The Real Battleground for AI Is Deep Inside Organizations
AI’s impact depends on embedding models with teams, data, and decision-making, not just deploying larger models or APIs.
This article originally appeared on Weijin Research on Sina on May 31, 2026. Original Chinese title: 「AI落地的真正战场，在组织深处」. It has been translated and adapted for an English-speaking audience.
Hello everyone, today I’d like to share some recent observations and preliminary thoughts.
Over the past few years, we’ve kept a close eye on how far large model capabilities can be stretched, and watched the market get excited about every performance jump. But at the same time, I feel more and more strongly that the friction in actually putting AI to work remains enormous. Turning AI’s technical ability into real enterprise value isn’t something you do by opening a chat window or plugging in a model. It demands a rethinking of business models, organizational processes, system integration, customer contexts, and even the way a company competes.
So we need to ask a more concrete question: how does AI value actually transfer inside an enterprise? How do AI technology companies work their way into a client’s organization? And how should traditional software firms and industry players turn AI into their own capability rather than being replaced by an AI platform?
The AI revolution is not just about large models. It’s a full technology stack.
When we talk about AI, we can’t just stare at the large models themselves. Jensen Huang once described AI as a “five-layer cake,” spanning energy, chips, infrastructure, models, and applications — a complete tech stack for the AI industry. But from an enterprise adoption angle, the stack needs to be sliced even finer.
Beyond hardware, infrastructure, models, and applications, getting AI into a company also requires a data layer, an execution layer, and a deployment layer. The execution layer, in particular, is critical.
Right now we pay a lot of attention to how a large model’s benchmark scores have improved. But when AI gets deployed inside a real company, the problems are a lot messier than model capability. Having a more advanced model doesn’t automatically solve enterprise problems. What companies need isn’t just a chatbox. They need a whole toolchain, a working environment, and an execution system built around the model.
Lately the industry has been tossing around concepts like Harness, Skills, and agent operating systems. All of them point to the same thing: the large model itself is only the core capability. Real-world deployment still requires a skills framework, tool calling, context engineering, permission management, workflow orchestration, and system integration. There’s a huge amount of concrete, tedious, even dirty work involved.
In that sense, AI is like every other general-purpose technology in history. Before it truly penetrates every industry, it has to go through a complex process of engineering, organizing, and adapting to real-world contexts.
From technical storytelling to real business spending
Looking at the Silicon Valley market since the start of this year — and including some Chinese large-model companies — AI company valuations have taken another leg up. A big driver behind that is agent technology, especially the breakthroughs in coding agents.
Products like Claude Code have essentially shown that agents can handle coding tasks end to end. Coding is a universal capability across digital technology. Once agents crack it, they will seep into a wide range of verticals inside knowledge work.
Companies like Anthropic have also let the market see that an AI-native business model is getting early validation in enterprise services. Its annualized enterprise revenue has surged over a short period, and that explosive growth has boosted capital-market confidence in AI enterprise services. Whether in the U.S. or China, the recent spike in large-model company valuations is closely tied to that confidence.
But that raises a new question: how do AI companies actually get inside a client’s organization and keep creating value?
A notable trend has appeared in Silicon Valley recently. OpenAI, Anthropic, and Google Cloud are all, in their own ways, doubling down on a particular role: the FDE, or forward-deployed engineer.
These engineers aren’t your standard pre-sales people, and they’re not consultants who just churn out slide decks. They go deep inside the customer’s organization, helping define problems, understand business processes, build systems, deploy models, and turn solutions into new product abstractions.
This approach borrows heavily from Palantir.
Why the Palantir model is getting a second look
Palantir is a controversial company, but it has built a deep track record in data analytics and agent deployment inside U.S. defense, intelligence, government agencies, and large Wall Street firms. Its core capability isn’t just selling software. It’s tightly integrating software, data, processes, and organizational decision-making.
Now OpenAI, Anthropic, and Google Cloud are all learning from that model. The reason is simple: enterprise B2B services have already proven to be a major direction for AI commercialization. And if you want to serve large enterprises, just offering an API or a model is nowhere near enough.
