The article delves into the common misconception about AI agents in the industry, which is that they solve all problems. IBM experts pointed out that although AI Agent is powerful, it is not omnipotent, and enterprises need to pay attention to its integration with core business and actual value. IBM’s watsonx Orchestrate provides a clear architectural design to help enterprises use AI agents efficiently, while emphasizing that technology should serve business scenarios, not just show off skills.
What is the biggest misconception about AI Agents in the industry right now?
“I think agents can solve all problems.” Wu Minda, a senior technical expert on data and artificial intelligence in IBM’s Greater China Technology Division, replied at a media roundtable a few days ago.
AI Agents are the “battleground” for almost all tech companies right now. IBM also launched an upgraded version of its AI agent solution, watsonx Orchestrate, during Think 2025: it provides pre-built, out-of-the-box professional domain agents (such as HR, sales and procurement agents, etc.); Support enterprises to build their own AI Agent in less than 5 minutes; Through the agent orchestration tool, the multi-agent and multi-tool coordination required for complex projects can be realized. It also provides observability across the entire AI Agent lifecycle, including performance monitoring, protection, model optimization, and governance.
The “inflection point” for AI agent-scale applications has arrived, and the industry consensus has arrived. However, IBM also emphasizes that AI does not need to be overly “deified”. “The essence of technology depends on whether it can solve the real problems of the enterprise, especially when it is bound to the core business, it is necessary to return to the business scenario to see whether the technology really generates value.” Zhai Feng, general manager and chief technology officer of IBM Greater China, said.
In other words, AI can’t solve all problems, and not all problems need to be solved with AI, and the same goes for agents.
“True and false” agents
Unlike traditional AI assistants (such as chatbots), AI Agents can not only understand instructions and generate content, but also autonomously plan task paths, call multi-system resources, and dynamically optimize strategies during execution based on real-time data.
B-end product manager’s ability model and learning improvement
The first challenge faced by B-end product managers is how to correctly analyze and diagnose business problems. This is also the most difficult part, product design knowledge is basically not helpful for this part of the work, if you want to do a good job in business analysis and diagnosis, you must have a solid …
View details >
These excellent characteristics have made the industry flock to AI Agents, and of course, many “old wine in new bottles” have appeared on the market – although they have changed the packaging of AI Agents, the core is still a traditional AI tool.
Wu Minda told reporters that it is not difficult to identify the so-called “fake agent”. “The computing power of pure ‘old wine’ is not when it is running (using), the agent has an autonomous ‘brain’, and it needs to keep thinking about things and calculating, and then it is necessary to stack computing power. However, in the past, automated processes or AI model calls were basically arranged in advance, calculated with historical data, and did not consume so many resources when running. ”
Through a unified platform on-ramp called AskIBM, IBM is also using AI Agent to empower employees internally. According to reports, AskIBM can automatically route to vertical field agents such as HR, IT, sales, and procurement according to employees’ query intentions, automating the whole process from problem resolution to system interaction.
In Wu Minda’s view, it is easy to build AI Agents, but if you want to do a good job in AI applications in enterprises, you need to achieve scale, which is difficult.
First, there are different frameworks, different applications, and different vendors behind the development of agents, how to connect them to each other?
Second, how can businesses find high ROI and suitable scenarios?
Third, how to manage the entire life cycle of an agent from construction, production, to O&M?
In response to these problems, watsonx Orchestrate has a clear architectural design, as shown in the following figure:
From a top-down perspective, IBM’s core ideas are three:
The first layer is the vertical agent matrix out of the box. Including the first batch of 3 AI agents released during the Think 2025 conference: HR agent, sales agent, and procurement agent. “For example, training, ID or permission applications after new employees are hired do not even require HR intervention, and the robot behind them can answer 90% of the questions. As for how many AI agents are called behind this, it doesn’t matter at all from a business perspective. Zhai Feng said. At present, the human resources agent has been officially launched, and the sales and procurement agent is also planned to be opened in June.
“For domestic enterprises, these agents can be used as templates when using them, and then adjust them according to their actual needs.” Wu Minda gave an example.
The second layer is multi-agent orchestration. Once the agent is developed and deployed to the Agent Catalog, it can be opened to the department or other employees, which is similar to an agent warehouse, supporting classification retrieval, permission management and version control, and administrators can also set access permissions and publish and share agents through the approval process.
