AI cloud: Nine out of ten enterprises have got it wrong.
My observations:
In the past year, the biggest pressure on the IT departments of many enterprises, especially those related to AI, has come from their bosses. The bosses believe that since we are involved in AI, we must achieve "cost reduction and efficiency improvement."
To put it more bluntly, they require everyone to use AI. In particular, they often ask the business teams to think about how to integrate AI into projects.
This pressure has also led to the emergence of a group of products centered around "technology demonstration" in the market. These products are not actually driven by "business pain points." For example, many are rushing to develop chatbots and automatic meeting minutes generators, but few stop to ask:
Is this really the biggest bottleneck hindering business efficiency improvement at present? How can we integrate old scenarios?
Some time ago, when browsing my WeChat Moments, I saw a report shared by a friend, the "Research Report on the Application of Artificial Intelligence by Chinese Entrepreneurs" issued by the Shell Finance of Beijing News in 2025. This report highlights this "panic" more clearly:
Nearly 90% of enterprises have embedded AI into at least one business operation process. It seems like they are moving very fast.
However, another set of data in the same survey is discouraging: Only 16% of enterprises have established a dedicated AI team, and more than 40% of enterprises haven't even started AI - related capacity training. Digging deeper, only about 12% of enterprises have established an AI governance system, and more than 60% are still in the preliminary exploration stage, or haven't even made a plan.
To put it simply, everyone is taking action, but most are doing it in a panic and a haphazard way. There are investments, but no system is established; there are projects, but no strategy is clarified.
Of course, it's already 2026, and the 2025 report is still relevant. The root cause of this panic is often that people have the wrong idea from the start:
They regard the AI cloud as a "more expensive traditional cloud," thinking that it simply means adding a few GPUs to the original cloud servers, running a large - scale model, and providing APIs externally. Essentially, it's still the same old way of renting computing power and buying resources.
If that were the case, there would really be no need to rush. One could just wait until the price drops and others have figured out all the pitfalls before following suit.
1
The AI cloud is not like that at all. Many enterprises have the wrong mindset. To understand what the AI cloud really is, we need to first clarify what problems the traditional cloud solves.
Actually, it's very simple:
Enterprises no longer want to build their own computer rooms. They move servers, storage, and databases to the cloud, using them as needed and paying according to usage. Ultimately, it's the outsourcing of IT resources, like buying infrastructure, which is essentially no different from renting a warehouse.
However, the AI cloud goes a step further. It changes the way you access services and solve problems.
How to understand this? Let me give an example of a scenario that a CIO has probably experienced.
A medium - sized enterprise is running more than a dozen SaaS systems simultaneously: CRM has one set of logic, ERP has another, and the HR system has yet another. The customer service department uses one tool, and the finance department uses another. They are barely connected through interfaces.
One day, the boss suddenly asks a cross - departmental question, such as "Why did the repurchase rate of large customers in the East China region drop last month?" Guess what happens? The three departments have to export data separately, piece together Excel spreadsheets, and hold two rounds of meetings to come up with a rough answer.
I remember that in 2020 when I was at the company, meetings were like this:
The marketing department would bring a report, the operations department would export a set of data from the system, and the growth department would export another. Then we would sit together and discuss. We either used the same system or pieced together data from several systems. Looking back now, it's quite funny.
In the past decade, when enterprises were digitizing, in essence, they "bought ten systems and then patched them everywhere." Each system solved a local problem, but the gaps between the systems had to be filled by people. These gaps are called IT silos, which are an inevitable product of the entire delivery model.
I once talked to some friends who work in SaaS and asked them:
Is it because people don't know how to do better, or is it due to limited technical conditions? Their answer was very honest: It's a compromise due to historical limitations when the technology wasn't mature.
Now the situation has changed.
Everyone is talking about Agents, but I've found that many people misunderstand them. They think an Agent means creating a fully automated system to replace people in running processes all at once. This is unrealistic and not how Agents really work.
The actual path is not that mysterious.
Think about it. Most enterprises already have a bunch of business systems, like CRM, ERP, and work - order systems, with all the data stored in them.
