The "easily overlooked" obstacles to creating large-scale model products in 2B scenarios

    Let’s get back to the point. I find that every time I write an article about the grand narrative of VC, it attracts a lot of attention, but I personally feel that VC narratives are all trying to “imagine” or “outline” the future in a cycle of several decades. We ordinary people, especially practitioners, should pay more attention to the impact of technology applications that are directly related to the present.     Under the wave of GenAI, you can see all kinds of exaggerated headlines every day when you wake up about what technology or company is like. Excessive media noise has largely obscured the key challenges of how to build high-value AI applications that can be implemented. Especially in highly complex 2B industry segmentation scenarios, most of the beautiful PPTs and Demos have no "then" in the end.

    Today, this article still returns to the "down-to-earth" positioning of the local blogger, and discusses the capability gap that is common in the application of large models in 2B scenarios but is often strategically ignored.

1. Background -  Overrated "Intelligence" and Underrated "Implementation"

   
   

    Generative artificial intelligence (GenAI) technology represented by ChatGPT is hailed as the "iPhone moment" of artificial intelligence. This view was first clearly stated by Jensen Huang of Nvidia in early 2014: " This is the iPhone moment of artificial intelligence ". If people still had doubts about the technological revolution triggered by this round of GenAI two years ago, now everyone is scrambling to seize every opportunity to get on this train.

    Over the past two years, the general-purpose large models (Foundation Model) represented by OpenAI at home and abroad really feel like " one week for the large model is one year in the real world ". The LLM technology route has evolved from "Pre-training" to "Test-time Compute", from "predicting the next token" to "taking more time to think/reason". While the intelligence capabilities are getting stronger, the intelligence costs are rapidly decreasing. This trend alone is in line with the early characteristics of the new technological revolution.

source: latent.space(as of Jan. 2025)

    However, compared with the popularity of basic large models (such as OpenAI GPT, Anthropic Claude, Google Gemini, DeepSeek, Alibaba Qwen, etc.) and 2C applications (such as ChatGPT, Gemini App, my favorite NotebookLM, Cursor, Devin, Manus, Midjourney, Runway, Sono, Monica.im, etc.), the difficulties of GenAI applications in many 2B industry scenarios are in sharp contrast.

    Despite the emergence of various experimental projects, the proportion of enterprises that have pushed GenAI from the PoC stage to the actual production environment may not exceed 20%. Harvard University's Jen Stave's "Jagged Frontier" theory on AI productivity is particularly prominent in the highly professional 2B field, that is, AI performs well in some tasks, but may complicate other tasks. In plain words, AI cannot do well in things that human experts think are very simple, but its performance in some complex tasks is amazing. I can't help but think of an interesting video I saw a few days ago. A 13-year-old child who is about 1.9 meters tall came home from school and complained to his mother that he forgot to bring his "Little Genius" watch.

    Now the question is:

Why is it that many GenAI projects that are highly anticipated in 2B scenarios find it difficult to create scalable and sustainable value after investing huge resources?

    This article attempts to analyze the "easily overlooked" obstacles to building 2B large-scale model products from the perspectives of technology, models, products, and organizations, as well as possible solutions to these problems. Due to my limited level, it is inevitable that there are some inaccurate or even wrong views, so listen to both sides to make a better decision.

1. Data Bad Land Rover - The Neglected "Data Alchemy"

    The view that "large-scale, high-quality data is the fuel of AI" has become "correct nonsense" from the traditional AI era to the big model era in recent years. One point that is generally overlooked is:

In the 2B scenario where data quality is uneven, structured, full of implicit knowledge and various data islands exist, the process of refining "raw domain data" into "high-quality fuel" that can be used by GenAI is a "data alchemy" that requires huge investment, long cycles, and extremely high technical content, but is extremely lacking in "sexy narratives."

    I have seen and heard of many companies and teams dying in this easily overlooked "first mile". Almost all large organizations face a common problem: data is highly fragmented and chaotic, and very difficult to integrate. Enterprise data is scattered in different systems (such as ERP, CRM, SCM, MES, sensor networks, etc.), with different formats and quality, and it is difficult to integrate to form unified, reliable, high-quality field data to provide the basic "data fuel" for AI applications.