However, as the number of enterprise agents increases from dozens to hundreds, the complexity of management will also increase. For example, after obtaining new customer leads, the sales agent will automatically trigger the market agent to analyze the dynamics of competing products, and then call the customer service agent to generate personalized follow-up strategies. In addition, whether it is built by the enterprise itself, by a partner, or by a specialized domain agent in the open source community, information can be shared and complex multi-step processes can be handled collaboratively.
The third layer is open ecology and open source collaboration. The front end is a unified entrance, and behind it is a very open agent ecosystem. watsonx Orchestrate integrates with more than 80 industry-leading enterprise-grade application tools from companies such as Adobe, AWS, Microsoft, Oracle, Salesforce Agentforce, SAP, ServiceNow, and Workday. For example, businesses can choose to call Salesforce’s sales forecasting agent directly in Orchestrate without having to repeatedly develop the interface interface.
Is the data AI-ready?
It should be noted that AI Agent requires a lot of knowledge behind completing a task, which may come from within the enterprise, the Internet or the large model itself.
Zhai Feng said that AI applications without data are empty talk, and enterprises must first ask themselves three questions if they want to implement AI: Is there high-quality data? Is this data in use? Does it really work?
In other words, having data does not mean being able to use it well. “More than 90% of the internal enterprises are actually unstructured data, but at present, everyone pays more attention to structured data.” “Therefore, helping enterprises improve the use of unstructured data is also IBM’s main focus. ”
Compared with the previous version, the data weft is integrated with data and latitude, and a semantic layer (watsonx.data intelligence) is added through unified metadata governance, allowing users to directly ask natural language questions, such as “What is the payable of a supplier”, and the corresponding data can be found through the semantic layer, which may come from structured data or various unstructured data in the document library. As shown below:
“We think this approach is more accurate than RAG because the documentation is not directly vectorized, and there is an extraction process in between. Specifically, we use watsonx.data integration to process structured and unstructured data, and for unstructured data, it extracts the entities and values in the process of vectorization, and then vectorizes the document. In the future, when the large model does knowledge base queries, it will not only return similar vectors, but also return relevant entities and values, and improve the accuracy through the assistance of entities and values. Wu Minda said.
watsonx.data integration is a comprehensive data integration tool, and unlike the data processing tools such as DataStage and Data Replication provided by IBM in the past, watsonx.data integration can support both structured data and unstructured and semi-structured data.
Further down, when the data is put into watsonx.data and integrated with watsonx.data integration, watsonx.data intelligence begins to “work”, its role is to provide unified data governance and data lineage capabilities. For example, Wu Minda said, “There are many ways to access the same batch of data, such as using the large model knowledge base – RAG to ask questions, or using traditional SQL queries – report queries, and using machine learning to model and extract data to train a model, which is also a way.” How to ensure that different access methods and permissions are controlled together? This can be managed through watsonx.data intelligence. ”
At the same time, IBM also encapsulates the managed data into API interfaces or vector databases for real-time calls by agents. For example, supply chain agents can directly access real-time inventory vector data to dynamically adjust procurement plans. This not only improves data availability, but also provides “nutrients” for the continuous evolution of the agent.
Looking at the entire link, if we compare the raw data in the enterprise source system to the raw materials of a manufacturing plant, then the role of watsonx.data integration is to manufacture and process the raw materials for production, put them in the warehouse of watsonx.data, and then manage them through watsonx.data intelligence to form an asset catalog, and finally provide them to the front-end AI and BI for use.
Is the process automated?
So, when business personnel without technical foundation can easily develop applications in a low-code or even no-code way, and the data is ready, can AI Agent be “unimpeded” in the enterprise process? I’m afraid it’s not that simple.
First, each enterprise has an average of thousands of applications, how does the AI Agent connect and connect with these systems and applications?
The dilemma of heterogeneous systems exists at any stage of enterprise development, with different interfaces and standards between systems and systems, and the problem in the past was how to connect different systems and how to break data silos, but now the problem is how to break the “dimensional wall” between AI Agent and these systems, how to call the data in it and perform related tasks.
For example, after receiving feedback on quality issues from customers, if an enterprise wants to hand it over to AI Agent for processing: in the first step, it needs to feed back the problem to the quality management system, and then call the data in different software systems such as manufacturing, design, and process for analysis to locate the root cause of the problem; The second step is to update the enterprise knowledge base after confirming the problem, so as to avoid the recurrence of similar situations. the third step is to notify the relevant person in charge, at which time it needs to call different communication applications such as email, DingTalk or Qiwei; Finally, if it is an issue with parts from an external supplier, external communication is involved, so it may be necessary to send the information generation document in EDI format.