The first step is to add an AI entry point, such as a Chatbot, on top of these existing business systems. Employees or customers can use natural language to search for information and ask questions, without having to jump between five systems. This significantly improves efficiency.
At first, this Chatbot may just be a question - answering window. Gradually, you can add more capabilities to it: it can query CRM data, generate reports, process Excel spreadsheets, and initiate approvals.
As these capabilities are added layer by layer, at a certain point, it will become an Agent that can handle an entire task for you. So, this is the logic of an Agent: First, let the old business acquire AI capabilities, and then build an Agent on top of these AI capabilities.
This difference is crucial.
In the past, software delivery was about "defining the structure first and making the business adapt to the structure." If you wanted to implement CRM, you had to re - organize the sales process according to CRM's logic. Now the situation is reversed: First understand the business intention, and then dynamically organize capabilities. Enterprises don't need to adapt to a certain software product; instead, let AI adapt to your business.
This is why the procurement logic of the entire industry is changing, from "What kind of cloud do I need to adopt" to "What kind of AI capabilities do I need." You are buying an intelligence that can be invoked, orchestrated, and integrated into your existing business processes.
I read a research report by Anthropic at the end of 2025, which surveyed more than 500 technology leaders in the United States. More than 80% of the surveyed organizations said that AI agents have already brought measurable return on investment.
Note, it's not "possibly useful," but "already generating returns."
So, if an enterprise CIO still evaluates the AI cloud using the "buying resources" framework, then what they are actually evaluating is an old concept that is disappearing.
2
After realizing that the AI cloud is different from the traditional cloud, many people's first reaction is quite reasonable: Since it's something new, I'll wait and see. I'll wait until the technology is more mature, the price is lower, and others have figured out all the pitfalls before entering the market.
I completely understand this idea. To be honest, five years ago when cloud computing was just emerging, it was okay to wait. After all, everyone was exploring, and there wasn't much difference between entering early or late.
However, the cost of waiting for the AI cloud is different from what you might think.
Most people think that waiting can save money and trial - and - error costs. This seems reasonable on the surface, but the problem is that what you lose during the waiting period is invisible.
What do you lose? I'll just mention three important things.
First, the time for data governance. AI can't be used right away; it needs data. How many systems is your data scattered in? Is the format unified? Is there a standard? Who is responsible for managing it? No cloud provider can solve these problems for you overnight.
It takes at least six months and up to one or two years for an enterprise to start data governance until the data is truly "ready to be fed to AI." This training time can't be rushed; it has to be accumulated day by day.
Second, the time for team learning.
The implementation of AI is not just the responsibility of the technology department. The business team needs to learn how to collaborate with AI, how to put forward requirements, and how to judge whether the results given by AI are reliable. This ability has to be accumulated bit by bit in real - world scenarios. If you don't start, how can you accumulate it?
Third, and most easily overlooked: the compound interest after a scenario is successfully implemented. AI has a characteristic that once a certain scenario is successfully implemented, the user's usage and feedback will make the system better and better.
Knowledge is accumulated, the model is optimized, and the process is refined. Enterprises that successfully implement the first scenario can replicate it to the second and third scenarios at an increasingly faster pace. And you haven't even started.
A survey by McKinsey and several consulting firms in 2025 made a judgment. To be honest, I found the data quite shocking: In enterprise - level AI projects, only about one - third have entered full - scale production, and only about one - quarter have achieved the expected revenue return.
On the surface, this seems to suggest that "AI isn't that effective," supporting the idea of waiting longer.
But if you look closely at the enterprises that have actually achieved returns, they have one thing in common: they completed their preparatory work earlier. Data, processes, talent, and governance - these four elements. The industry calls such enterprises "AI - Ready organizations."
Many enterprises that are still at the stage of "creating a demo, holding a few rounds of reporting meetings, and getting some PPTs showing successful pilot projects" are directly warned in the report: They will be left behind by their peers starting in 2026.
There is an even more startling statistic. Research by Gartner and MIT shows that 95% of generative AI pilot projects have failed to go into production. Yes, 95%. This figure really shows that most enterprises have taken action but have the wrong mindset.
3
So where exactly did they go wrong? After looking at many cases and surveys, I found that although the stories of enterprises that have fallen into pitfalls are different, they all make the same fundamental mistake: They use the mindset of buying equipment to buy intelligence.