    Many organizations view data preparation for GenAI as a one-time “data cleansing” (simple ETL or data lake build) rather than an ongoing, highly complex “alchemy.”

    Little do people know that the establishment of high-quality domain data advantages is the prerequisite for AI to be truly implemented, and it is also a key step that cannot be skipped. However, the construction of domain data advantages requires continuous strategic investment. These tasks are often the "dirty and tiring work" of enterprise digitalization and intelligence. Senior management cannot see or look down on them. Whether the implementers can do a good job depends highly on the deep understanding of the business, cross-departmental collaboration and effective management mechanisms. However, once successful, it will have huge long-term value to the enterprise. To use a local saying we often mentioned before, it is " sitting on a gold mountain of data to beg for food ."

    The challenge of the "first mile" lies not only in data access, but also in achieving semantic unification and alignment of business logic. This is an ongoing and costly process. For example, I believe that Palantir's Foundry and Ontology are the company's invisible "moats .     "

    The strategic underestimation of "data alchemy" has led to the fact that the 2B scenario large-scale model application project itself is built on an unstable foundation, wasting money and human investment, and failing to release real business value.

    So what is the solution? Although each company has its own industry and company characteristics, the practices of Palantir's Foundry and Ontology are definitely worth learning from. The reason why Palantir's Foundry and Ontology are considered "invisible moats" is that they face the original sin of enterprise data chaos and invest more than ten years in "the dirty work of data governance." In the era of GenAI, is this "slow work" an important direction for building 2B competitive barriers? Personally, I think it is still.

    Many blind big model technology optimists may hope that GenAI can provide a "microwave-like" shortcut to automatically understand and process chaotic enterprise data. Understanding is beautiful, but the reality is very bleak.

    Trying to bypass the “slow cooker” stage of data governance and directly expecting GenAI to solve its data chaos problem is likely to result in applications that are superficial and lack reliability. GenAI may be able to speed up the “slow cooker” process, but it cannot completely skip it.

2. Model obstacles - overestimated "general intelligence" and underestimated "domain intelligence"

    Chasing bigger and more universal models seems to have become a "technological religion". But what is overlooked is:

In 2B scenarios that require professional industry knowledge, domain logic reasoning, and high-reliability decision-making, the capabilities of general large models are fully exposed. Its characteristics of "knowing a little about everything but mastering nothing" make it difficult for it to directly handle core business. How to fill the huge gap between "general intelligence" and "domain intelligence" is currently the most underestimated and challenging part of GenAI in 2B implementation.

    I believe everyone can feel it to some extent. The lack of industry depth, industry expertise and industry-specific knowledge in the vertical industry field makes these generic big models feel a bit uninteresting to use. Enterprises can easily fall into the "big model capability trap", that is, overestimating the actual utility of the generic model in specific complex industry scenarios. This is also the "Jagged Frontier" dilemma mentioned above.

    This blind optimism about the capabilities of general big models will lead enterprises to believe that they can solve complex industry problems by simply calling LLM (such as the most common prompt engineering). However, the depth of general big models' knowledge in specific fields, their ability to follow complex business logic, and the accuracy of key details are often orders of magnitude different from those of human experts. The characteristics of general big models that "know a little bit about everything but are proficient in nothing" make it difficult for them to directly handle core business.

    From the specific technical challenges, the core challenge is how to effectively "inject" or "align" the implicit, unstructured, and highly context-dependent "industry wisdom" (Domain Expertise) into the large model. This is far from being completely solved by simple SFT Fine-tuning or RAG. It requires systematic innovation in the internal mechanism of the model, domain knowledge representation, and human-computer collaborative reasoning. The exploration of companies such as Harvey.ai and OpenEvidence is essentially trying to build a new type of "cognitive architecture" , and LLM is just an important component of it.

    Harvey AI, a star company in AI applications in the legal industry, is trying to build professional AI applications for the legal field. It integrates multiple basic models (from OpenAI, Anthropic, and Google) and deeply optimizes specific legal tasks and business workflows. This is more like building a new "cognitive architecture" for a specific industry.

    OpenEvidence, a star company in AI applications in the medical industry, is trying to provide evidence-based support for clinical decision-making by integrating a large amount of peer-reviewed medical research, which also reflects the ability to deeply process knowledge in specific fields. Although RAG enhances factuality by retrieving external data, and fine-tuning can embed domain knowledge at a deeper level, more cutting-edge hybrid methods such as RAFT (retrieval enhanced fine-tuning) may also be a technical route worth trying, which can combine the advantages of both.