What large enterprises really need is a “software engineering construction crew.” This team goes into the client’s organization, helps nail down the problem, untangles the data and processes, embeds AI capabilities into business systems, and continuously forms new product abstractions through the deployment process. In the end, it creates value for the client while also strengthening the AI company’s own platform.
It’s a closed loop: enter the customer’s organization, understand the business problem, complete system deployment, form product abstractions, then strengthen the platform.
Some large Chinese tech companies and software firms are now studying the Palantir model. It looks a bit like a consulting firm, a bit like a software company, a bit like a system integrator. But what it delivers is not a one-off project; it’s an AI system that can run continuously, receive feedback continuously, and evolve continuously.
This may be one of the most important shifts coming next in the AI value chain.
Companies deploy AI for reasons beyond cutting costs
For the past three years, the AI industry has mainly told a technical story: stronger models, more parameters, faster inference, lower costs. But as we move into 2026 and agent technology becomes more usable and accessible, the market is starting to ask a different question: how does AI create real business value?
Many people assume that companies deploy AI primarily to reduce costs. In reality, adopting AI often increases costs at first.
Companies need upfront technology investment, have to hire strong talent, must rework legacy systems, and need to reorganize data and processes. In some scenarios, token costs may even exceed existing labor costs. So explaining enterprise AI adoption purely through “cutting costs” often falls short.
What companies care about more is: Can AI change how I compete? Can it stretch my business boundaries? Can it reshape my organizational processes? Can it get me into markets I couldn’t reach before?
At the same time, large models are becoming more and more similar. For many companies, picking a particular model may not, on its own, create a clear edge. The real differentiation comes from how the model is embedded into a company’s workflows, how it understands industry semantics, and how it drives organizational action.
That’s why Palantir’s concept of Ontology (本体), an organizational twin, deserves attention.
Palantir’s core idea is Ontology. The word means many things, but I prefer to think of it as a kind of “organizational twin.”
Ontology is not just a collection of enterprise data, and it is not a static database. It’s a unified model inside the platform that defines business objects, their attributes, relationships between objects, action rules, and permission logic.
Its goal is to create an organizational twin that reflects and drives a company’s real business processes in real time.
That means Ontology does not start from data. It starts from decision-making. It’s not built to store data. It’s built to abstract how an organization runs, makes judgments, and acts into a system model that AI can understand and execute.
Without Ontology, a software company can easily degenerate into a giant systems integrator. It can do data ingestion and system deployment, but it struggles to genuinely change how its customers make decisions and take action.
When AI is deployed inside an enterprise, depth of embedding is crucial. I see five layers to that depth:
The first is the data layer. The company brings various data together. This is the shallowest layer.
The second is the semantic layer. AI must not only see the data, but also understand what that data means inside the business.
The third is the process layer. AI must enter the company’s specific workflows, not stay outside as a question-and-answer tool.
The fourth is the action layer. AI must be able to guide the next step, and even execute actions within authorized limits.
The fifth is the governance layer. The AI system must be controllable, auditable, and accountable. It has to work in heavily regulated environments like finance, risk control, healthcare, energy, and government services.
Governance is especially important in finance and risk management. AI cannot be just a large model bolted on to existing work, nor can it simply turn data into a pool. It must be embedded in processes, understand the relationships between different roles, departments, and permissions, and guide the next step in business operations.
The depth at which agents reorchestrate organizational processes determines whether they can create real, sustained value for clients.
The value of an AI platform lies in turning a large model into a controllable system.
Palantir’s AIP, its AI Platform, points in an important direction. An AI platform does not just hand a large model to the enterprise. It turns that large model into a controllable, auditable, and executable agent system inside the enterprise.
That process includes model access, context engineering, ontology modeling, business applications, and automated execution. The aim is to transform model capabilities into operational capabilities within enterprise workflows.
From this angle, AI deployment is not about “models replacing people.” It’s about “models entering the system.” It’s not a single-point tool. It is an operating system embedded in an organization’s structure, processes, and governance.
Two kinds of companies in the AI value chain: “AI+” companies and “+AI” companies.
We can understand company types along two directions.
The first type is the “AI+ company.” These companies are themselves providers of AI technology and intelligence. They primarily make money from models, tokens, APIs, and agent platforms. OpenAI, Anthropic, Google DeepMind, and DeepSeek are all in this category.