“It’s not easy to string all these processes together, and you have to integrate the AI Agent with your existing systems effectively.” Zhang Cheng, a senior automation technical expert at IBM’s Greater China Technology Division, said that although integration is a relatively “old” concept, its importance and gold content are still increasing. The Hybrid Integration released by IBM during the Think 2025 conference is mainly to provide complete cross-platform integration capabilities on and off the cloud.
Second, how to visualize multiple agent links and how to deal with possible errors during task execution?
For example, if an AI Agent is launched, if there are problems such as delays, network interruptions, memory overflows, and downtime, how can the entire link be monitored in real time, actively detected, and quickly diagnosed, analyzed, and dealt with? This process relies on automated IT operations.
At the O&M level, IBM proposed the concept of AgentOps, which aims to visualize the entire link from the construction and deployment of AI Agents to O&M, optimization and iteration. For example, through Instana’s full-stack monitoring tool, real-time tracking of agent call links, resource consumption, and decision-making accuracy can be tracked. When the response delay of an agent exceeds the threshold, the system can automatically trigger the scaling mechanism.
Third, how to allocate resources reasonably to ensure efficiency and minimize costs?
IBM believes that automation at the infrastructure level is needed to enable AI to use foundational resources with a high degree of automation and elasticity. Observability tools have become critical to this end, including discovering, managing, monitoring, and optimizing agent usage across the enterprise, ensuring efficient and responsible technology adoption.
The watsonx portfolio provides a suite of monitoring tools to monitor AI performance and reliability, enforce AI guardrails, and use AI resources effectively, for example, to evaluate and select AI models based on specific goals such as cost-effectiveness or performance. In addition, IBM announced the acquisition of HashiCorp last year, which is positioned to help enterprises achieve automated, on-demand full lifecycle management from the underlying infrastructure level.
“Relevant data shows that on average, 27% of enterprises’ cloud computing spend is wasted, and this waste can be analyzed by the platform to better realize the deployment of underlying resources, which is something that IBM automation software focuses on solved.” Zhang Cheng emphasized.
Return to the essence of business to look at the value of technology
Writing here, I don’t know if you have noticed that in addition to the new concept of AI Agent, data, automation, etc. are actually clichéd topics for enterprises in the era of informatization and digitalization. In the end, whether it is humans to perform tasks and make decisions, or AI to perform tasks and make decisions, a relatively complete and mature IT infrastructure is essential, which is a “lesson” that enterprises must make up for.
Solve this problem, and then look for the scene.
“Each enterprise has a different stage of development, and the bottlenecks encountered are also different, enterprises must first think clearly about which place is the real pain point, whether you want to reduce costs and increase efficiency or business innovation, the demands of enterprises must first think clearly. It’s not that HR Agent is here today, your company only has 3 HR to use. Zhai Feng said.
So, how can enterprises accurately grasp business needs and make internally built AI agents more targeted, high-value, and continuously optimized? Zhang Xun, manager of the garage innovation team of IBM’s Greater China Technology Division, concluded that enterprises should do it through an advanced path and achieve intelligence step by step through the process of continuous iterative optimization.
“Businesses first need to make sure their investments are manageable, so our team usually deploys our idea and product through a POC (proof of concept) with the customer to select the most typical scenario, and then validate its ROI and deploy it at scale if it meets expectations.” Zhang Xun told InfoQ, “The entire POC process is about 30 days, but the plan is iterated every week, and this process requires full customer participation, timely feedback, and verification that we are on the right path, and if not, we need to constantly adjust.” ”
Taking the manufacturing industry as an example, the IBM garage innovation team has summarized four scenarios that can bring the greatest ROI to enterprises after researching, in-depth discussion, and co-creation: R&D, production, supply chain, and finance. Taking IBM itself as an example, why the three agents of HR, finance, and procurement were released first is actually a scenario with a better ROI that has been verified by IBM itself.
All in all, in IBM’s view, the essence of enterprise-grade AI is not to show off skills, but to reconstruct the business. As IBM Chairman Arvind Krishna said, “The era of AI experimentation is over, and enterprise competitive advantage depends on tailored AI applications and quantifiable business outcomes.” On the Agent track, IBM still plays a combination of “full-stack technology + industry know-how + open ecology”, which does not pursue speed, but emphasizes technical depth and landing accuracy.
For enterprises, it is necessary to realize that no matter how cool the technology is, it cannot solve the problem of business essence, and the more dazzling the technical concepts and products, the more you must be determined to practice “internal skills” and quickly complete IT infrastructure capabilities, which is the basic prerequisite for catching the express train of AI.