    From a scalability perspective, one thing that is easily underestimated strategically is that building, verifying, and continuously maintaining a "truly knowledgeable" industry domain model for each industry segment (or even each large enterprise customer) requires long-term, huge technical and talent investment. Most senior corporate leaders, especially business-related leaders, find it difficult to really make this decision, unless the startup company can only rely on this to survive in the fierce competition.

    How to solve it? Here are some possible solutions.

    Should we re-examine "big model + big data (domain)", "big model (generative) + small model (professional model)" or build our own domain big model in the long term ? These technical routes may have their own feasibility in different stages and scenarios. Technical leaders need to find the optimal balance between "model scale-domain depth-cost-effectiveness". What you must pay attention to is not to hammer nails everywhere (industry scenario) with a hammer (big model).

3. Product obstacles - neglected "business process reshaping" and "user mind re-education"

    The most common misunderstanding in 2B scenarios is:

    Considering big model technology as a silver bullet, we try to completely overturn and subvert the previous business system and process. We think that through some data embedding + RAG, based on some open source agent products/frameworks and some prompt templates, we can quickly rebuild a "new" business system.    

    The biggest disadvantage of this misunderstanding is that people use the excuse of building a large model application to make random changes. What is even more terrible is that they use simple "chat is everything" interaction in the name of "improving customer experience" to cover up the complexity of business systems and user needs . To some extent, the harm of using the wrong person is far greater than the impact of not using the person.

    In the 2B industry, if GenAI is deeply integrated into the existing, complex, and deeply rooted core business processes, and only provides a "plug-in" chatbot or text generator, its value will be greatly reduced, and it may even be regarded as an "efficiency disruptor." I have always insisted that no matter what technology and product form, as long as it cannot truly realize customer and business value conversion, it is "playing rogue."

    Many GenAI applications in 2B scenarios become "shiny toys" after the POC demo, but they do not fundamentally change business processes and working methods. Personally, I believe that the real transformative value of AI technology in the 2B field lies in making GenAI a "full-process embedded engine" that is seamlessly integrated into core business processes and reshapes them, which also requires a deep understanding of business processes, stakeholder recognition, and intelligent organizational management. Compared with superficial single-point automation, fundamental intelligent process reengineering can bring more far-reaching impacts.

    Strategically ignoring deep business process integration and business value conversion will result in GenAI technology only providing "dispensable" marginal utility on non-critical business paths, and instead become an "efficiency disruptor."

1. Product positioning perspective: From "LLM general capability API" to "business operating system OS" is the key to building a moat

    Many GenAI products only use AI to "empower" a certain link in the existing process, but fail to fundamentally consider whether GenAI can give rise to new, more efficient, and smarter business processes. Real change often comes from the reconstruction of the entire value chain, rather than local optimization.

    So how can we "invisibly" embed GenAI's capabilities into the core business systems (ERP, CRM, SCM, PLM, industry-specific software, etc.) that users rely on daily, making it the "default option" and "smart base" for users to complete tasks, rather than a "new tool" that requires additional learning and adaptation?

    Another common misconception is:

    Strategically underestimating the technical investment, time cost, and organizational coordination difficulty required for deep integration with corporate customers' large and heterogeneous IT systems is often the "heaviest" link in the large-scale promotion of GenAI applications in 2B, and these things are considered "things that don't scale" in the early stages.

    Will the ultimate form of 2B GenAI products tend to be an "industry operating system" or "intelligent business platform" that not only provides AI capabilities but also defines new industry standards, data interfaces, and collaboration paradigms? How to build such a platform? Who can lead it?

    Building such an "industry operating system" requires deep industry insights, powerful data ontology modeling capabilities (such as Palantir's Ontology), and the ability to cultivate an entire ecosystem. It would be a strategic misjudgment to simply use the big model as a black box.

2. User experience perspective: The “chat is everything” experience is definitely not the best experience

• The neglected "interaction efficiency trap"

      Initial over-optimism about the "chat is everything" interaction method quickly hit a wall in 2B complex task scenarios. Trying to use a universal "Chatbot thinking" to design the interaction interface of all 2B GenAI products, while ignoring the unique requirements of professional users in different industries and positions for information density, operation accuracy and decision-making efficiency. 

