of the value of AI may exist in consumer surplus form, not fully entering the GDP. However, capital markets ultimately still require revenue, profit margins, and cash flow to support valuations; and for the latter, the heavy social costs (unemployment) that come with it limit the speed of AI penetration. Therefore, observing AI should not only focus on product releases, user growth, enterprise subscriptions, ARR, or the profit elasticity of individual companies; instead, attention should be paid to the following four dimensions: Firstly, has AI entered the core production stage of the enterprise, rather than just being an office assistant? Secondly, has AI changed organizational processes, rather than just improving individual efficiency? Thirdly, has AI generated sustainable ROI, rather than just reducing the costs of specific tasks? Finally, has AI brought about macro-level productivity improvements, rather than just reassessing the prices of certain assets? Model capabilities are fast variables, organizational structures are slow variables; capital demands are fast variables, and total factor productivity is a slow variable. The diffusion of AI is not solely determined by model capabilities but also by the speed of enterprise adoption. The key turning point of a technological revolution often lies not in the moment when technology is stronger but when society and enterprises finally find new organizational forms that match the new technology. Risk warnings: - AI technology exposure updates are not timely and comprehensive, leading to data statistical biases; - AI Agent capabilities develop weaker than expected, resulting in noticeable changes in labor scale; - Rapid shift in global central banks leads to global second-round inflation risks, suppressing global demand, where labor redundancy surpasses AI impact, diminishing the cost-saving attributes, triggering greater concerns over AI investment returns.