They possess underlying model capabilities and may expand upward into knowledge-work fields like coding, finance, consulting, and enterprise services. Products like Claude Code are already unsettling many software tools, development platforms, and SaaS companies.
The second type is the “+AI company.” They come from every industry. They bring their own industry customers, business processes, and domain expertise, and they are now using AI to strengthen their business capabilities. Most of these companies are not model companies, but they control industry entry points and application scenarios.
Between these two groups sits a large number of vertical software companies and industry SaaS players. They are neither underlying model companies nor end-industry customers; they are the ones who own industry workflows and software access.
These companies face direct and pressing challenges from AI.
First, their existing software interfaces could be bypassed. Instead of opening one piece of software after another, users can accomplish tasks directly through agents.
Second, workflows could be absorbed by large platforms. Once an agent platform masters process orchestration, it can take away value that once belonged to vertical software.
Third, differentiation could weaken. If model capabilities grow more and more alike and agents reorganize the original functions of software companies, those companies will have to answer a new question. Where exactly does their moat lie?
But there is no need for excessive pessimism. These companies also hold very important and inherent advantages.
First, they know their industries. Large model companies possess huge amounts of data, but vertical software companies understand industry semantics, business objects, rules, and real-world processes.
Second, they control the entry point to customer systems. They are already deployed inside client organizations, with established relationships, usage habits, and trust.
Third, they have compliance and governance experience. In highly regulated sectors like finance, healthcare, energy, and government services, the organizations that truly understand industry rules, lines of accountability, and auditing requirements are often not the general-purpose model companies. They are the software and technology firms that have served those industries for years.
Take healthcare as an example. The large model companies keep emphasizing that AI can assist doctors and improve health services, but labor productivity in U.S. healthcare has not meaningfully risen because of it, while costs continue to go up. This shows that in fields with heavy regulation, deep processes, and strong accountability, AI does not automatically create value just because the model is powerful. What creates real value is systemic embedding that goes deep into the industry, understands the processes, and masters governance.
For vertical software companies and industry technology service providers, the key is not simply plugging into a model. It is turning their own strengths into new moats in the AI era.
First, business objects need to be turned into an ontology. Every company possesses the objects, relationships, rules, and actions within its industry. Vertical software companies are best positioned to abstract these elements and form their own industry ontology. That means turning business processes, decision logic, permission relationships, and action rules into system models that are programmable, callable, and executable. This may be the key for industry software companies to maintain their advantage.
Second, AI must be embedded into workflows, not confined to a chat window. When enterprises deploy large models, they cannot focus solely on the model itself. Bringing a model into the enterprise requires supporting tools, work environments, operating systems, permission structures, and industry knowledge. In many cases, these capabilities are more important than the model itself. Truly valuable AI does not just answer a question. It enters the process, understands context, invokes tools, drives action, and generates feedback within the system.
Third, model suppliers must be diversified. Enterprises should not rely excessively on a single model provider. As the model provider delivers services, it is also learning and understanding industry knowledge. If an enterprise hands over orchestration authority, process control, and data entry points entirely to a single model platform, it risks losing its core control. Therefore, enterprises should be compatible with both closed-source and open-source models, maintaining flexibility at the model layer. The true moat lies not in the model itself, but in industry data, industry semantics, business closed loops, permission systems, and customer relationships.
The core of AI implementation is not a single-point breakthrough in model capabilities. It is how the model enters enterprise systems, understands industry semantics, reorganizes business processes, and becomes a controllable, auditable, and executable agent system. For AI companies, the key to future competition is not just the model, but the ability to enter customer organizations, solve real problems, and form product abstractions. For vertical software companies and industry technology service providers, AI is both a challenge and an opportunity. Interfaces may be bypassed, processes may be absorbed by platforms, but industry semantics, customer trust, compliance experience, and process control rights remain crucial assets.
In the AI era, true enterprise value will not come solely from a more powerful large model. It will come from the deep integration of models, data, processes, actions, and governance. Whoever can orchestrate these capabilities will capture real business value in the next phase of AI implementation.