Large language models (LLMs) are entering core enterprise business scenarios, but the true barrier to their deployment is not the model capability itself, but whether data can be accurately understood and invoked by the models. A massive amount of scanned documents, PDFs, images, videos, and complex tables are accumulated in business systems, retaining critical knowledge, yet they have long remained in an unstructured, hard-to-parse, and hard-to-correlate state.
After investing substantial resources in structured governance, enterprises have realized that data formats designed for human reading are not naturally suited for LLM reasoning. Only by restoring semantics and structures as losslessly as possible can a truly AI-Ready data foundation be built. EasyLink is an AI-native enterprise connecting AI, data, and scenarios. Its founder, Chen Siyin, is a cross-boundary entrepreneur with years of investment experience in financial technology. She keenly identified data as the hardest nut to crack in AI application deployment and led her team to delve deep into unstructured data processing. In vertical industries with high-precision requirements, such as financial asset management and healthcare, she has pushed LLM applications from toy-level to production-level.
Recently, Ai Fenxi had an in-depth conversation with Chen Siyin, focusing on several core issues: AI's data preferences, pathways for extracting tacit knowledge, and the future evolution of AI-native organizations.
Core Insights
- Enabling AI to understand data is the biggest barrier to industrial AI deployment for enterprises.
Against the backdrop of enterprises self-deploying open-source models and model capabilities becoming increasingly homogenized, the key determinant of AI application quality is the quality of content fed into the LLM. The model determines the baseline, while data determines the ceiling.
- AI is not accustomed to human structured thinking.
Traditional tables and rigidly structured data are designed for human consumption. For AI, however, lossless semantic descriptions and natural network structures are the optimal nourishment. Transforming data into natural language context is what makes it truly AI-Ready.
- There is no middle ground in enterprise-level AI scenarios.
General office scenarios can utilize lightweight or open-source architectures, but core business scenarios such as finance and healthcare demand extremely high precision. A score of 90 is merely a passing grade in enterprise scenarios; without achieving an accuracy of over 95, AI cannot be deployed to the front lines of production.
- Future AI-native organizations will be extremely flat.
As AI deeply penetrates various job roles, traditional complex reporting lines will be dismantled. Future enterprises may be streamlined to just two layers: the executive level and the expert level. The expert layer will consist of a few super-individuals with extensive experience, responsible for managing and auditing Agents.
The following is a transcript of the interview, slightly edited without altering the original meaning.
01. What is AI-Ready Data?
Ai Fenxi: It is often said in the industry that enterprises should build an AI-Ready data foundation, and the core of AI-Ready is Data-Ready. What exactly does AI-preferred data look like?
Chen Siyin: AI's advantage in processing information lies in understanding semantic associations. When processing unstructured data, preserving sufficient context and relational expressions enables AI to exert stronger comprehension capabilities. Our team started our business in 2023. Our first client was a financial institution that gave us a wide variety of data, including videos, images, scanned documents, and PDFs. They wanted to build an internal employee assistant, hoping to use a knowledge base approach to enable the LLM to utilize this data. That was when we discovered the pain point: although the model is powerful, it lacks the enterprise's private internal data, and enterprise knowledge is embedded in massive amounts of unstructured data. To truly enable LLMs to utilize this knowledge on the reasoning side, two core issues must be resolved. First, how to provide the semantics in unstructured data to the model losslessly. This semantics includes not only text but also the semantics of tables and images. What form does AI actually prefer? Humans particularly like tables and structured expressions, whereas AI is better at capturing causal and logical chains from natural language containing rich context. Second, structure must be considered. When facing large-scale data, on the one hand, the natural structure among internal semantics of the data must be considered, which is helpful for model reasoning; on the other hand, semantic relationships also exist between different pieces of data. How to completely restore and express these relationships to the model is crucial. When building underlying unstructured data capabilities, we have always focused on the complete restoration of semantics and structure.
Ai Fenxi: Can you give an example?
Chen Siyin: The most typical example is transforming lengthy documents into a hierarchical structure tree that retains core levels and associations. This format preserves the rich context of the original text while restoring clear mappings between data nodes, which is truly AI-Ready for AI.
Ai Fenxi: Where do these tasks stand in the overall AI infrastructure? How important are they?