  
  

      I think I have used various GenAI applications in depth, but for a user like me who has some knowledge of large model technology and is a half-baked amateur "product manager", I feel that the threshold for complex scenarios and accurate expression of needs is still extremely high. I also find that the instruction following for some vertical scenarios is still far from expectations . For professional tasks that require precise input, multi-step operations, and structured output, pure open chat is often inefficient, error-prone, and difficult to standardize.

      Pure open chat interaction seems to lower the threshold for use, but it often leads to inefficiency, is prone to errors, and is difficult to standardize. The cognitive load brought by pure conversational UI and its shortcomings in functional usability are very obvious.

  
  

• Product interaction challenges

      How to find the best balance between the flexibility of GenAI CUI and the structured control interaction of traditional GUI? I think the more realistic and effective interaction in 2B scenarios must be "hybrid interaction", such as "natural language initiation + structured parameter adjustment + visual result verification" or "context-aware active interaction", that is, AI actively provides relevant suggestions or operation entrances based on the current business process and user portrait.

  
  

• The concept of DigitalMe

      In the 2B scenario, should GenAI's ideal interaction pursue "getting to know you better the more you use it" and achieve a highly personalized and adaptive "intelligent workflow companion" by continuously learning users' operating habits, preferences, and domain knowledge? If so, what kind of technical and product design support is needed?

  
  

• The neglected “fragility of user trust”

      In the 2B scenario, a critical AI decision error or interaction barrier may permanently destroy the trust of users (especially conservative industry experts) in new technologies. Building user trust in the 2B scenario is a slow and difficult process, but its collapse can happen in an instant. A critical AI decision error or a bad interaction experience, especially for industry experts who are conservative about new technologies, may permanently destroy their trust in GenAI products. This fragility of trust is something that startups must always be vigilant about during product design and iteration.

3. Entrepreneurs’ perspective: “ Products that cannot generate real money are just scams”

• The Neglected “Customer Success”
  

      In the 2B GenAI field, "customer success" is far more than product delivery, but also about deeply participating in the customer's business process transformation. If it cannot bring real money and a qualitative change in user experience, it is just a hooligan. For successful GenAI products in 2B scenarios, will the connotation of "customer success" evolve into a kind of Business Evolution as a Service?

  
  

• Embrace of smart organizational culture

      For startups trying to transform traditional industries with GenAI, the biggest challenge may not come from the technology itself, but from the huge resistance to changing the deep-rooted work habits and organizational culture of users (especially large corporate customers). How to overcome this "soft barrier"? Overcoming this "soft barrier" requires startups to have strong strategic patience, excellent communication skills and professional change management knowledge. It is far from enough to rely solely on technological leadership.

  
  

• The difficulty of proving value

      Proving to 2B customers the true business return (ROI) of GenAI in a clear, credible, and quantitative manner is often more difficult than implementing the technology itself. This challenge is particularly prominent when the value of GenAI is indirect (such as improving employee satisfaction), long-term (such as accumulating knowledge assets), or difficult to measure with traditional financial indicators (such as short-term profit growth).

  
  

• 2B GenAI’s business model

      I personally believe that future GenAI products will be priced based on outcomes , that is, charging according to the proportion of cost savings or improved business indicators (such as sales conversion rate, customer retention rate) that GenAI applications save for customers. The feasibility of this business model lies in that it can more closely bind the interests of startups with the success of customers, and encourage startups to provide solutions that can truly create value.

      For example, Intercom's Fin charges by the number of conversations solved by AI, and Saleforce Agentforce charges by the number of problems solved in customer service scenarios. These are all attempts to move towards a pay-by-results model  . Although this model itself has challenges in how to clearly and fairly define and measure "results", the development of a value-driven business model must be the key to achieving sustainable profitability for GenAI products in 2B scenarios.

  
  

• The neglected “profit window”

       GenAI technology is still evolving at a high speed, and the cost and capabilities of basic large models are also changing rapidly. It often happens that companies with basic large models move up to the application and business parts, while some vertical scenario GenAI applications want to move down to large models. 2B GenAI startups need to quickly find a sustainable profit model and establish barriers before the technology dividend disappears or the competition becomes fierce, otherwise the profit space will be severely squeezed.