The most typical example is that the IT department makes a technically impressive product, but the business department doesn't care at all.
There is a large manufacturing enterprise (I won't mention its name to avoid offending anyone) that spent millions on an "enterprise knowledge assistant."
The functions were quite comprehensive. Employees could ask about company policies, check leave regulations, and learn about the reimbursement process. When it was launched, three rounds of promotional emails were sent internally. In the first month, some people tried it out, but by the third month, when checking the backend, the daily active users were in single digits. The project was quietly shut down.
It's not that the technology was bad. What employees really struggled with was the fault diagnosis of production - line equipment, which was a money - burning issue every day. But it was the IT department that made the decision, and they chose the scenario that was easiest for them to implement and showcase, not the one that was most painful for the business.
This kind of thing happens a lot. Another common situation is that after the boss decides to "fully embrace AI," the subordinates try to connect all business processes to AI at once.
As a result, they find that the general large - scale models have serious "hallucinations" in their industry. The information they provide is unreliable, and they can't connect to internal data and APIs, so they are useless.
This is the procurement mindset: Thinking that buying the most powerful thing will solve all problems. But AI is not a piece of equipment that can start working as soon as you plug it in. It's more like a new employee. You have to tell it what the business is like, where the data is, what is correct, and what mistakes it can't make.
Then, when choosing a platform, everyone compares the number of GPUs, the price of Tokens, and the benchmark scores.
But a Gartner survey shows that in more than 68% of enterprise platform selections, the real bottlenecks are that the platform can't connect to existing systems, can't ensure data security, and the operation and maintenance costs far exceed the budget.
So when choosing a platform, don't just compare parameters. Just ask three questions:
How difficult is it to integrate? Who will be responsible if something goes wrong? Can my existing systems work with it? These three answers are more useful than any benchmark scores. In the end, although these pitfalls seem different, they all stem from the same root: still using the mindset of buying hardware or software to make a decision about "buying capabilities."
So how should you think about it?
Just one question: In your business, which process takes the longest time and the highest cost to solve when a problem occurs? Start from that process.
Don't think too big; don't pursue full - scale intelligence. Use the minimum investment to successfully implement a real - world business scenario and get a quantifiable result. Use this result to convince yourself, your team, and your boss.
Before taking action, go back and do one thing: Inventory your data. Can your data be used by AI? How many systems is it scattered in? Is there a unified standard? Who is in charge?
After chatting with dozens of global enterprise executives, Accenture reached the same conclusion: The bottleneck for the large - scale implementation of AI lies in the organization. The approval processes and control mechanisms designed for business stability have now become the biggest obstacles.
90% of enterprises are already taking action. But the difference between taking action and taking the right action lies in the way of thinking.
I have a similar experience myself. I often attend industry conferences. Some conference organizers now put the registration forms directly on the AI platform. How do they do it? They set up a landing page in the front, connect it to a form in the back, and link them together with two developer tools. Participants scan the QR code, sign in, and the data is collected.
This is a very simple thing, just a registration process. But if you think one step further: After the data is collected, some of it can be directly analyzed by AI, such as:
The industry distribution and areas of interest of the participants; some can be given to the sales team to follow up by phone for targeted outreach. Later, the sign - in data, interaction data, and conversion data from follow - up can be used to optimize the planning of the next conference.
You see, starting from a simple registration form, a whole set of customer acquisition and conversion processes can gradually be developed. This is the logic I mentioned earlier.
Well, I've explained it in plain language. Whether it's MCP or various new technologies, the development of external terms and technologies is always uneven. But from the perspective of practical use, the key is whether you can successfully implement a business scenario and use this scenario to drive returns. That's definitely the right approach.
When everyone enters the market, it will be about who has a deeper understanding of their own business. I think this understanding mainly boils down to three things:
- How to process data
- How to process knowledge
- How to develop an Agent based on these
Once you figure out the first two, the organization will evolve quickly, as the value of an organization is to serve the business.
This article is from the WeChat official account "Wang Zhiyuan" (ID: Z201440), written by Wang Zhiyuan and published by 36Kr with authorization.