Chen Siyin: These data preprocessing tasks are the first step in the entire AI application construction chain, and also the most difficult and core step. This includes the memory capabilities of Agents, which rely on the processing of the data itself. If data parsing is incorrect, no matter how strong the reasoning model is, it will be a castle in the air. As for general LLMs being able to do data processing, general capabilities face huge challenges when confronting different industries. For example, the financial industry, which we interact with frequently, naturally has extremely high requirements for accuracy and reliability. Any hallucinations generated during data processing could bring compliance risks. The healthcare industry is the same; massive amounts of medical records, papers, and images cannot have any parsing errors when converted into fuel usable by AI. When enterprises perform data preprocessing, the precision requirement for these complex data is 100% restoration of semantics and structure. It is difficult for general models to achieve perfection in all scenarios.
Ai Fenxi: Can data preprocessing be understood as document parsing?
Chen Siyin: Merely doing document parsing is not enough; a hierarchical structure must also be established. Specifically, after extracting key information from various formats, it is necessary to identify the structured relationships of this information and build a logical structure tree containing chapters, paragraphs, and levels to provide comprehensive context information for model reasoning. This is the structure among internal semantics of the data. Semantic relationships also exist between different pieces of data. Further, a larger and more comprehensive structured relationship can be established to solve the pain point of enterprise data being scattered everywhere. For example, if you want to understand a certain enterprise's revenue situation in a specific dimension, there might not be any complete data directly stating what the revenue is; it is scattered in various corners. At this point, if this structured relationship can be built, the model can reason based on the associative relationships, significantly improving retrieval capabilities and the accuracy of final results.
Ai Fenxi: Besides PDFs and forms, can unstructured data like audio and video files be transformed in this way?
Chen Siyin: Whether it is PDFs or videos, we support semantic parsing. Based on intelligent vision and semantic analysis, the system can automatically identify content levels and establish contextual associations, achieving high-precision information presentation.
Ai Fenxi: There are many parsing vendors, and there is also competition from open-source parsing models. How do you build a competitive advantage?
Chen Siyin: Focusing on precision in vertical fields to build a data flywheel. In customer scenarios, we encounter corner cases with industry characteristics. Through continuous iteration and optimization, we gradually form our own moat at the data layer. Becoming increasingly refined and accurate in certain vertical fields is the most core data flywheel effect. With this task-specific LLM, we can break through the limitations of model performance and computing power costs. Taking the financial industry as an example, they have extremely high requirements for real-time processing. When listed companies release a massive amount of annual reports on the same day, they need to be fully processed within T+0 time. If directly using a general LLM with ultra-large parameters or calling public cloud APIs, the consumption of computing resources and costs would be astronomical, and performance would not be able to keep up at all. Our visual semantic model is an industrial-grade model focused on the specific task of data parsing, specifically designed to solve performance and cost issues during the data parsing process. Our self-developed multimodal data parsing engine is deeply optimized and adapted to domestic computing power. The single-card inference efficiency can reach 1.5 times that of NVIDIA cards, and the inference acceleration engine is twice as fast as open-source solutions, achieving a parsing speed of about 1 second per page.
02. Enterprises Dare Not Use AI That Only Scores 90
Ai Fenxi: What are the advantages of Agents rooted in an AI-Ready data foundation compared to those without this foundation?
Chen Siyin: The difference between the two is the difference between production-level and toy-level. What enterprises corely value is still precision. In industries with extremely high precision requirements like finance and healthcare, a score of 90 might just be a passing grade before people dare to start using it, and core businesses must achieve at least over 95. We recently found through communication that many enterprises have used open-source architectures or introduced relatively lightweight products, only to find it difficult to move from the validation stage to core business scenarios when underlying data processing is inadequate. The reason behind this is that if the underlying data is not processed well and the input data quality is low, no matter how much you tune retrieval or the model, the output quality will still have problems. Therefore, if accuracy cannot be improved, there is no hope of pushing it into enterprise core business scenarios.
Ai Fenxi: Speaking of Agent construction, many enterprises now use the Harness architecture. What specific role can an AI-Ready data foundation play in it?
Chen Siyin: The data foundation is the core cornerstone for enterprises to build complete underlying Agent capabilities. We understand that enterprises need several complete capabilities to build Agents: the first is on the data side, where both structured and unstructured data need to be prepared to be AI-Ready; the second is on the knowledge or memory side, where after having data, it needs to be precipitated into knowledge usable by the enterprise. Actually, internal enterprise knowledge is divided into two categories. One is explicit knowledge, which is the various data just mentioned; the second is tacit knowledge, which includes industry cognition, experience, and feedback during use across various industries and job roles. For example, when a person sees a bunch of data, they will naturally string this data together based on past experience to summarize it; this is tacit knowledge. The task on the knowledge side is to integrate all explicit and tacit knowledge so that the LLM can better understand and express it. On top of this capability, combined with Harness and all internal enterprise knowledge, a complete, production-level underlying Agent capability can be assembled.
