AI Reshapes the Flow of Capital: Budget, In-house R&D and the Last Mile

The ToB market is undergoing drastic changes, and service providers have adopted brand-new operational models.

"80% of e-commerce customers reported that they have clear reduction targets for software and infrastructure budgets."

"For information systems composed of menus and hyperlinks, AI makes their development extremely easy nowadays."

"Previously, what enterprises pursued was a tool, but now what they want is a verifiable result."

……

This year, the flow of capital in part of the enterprise market is shifting: in the fields of e-commerce, marketing, and office-related lightweight systems, budgets for software and infrastructure are being cut.

The freed-up capital is flowing into AI. In the past, enterprise procurement was driven by mature business systems, where AI features were merely highlights or free add-ons, but now AI itself is driving enterprises' procurement decisions.

At the same time, AI Coding has sparked the impulse of some large enterprises to develop in-house solutions in multiple scenarios such as marketing, office administration, and customer service. Low-threshold scenarios with short links, low barriers, and no reliance on enterprises' deep private data have been broken through by self-built systems, while business owners are asking: "Where is the tangible outcome of the money we spent?"

Changes in budgets and capital flow act like a baton, driving adjustments in part of the software and service market. According to observations from Digital Intelligence Frontline, many service providers have clearly realized that the original service model no longer works.

Split products into finer parts? Make services more in-depth? Deliver tangible results?

A full reconstruction is underway.

01

Changes and Invariance After AI Reshapes Budget Perceptions

In the first half of this year, a digital intelligence service provider for enterprises observed a very obvious signal from its e-commerce clients: About 80% of e-commerce customers reported that they have clear reduction targets for software budgets.

"There were signs in the previous two years, but this trend is particularly prominent this year." The vendor told Digital Intelligence Frontline that this trend not only appears in enterprises under profit pressure, but also in some enterprises with relatively good profits.

However, some people hold the view that budgets have not disappeared from these enterprises, but the allocation method inside many enterprises has changed, and some enterprises are moving budgets to AI-related fields.

Digital Intelligence Frontline learned from the industry that the penetration of AI agents is accelerating this year. The headquarters or top leaders of many large enterprises have generally put forward new requirements: some work links must be solved with AI — a home appliance enterprise requires certain processes to be fully AI-enabled, and a large clothing enterprise requires a certain proportion of new product designs to be completed by AI.

AI coding has greatly reduced software development costs, and is also rewriting the mindset of budget decision-makers. "They may not know whether the specific effect of using AI is good or not, but they feel that large models have lowered the barriers of many types of software, so the budget for these software should be reduced and allocated to AI." observed Liu Jingyi, senior product operation expert of Linyang AgentOne.

This change driven by top-down mandatory indicators has been confirmed from multiple sources. Sun Linjun, CEO of RealAI, an enterprise-level AI agent vendor, also found that agents are greatly replacing the construction of traditional IT systems in industries such as e-commerce. In Q1 this year, RealAI's agent business in the e-commerce sector achieved 150% growth, with a large number of new customers who signed contracts immediately after the first meeting. Sun Linjun analyzed the reason and said, "The explosive popularity of crawfish has educated the market. Enterprise owners have done relevant market research and usually come with clear demands."

Many owners of traditional industries also responded quickly. "This year, many traditional industries such as exhibition and real estate have shown obvious interest in AI, and are willing to invest large budgets in project transformation," a person from the enterprise market feedback.

Wang Ting, head of customer AI success at Linghe Digital Intelligence, an AI service provider for the manufacturing industry, has contacted a large number of traditional enterprises. She observed that AI has become a very clear decision point in enterprise procurement this year.

Wang Ting experienced the change node when technologies such as ERP entered enterprise procurement decisions during the process of enterprise digital transformation. She deeply perceives that the current enterprise procurement decision for AI also has such periodic characteristics. "In the era when ERP rose in 2008, the industry said that implementing ERP is seeking death, while not implementing ERP is waiting for death. Today, AI has reached the same decision node. Many business owners have reached a consensus on whether to adopt AI, and they are only struggling with how to implement it, when to implement it, and what conditions are required."

Some overseas survey data even show the proportion of budget reallocation, where enterprises shift recruitment and software expenses to AI. Goldman Sachs released a more specific figure in a CIO survey in May 2026: only 33% of enterprises' AI token budgets come from new appropriations, and the remaining 66% come from the reallocation of existing budgets. There are two main sources: one is the labor budget, and the other is the application software budget.

