Enterprise-Level Artificial Intelligence ｜ How Can Your Enterprise Become Intelligent?

Today, I conducted a workshop on enterprise AI applications with a team of managers at a large enterprise.

First of all, I'd like to say a little bit about my preparation for this seminar: in the last two years, I've been talking about enterprise-level application of AI on various occasions, and I feel that the iteration of AI technology is too fast - last year, I was still talking about how to use generative AI in the enterprise , but today, I think that these contents are almost all out of date; with the enhancement of the model's inference ability and the enhancement of the thinking mode, while the large language models have function function, I think that the AI application of enterprise-level AI can be improved. With the enhancement of the model's reasoning ability and thinking mode, and at the same time, the large language models have function-calling ability, the intelligent body application has become a mandatory option for enterprises to use the large language models.

To prepare the PPT for this workshop , I looked at the PPT I wrote six months ago , and most of the pieces are no longer usable today, so I decided to turn to Manus .

I wrote a long prompt, "Let's say I'm a consultant and I want to talk to the business management team about the latest technological developments in Artificial Intelligence and how these new technologies can be applied to the business to innovate the IT architecture of the organization ...... Please help me generate a set of PPTs ".

Manus worked for half an hour and outputted a document (below). I feel that this is not up to my requirements - although the total logical structure of the document looks good, but the depth of analysis is not painless, the format is dull, the case looks like hearsay from the Internet, in short, is the level of an MBA student's homework, sorry for the client to pay my consulting fees, so I decided to rub a document by hand.

From this perspective, it is impossible for AI to replace consultants.

I started talking about BOAT last year , from that time onwards, I think the mainstream form of large models applied to enterprise information systems has been gradually clear, that is: will be based on workflow technology and deployed on top of the workflow process rules engine, RPA and so on the formation of the business process automation ( business process automation ), partially or completely by the large model-driven (LLM Agent ) to replace.

Intelligent body does not replace the traditional ERP, SaaS, microservices, and other record-based systems, the following figure is the principle of intelligent body. I often make an analogy, if the enterprise system is compared to a human body, artificial intelligence is like a brain, but the brain alone can not complete the task, but also need tools, the brain also needs blood nourishment, then ERP and so on is to complete the task of the tools and the human presence of the external environment, and enterprise data is equal to the blood:

From my experience with Manus , at least for the time being, intelligence can't completely replace human work, and the means of digitization is to complete the "mashup" of workflows and intelligence in the completion of a complex, long-cycle business process.

Each activity unit that constitutes a business process is called a "task", and a string of tasks constitutes a workflow. If it is relying on people to arrange , then it is a "workflow ( Workflow )", if in the process, to complete a specific goal, you need to rely on a large model of the dynamic arrangement of several tasks to achieve the goal , then the completion of the goal of the software unit is an intelligent body ( Agent ).

Two years ago, when the emergence of intelligent body, the architecture of intelligent body is: the large language model is used to talk to the user, according to the user's request for the goal to disassemble the task, need to be through the complex middleware framework to link the various tools to complete the task, see my two years ago, " Artificial Intelligence and ERP ｜ large language model how to reshape the enterprise-class IT applications ".

From the beginning of 2024 , the ability of new versions of large language models began to evolve, first allowing the model to generate structured tool call instructions, but still need to be manually parsed and executed by the developer , to the second half of 2024 , many of the new versions of the large language model has the native ability to directly call external tools and return the results, the degree of dependence on the middleware framework is reduced (the middleware is used to do the complex process orchestration); the use of these new types of With these new large language models of intelligences, the user only needs to declare the list of tools in the prompt , and the model automatically generates the invocation commands and processes the results , significantly simplifying the construction of intelligences.

The current leader in the development of intelligence is Anthropic, a macromodeling company that spun off from the founding team of OpenAI, and whose "enhanced macromodeling" to improve the capabilities of intelligence, shown above, was published at the end of 2024, is currently the most influential.

In the last two months, open-source protocols such as MCP by Anthropic, A2A by Google, and ACP by IBM have significantly accelerated the construction of an ecosystem of intelligence by standardizing tool calls and inter-intelligence communication. MCP lowers the barrier to integration of intelligences with external systems, while A2A and ACP promote cross-vendor collaboration of intelligences to form an Internet-like layered architecture. These protocols have to be supported by both industry vendors and the public.   These protocols still need to solve protocol compatibility and security governance issues to be supported by industrial vendors and engineer communities, but the technical foundation of the 'Internet of Intelligent Bodies' has been initially established, and it is expected to enter the stage of scale application of intelligent bodies by the end of 2025."

The current bottleneck for intelligent bodies to handle enterprise business lies in the intelligence of the large model that drives the intelligent body itself, i.e., its ability to think and reason.

The development of large language models applied to intelligent bodies has gone beyond the stage of comparing model parameters two years ago and doing math problems one year ago; benchmarking of large language model capabilities is shifting to measure the ability of intelligent bodies to use tools and handle end-to-end tasks over time in different domains; test cases focus on the robustness exhibited by intelligent bodies in edge scenarios (e.g., missing tools, extraneous queries, incomplete inputs); and attention is increasingly being given to The increasing focus on multi-round tasks requires intelligences to manage context, sequence actions, and adapt to goal evolution.

By comparing the performance of human experts with 13 cutting-edge AI models (such as the GPT series and Claude 3.7) at the end of 2024, the nonprofit organization METR found that AI had a nearly 100% success rate in tasks that took humans less than 4 minutes, but a success rate of less than 10% in tasks that took more than 4 hours. A research report in March 2025 showed that based on 170 real tasks (covering fields such as programming and network security), the core indicator is "50%-task completion time span", that is, the time it takes humans to complete tasks that AI can complete with a 50% success rate. In 2019, AI's ability to complete tasks that take humans 10 minutes can be expanded to tasks that take humans 20 minutes after 7 months; in 2024, this cycle will be shortened to 3 months. By analyzing the growth trend, it is predicted that by 2028, AI may be able to complete complex projects that take humans one month.

The ability of intelligence to handle complex tasks doubles every seven months, which is known as " Moore's Law for Intelligentsia ".

Currently, AI has surpassed humans in standardized short tasks (e.g., code optimization, data querying), but still lags significantly in long tasks (e.g., complex software development, design of scientific experiments) that require cross-domain integration and dynamic decision-making. Currently, AI is better suited as a "high-efficiency tool" than as a "full-process replacement".

At the seminar, a leader asked me, Technology iterates so fast, is it outdated if we invest in making intelligent systems now and do it well?

I agree with this point of view, so my advice to him was:

First, enterprises to introduce AI, in addition to technical breakthroughs at the level of modeling capabilities (I see that there are indeed enterprises investing in this), the use of existing AI capabilities to introduce intelligent body applications.

Second, don't carry out the kind of so-called AI planning that the leadership initiates, top-down, and looks for scenarios everywhere , and don't use the traditional heavy IT planning methods (such as enterprise architecture methods) to plan for intelligent body applications.

Third, to comprehensively popularize the ability to build an intelligent body in the internal staff, or to buy the intelligent body of external vendors, RaaS services, do not want to build everything themselves; first, use up, in practice, to innovate.

Fourth, the establishment of a mechanism to accelerate the release of the commercial value of innovation, so that the commercial value of intelligent applications in the enterprise snowballs.