Ai Fenxi: The extraction of tacit knowledge has always been a difficult problem. How much of a role can AI play in knowledge extraction in the future?
Chen Siyin: Knowledge extraction is a rigid demand in specific vertical scenarios such as investment research. The core role of AI here is to act as a navigator for underlying data. We have worked on some vertical scenarios, such as investment research in the financial industry, which naturally requires knowledge extraction around specific themes. Different industries have different analysis chains and frameworks, which means the invocation of underlying data has strong industry characteristics. This requires expressing the semantic relationships in the data explicitly in the middle layer between data and reasoning, in a way that is more friendly to the model and does not require rigidly defining a Schema. This is like giving AI better navigation capabilities on the underlying data. As for whether this process is completely AI-native or requires human-machine interaction depends on the complexity of the task. If it is just writing a formal report in a fixed format, you can directly throw a template at it, and AI can understand and implement it itself because it clearly knows which data to fetch from the underlying layer. But for some open-ended or decision-oriented complex scenarios, human-machine interaction is needed to capture and precipitate human feedback during task execution. At the same time, we also need to perform more complex network structure processing on the data itself to support the model in deep reasoning.
Ai Fenxi: What is the difference between this network structure processing and knowledge graphs? In the past, building a knowledge graph required defining various metadata and Schemas first.
Chen Siyin: There is absolutely no need to define Schemas in advance now. Traditional knowledge graph construction required defining Schemas and rules in advance, which has become a limitation in the LLM era. Now, we let the model automatically understand the associative relationships between semantics without needing to define rigid frameworks in advance.
Ai Fenxi: So for the ultimate client customers, what the underlying graph architecture specifically looks like is actually a black box?
Chen Siyin: Clients care more about the quality of the final output; underlying technical details are not their focus. Business users are unaware of which underlying architecture is adopted. If it is a knowledge base, clients care about the quality of the answers; if it is an Agent, clients only care about the quality of task execution.
Ai Fenxi: Financial institutions have always been doing data governance. Does this kind of data and knowledge governance for AI applications now need to be integrated with traditional data governance?
Chen Siyin: At this stage, the two are more advanced in parallel. As AI applications deepen, there may be a need for integration in the future. Banks should continue to do traditional underlying data governance as they always have. The AI infrastructure or AI governance built at the upper layer around LLM applications are currently two different levels.
03. AI Prompts Enterprises to "Lose Weight"; Future Organizational Structures Will Have Only Two Layers
Ai Fenxi: How do you view the impact of AI on organizational forms? Can it also establish some kind of competitive advantage?
Chen Siyin: AI is already penetrating every job role, and behind every employee, there might be a large number of Agents being invoked. It is highly likely that future organizational structures will evolve to have only two layers: below the management level, there will be a small number of highly experienced professional talents responsible for managing various functions. They will be in charge of managing various Agents and checking results, ultimately collaborating with the management level. The complex reporting lines in the middle will all be replaced by Agents. Our company is now developing in this direction. As an AI-native enterprise, we do not want future organizations to become particularly bloated and complex.
Ai Fenxi: What are the main vertical industries that EasyLink is currently focusing on?
Chen Siyin: Finance and healthcare are currently the two industries where we process the largest amount of data. Especially in the financial asset management industry, we have achieved full-chain coverage from underlying data processing, knowledge precipitation, and memory layers to upper-layer Agent application scenarios. This is actually driven by our clients. Based on client feedback, we found that after the underlying data is done well, extending upwards to Agent scenarios is a very natural progression.
Ai Fenxi: The healthcare industry itself has a long history of data governance and an extremely large amount of unstructured data. What is your starting point for entering the healthcare industry? What are the core plans for 2026?
Chen Siyin: Actually, neither finance nor healthcare was deliberately chosen by us; rather, clients came to us proactively with their pain points. The common characteristics of these two industries are very obvious: first, the volume of unstructured data is massive. Asset management has a large number of equity and bond reports, while healthcare has massive amounts of medical records, papers, and images. Second, both industries are extremely eager to transform this data into AI capabilities but have encountered bottlenecks at the data foundation stage. Regarding the plans for 2026, we will closely focus on promoting two main lines. The first line is the data line. We will combine the vertical industries we have already entered to build more high-quality industry data processing capabilities, helping enterprises create AI-Ready data foundations. The second line is the Agent line. In early July, we officially released Tiri, our Agent product for enterprise-level scenarios, directly integrating our most core data parsing capabilities to help clients deliver final business results.