However, senior industry practitioners believe that this budget shift mostly occurs in relatively lightweight software system departments such as marketing, customer service and office administration. In fields involving complex production and manufacturing, bank credit finance, and core ERP, finance and tax, legal affairs of large enterprises, mature enterprise system software still dominates, and AI penetration is still in the early stage.

Digital Intelligence Frontline contacted the big data department of a leading bank, and they did not even use AI systems to process data, because traditional data analysis methods are very mature, and the financial system has high requirements for accuracy, so they are very cautious about introducing AI.

An ERP vendor also mentioned that they are promoting AI capabilities in the ERP field to build benchmark customers, and they have co-created with benchmark customers for a year, but the speed of covering large customers is very slow, and it is difficult to achieve rapid volume growth like in the office field.

02

Customer In-house Development Trend: Lightweight Scenarios That Are First Broken Through

Closely related to the changes in enterprises' budget perception and procurement willingness for "lightweight systems" is the strong in-house development impulse of demand sides after AI Coding capabilities are enhanced.

Liu Jingyi observed, "Customers are educated to get used to conversational interaction, and when they encounter some common products, they will think that they can buy some cloud servers and computers to develop them by themselves." Moreover, due to the improvement of agent capabilities, the internal demonstration speed of enterprises is extremely fast, which leads to the fact that in the POC stage, service providers no longer only compete with external opponents, but often have to compare effects with tools built by customers themselves.

Xiao Yuyan, deputy general manager of NetEase Enterprise Services and head of cloud business, also told Digital Intelligence Frontline at the end of May this year that the very direct challenge they felt this year did not come from competitors, but that some large enterprises are considering developing some modules and product functions in-house. "There may be a considerable technical team inside the customer, and they need to prove their value."

The in-house development impulse has also spread to the government side. Fang Yi, founder of Merit Interactive, gave an example to Digital Intelligence Frontline that the person in charge of a local development zone no longer approves the information construction requirements submitted by subordinates, because someone in the office has directly developed the system with AI. When planning and constructing similar government systems, some government departments will actively learn from the mature practices that have been implemented in other regions, use AI tools to shorten the early-stage research and prototype construction cycle, and then make adaptive adjustments combined with their own needs.

Fang Yi believes that AI has also impacted the bidding and tendering form itself. "The detailed function point descriptions in the bidding documents can speed up the early demand sorting and development verification with the help of AI tools", and even the bidding documents and tender documents themselves have introduced AI assistance.

The scenarios that are first broken through share common characteristics: short links, low barriers, and no reliance on enterprises' deep private data. For example, in the after-sales evaluation link of e-commerce, Liu Jingyi introduced that for some scenarios of analyzing individual product reviews, customers tend to build an agent by themselves, "It only costs at most a few hundred yuan in computing power, and the old logic of selling a complete product worth tens of thousands or hundreds of thousands of yuan no longer works."

For complex systems such as energy and chemical industry, industrial software, and core ERP modules that are heavily bound to process flow, industry knowledge and private data, the possibility of AI replacing them is still low. An enterprise-level AI agent vendor introduced that in some traditional manufacturing industries, for high-complexity scenarios and systems that support the current core business operation of enterprises, enterprises usually dare not directly use AI to replace them, but use interfaces or RPA methods to call relevant data. "The interaction method may change, but they dare not modify these assets and systems casually."

In the process of exploration and trial and error, we can see that the supply-demand relationship in the market is changing rapidly. Service providers can also observe that many enterprises are swinging back and forth between in-house development and external procurement.

Wang Ting from Linghe Digital Intelligence met an enterprise with annual revenue of nearly 20 billion yuan, which formed a 5-person AI team and used an open platform to build various tools for more than a year. The chairman commented that there was no effective result except that marketing could quickly generate images. Then this enterprise turned to seek external teams to take over. "We have encountered more than one or two such enterprises," Wang Ting said.

As the threshold of software development drops sharply, IDC states that enterprises' AI procurement is increasingly focusing on use efficiency, cost control and quantifiable implementation results. Many service providers have felt that the procurement standards of enterprises for external service providers' products also require clear output and "results".

Cheng Weizhong, founder of Zhongke Deepvision, does not shy away from the challenges he encountered. In the past six months, Zhongke Deepvision's traditional digital human business with tool attributes has been hit greatly. "Previously, what enterprises wanted was a tool, but now what they want is a result. Having a digital human in the e-commerce live streaming scenario does not mean that goods can be sold." Cheng Weizhong believes that not only digital humans, but the entire SaaS industry needs to face the interrogation from customers about whether the products of service providers can bring tangible results and effects.