  
  

4. Organizational obstacles - neglected "cultural inertia" and "capability gap"

    The impact of GenAI on organizations is far more than just a new technology. It requires enterprises to undergo a profound "genetic reorganization." From rigid hierarchical structures to overly specialized divisions of labor, from experience-driven decision-making to data and intelligent collaborative decision-making, from sticking to traditional skills to embracing "Collapsing the Talent Stack" and lifelong learning.

    The biggest obstacle is often not technology, but the deep-rooted "organizational cultural inertia" and the resulting "capability gap":

• The neglected “friction cost of collaboration” 

      GenAI products naturally require a deep coupling of talents with diverse backgrounds, such as algorithms, engineering, data, products, design, and domain experts. However, there may be huge differences in the "language system", "thinking paradigm", "work rhythm", and even "value ranking" between different professions, resulting in a large amount of hidden "collaboration friction costs". A possible solution to this problem is "Collapsing the Talent Stack", which I have discussed in detail in my previous article " A Little Thought (2) - "Collapsing the Talent Stack" ".

  
  

• The neglected “superpower requirements for AI product managers”

       In the GenAI era, AI product managers must not only understand users, markets, and design, but also have a considerable understanding of AI’s technical principles, capability boundaries, data requirements, and ethical risks, and be able to effectively “translate” and “glue” team members from different backgrounds. This “super connector” role is extremely rare, and in fact, this role is not necessarily limited to product managers.

  For example, the head of product at Revolut, a global financial technology company, recently proposed a " Local CEO " model     in an interview . In plain words, there is one person who is responsible for a product or project end to end. I think in the future intelligent era, no matter what role you played before, you may become this "Local CEO". This kind of "super connector" talent who can handle multiple complexities is extremely scarce in the current market (I haven't seen a very good one yet).

  
  

• The future of smart organizations

      Some time ago, Professor Zeng Ming shared two most important points in his latest speech "Smart Economy and Organization":
  

      1. “What percentage of the company’s business is independently operated by AI?”

      2. “What percentage of the company’s internal employees are silicon-based?” 

      

      When AI can automate many repetitive mental tasks, what will be the core talent needs of future organizations? Will it be stronger critical thinking, creativity, emotional intelligence, or the ability to deeply collaborate and co-evolve with AI? How should companies plan their talent strategies in advance?

Summary: Key Takeaways

    Building a 2B big-model product is not an easy task, whether it is within an enterprise or a startup. Before you start, it is recommended to seriously consider those "easily overlooked" obstacles:

1. From the data perspective , the difficulty of "data alchemy" to transform raw domain data into high-quality "AI fuel" is generally underestimated. The data heterogeneity, isolation, and business-model semantic gap unique to the B2B industry place extremely high demands on the construction of "domain data advantages". If this is ignored, GenAI products will be like building on quicksand.

2. From the model level , blind optimism about the capabilities of general large models has led to a "capability trap". How to fill the gap between general intelligence and domain intelligence and ensure the applicability, controllability, stability and security of the model in a specific 2B scenario is far more complicated than imagined. There will always be a big contradiction between the excessive reliance on "black box" large models and the needs of 2B businesses and customers for certainty and explainability, which needs to be seriously resolved.

3. From the product level , the challenge of the "last mile" lies in how to invisibly integrate GenAI into and reshape the core business process, rather than simply providing a "plug-in" tool. This requires profound business process reengineering and user mind reeducation, as well as vigilance against interaction efficiency traps and careful construction of user trust. The difficulty of value quantification and the exploration of sustainable business models are the key to the survival of 2B GenAI products.

4. From the perspective of organization and talent , the impact of GenAI goes far beyond the tool level. It requires enterprises to "collapsing the talent stack", breaking down organizational silos, cultivating cross-border talents with "AI thinking", and promoting cultural change from top to bottom are the only way to overcome "cultural inertia" and "capability gaps".

    These challenges do not exist in isolation, but are intertwined, forming a systemic dilemma for the implementation of B2B GenAI. Any "ignorance" of any single link may trigger a domino effect, leading to the failure of the entire GenAI product or project.