The emphasis on ROI and value makes a large number of traditional industry owners hardly promote AI in their enterprises with the "token maxing" attitude adopted by large Internet companies at the beginning of this year.

Wang Ting from Linghe Digital Intelligence relayed a case of a manufacturing customer's boss to Digital Intelligence Frontline. The enterprise spent 50,000 yuan on an office AI product and stopped the cooperation after half a month. "We didn't even hear any echo after spending 50,000 yuan. It seems that even investing 500,000 yuan may not bring any tangible results." In her opinion, this is not a problem of tools or platforms. If enterprises cannot measure the value brought by AI after purchasing it, it is naturally difficult for them to really pay for AI.

03

Responses of Service Providers: Refine Products, Deepen Services, Deliver Tangible Results

Based on the real actions of different types of vendors in the past six months, Digital Intelligence Frontline found that agent service providers, in order to adapt to the market's in-house development trend and the requirements for delivering tangible results, are making adjustments from multiple levels such as customer stratification, finer product granularity and result delivery to match market demands.

Take Linyang AgentOne as an example, they are cutting the granularity of service scenarios into smaller parts. Previously, they used the entire product or the entire platform to serve customers. Now, in line with the AI era, they need to split product skills and MCP interfaces. Linyang told Digital Intelligence Frontline, "Only with smaller granularity can we find more customer entry points in the market."

In this process, customer stratification is inevitable. Digital Intelligence Frontline learned from Linyang that a small number of enterprises at the top of the pyramid have built many agents internally, and they look for service providers to make up for "the last missing skill in the last mile". Linyang's agents can be connected to the enterprise, and realize skill complementation through agent-to-agent calls.

The second category is the current mainstream type. For example, some e-commerce enterprises have received the top-down proposition of "using AI to improve e-commerce operation efficiency", and opened the operation scenarios of several e-commerce stores. A dozen of operation specialists need to upload images, modify prices, sign up for activities, and launch advertisements every day, and the enterprise needs to improve efficiency and reduce costs in these deterministic scenarios. For such customers, Linyang can provide a managed digital employee solution to serve customers, similar to "AI operation employees", allowing digital employees to "join the job" like real people, assigning a cloud-based employee computer, with factory-configured post skills, and then the enterprise configures the rules and permissions of the digital employees.

There are also some large customers with relatively vague demands. They only have the need to use AI, but they do not have clear ideas about how to use it and where to use it. For these scenarios, Linyang will send the FDE (Frontline Deployment Engineer) team to enter the site as expert roles for consultation and diagnosis, "accompany customers for the last journey", and then precipitate the scenario demands and functions back to the standard products in the enterprise.

Quite a few service providers are redefining customer boundaries.

Xiao Yuyan from NetEase Enterprise Services also saw two typical customer portraits this year, "One type says they are going to develop in-house, and the other type says they failed in in-house development and come back to us." For leading enterprises with in-house R&D capabilities, Xiao Yuyan believes that there are a large number of long-tail demands in these enterprises, which can be independently developed through the general middle platform capabilities inside the enterprise. In fact, the fully customized development model close to leading customers does not conform to the business model of most software service providers. Therefore, NetEase Enterprise Services will focus more on serving mid-to-low tier customers whose self-development resources and capabilities are inferior to large enterprises. In order to serve such enterprises well, they not only polish in landing scenarios and product capabilities, but also provide a series of tool combinations to help improve the AI operation capabilities after product delivery.

Deliver results in a more closed-loop way is also the strategy that Cheng Weizhong from Zhongke Deepvision is currently adopting. Cheng Weizhong introduced that around result-oriented product delivery, in the past six months, Zhongke Deepvision is transforming from pure SaaS sales of digital humans to effect-oriented services, focusing on building a closed-loop service link from traffic acquisition to conversion.

Cheng Weizhong believes that there are only two models for enterprise-level agents in the future: one is effect-oriented, and the other is light customization and private deployment, which are deeply connected with customer data and SOP. "Short-link workflows will basically be eliminated as the capabilities of large models upgrade."

Therefore, on the original position of digital human live streaming, the energy distribution of the Zhongke Deepvision team is different from the past. Only 20% to 30% of their energy is spent on the digital human product for live streaming rooms, and 70% of their energy is used to make up for the capability gap of the enterprise's goods sales business chain outside the live streaming room. For example, how to place advertisements, how to generate high-conversion-rate hit short videos, and then to the reception of AI digital human live streaming rooms, striving to achieve more complete link coverage, and help customers complete the closed-loop of product sales with digital humans.

In order to complete this closed loop, they have also opened up new product directions. At present, AI assistants are becoming new traffic entrances for various services