The judgment of top talents is the greatest treasure in the AI era.
The focus of competition in the first phase of AI implementation is who can access top-tier large language models. This phase has basically come to an end, and the threshold for using large language models is becoming increasingly accessible to all. The core of competition in the next phase lies in who has completed the more difficult task: encoding and solidifying the real thinking patterns and working methods of enterprises.
Some enterprises have truly reshaped their work patterns with the help of agents. Others are trapped in small-scale, low-risk pilots and cannot achieve stable, large-scale implementation of agents.
The gap between the two types of enterprises is not at the technical level as most people imagine. Today, the vast majority of institutions can access the same large language models, tools, and roughly equivalent infrastructure. The real watershed lies in how to address a topic that most managers have never faced before: making subjective judgment explicit. This has become a new bottleneck in AI application implementation, catching many enterprises off guard.
Over the past decades, enterprises did not need to sort out the decision-making logic of top talents. Professional experience has been passed down from generation to generation through mentorship, observation, imitation and personal practice. New employees gradually internalize the enterprise's way of thinking by observing and listening. This model works well when humans perform execution tasks.
However, AI agents have broken the original balance. Different from traditional software, agents can operate in environments with ambiguous information and make decisions in real time. But unlike humans, they cannot understand behavioral norms through observation, nor can they deduce background information from organizational culture. Agents can only work according to explicitly written instructions, and nothing more.
As a result, a typical failure scenario has emerged widely across all industries: an enterprise launches a customer-facing AI agent without first sorting out the decision-making logic of top customer service staff — for example, how to handle special price cases, appease dissatisfied long-term customers, or respond to demands that just exceed the policy scope. Eventually, the agent gradually deviates from the direction and gets out of line with the enterprise's goals, because no one has put the real judgment criteria for such decisions and the implicit background conditions required to make correct choices into written documents.
What does it really mean to make judgment explicit and systematic? It means converting the enterprise's implicit decision-making principles into standardized guidelines that agents can execute. Such principles cover risk tolerance, brand tone, escalation trigger thresholds, quality standards, and the subtle logic behind exception handling. For a long time, these experiences have only existed in the minds of senior employees. To make AI generate value, these experiences must be transferred elsewhere.
The Core Gap Between Leading and Lagging Enterprises
Enterprises at the forefront are building what we call judgment infrastructure, relying on this system to realize the large-scale replication of professional capabilities.
To build this infrastructure, managers need to complete three structural transformations:
1. Collaborative Governance by Business Departments, Human Resources Departments and Information Technology Departments
Defining risk tolerance boundaries, setting agent performance standards, and managing the launch and phasing out of agents are not only technical issues, but also organizational issues. The key point is: this work cannot be outsourced.
Business leaders, human resources teams and information technology departments must establish a collaboration mechanism to jointly manage digital workforces. Enterprises should change their mindset: do not simply regard agents as software licenses, but as operation participants, to continuously guide and optimize their behavior patterns.
ITA Group, a global event planning, incentive and employee recognition service provider, experienced this rule firsthand when trying to build an air ticket booking AI agent for its event business. The difficulty of the project was not developing the agent itself, but clarifying the information that the agent needs to master to make credible decisions: when to prioritize cost control, when passenger experience is more important, which special cases can be approved, and what scenarios require human intervention.
This practice exposed a common pain point: a large amount of decision-making experience needs to be relayed by business experts to technical personnel. ITA Group then adjusted its operation model, providing tools to developers, managers and knowledge workers, so that they can independently adjust the agents that represent them to carry out work. Chief Operating Officer Maura McCarthy worked closely with Chief Information Officer Jason Kacher, and with the support of the CEO and CFO, promoted the management to reach a consensus to ensure the implementation of the transformation.
McCarthy said: "The most valuable experience we have gained is the importance of combining professional experience with AI. The core idea is that business experts first adjust the behavior logic of the agent before large-scale promotion; and continuously iterate and optimize the agent's behavior to adapt to changes in business, work content and evaluation criteria."
2. Managers Transform into "Judgment Architects"
This is the most far-reaching transformation, and also a link that most enterprises seriously underestimate.
Take Debbie Riazi as an example. She is the Director of Compliance and Labor Relations at AWP Safety, the largest on-site safety company in North America. The enterprise covers 33 states and has nearly 9,100 employees, and Riazi's department has only herself. She has built a series of AI agents, each of which solidifies a part of her professional experience. One of them is responsible for handling employee job matching applications: calling up the corresponding job descriptions, displaying similar historical processing cases, and completing information collection in accordance with the standardized process she has polished for years. Another agent handles all kinds of information inquiries received by the enterprise: sorting out the content of demands, assigning responsible persons for docking, and drafting reply documents. These agents save her hundreds of hours of work every year, but the more important value lies in the ripple effect brought by the freed-up time. "I can intuitively prove that all matters are handled with unified standards," she said, "which naturally reduces the legal risks of the enterprise."
Nathan Mapp, Financial Director of a global venture capital and applied technology enterprise, went a step further. With more than ten years of experience in the financial field, he has precipitated a complete set of professional methodologies, organized them into Markdown documents, and agents built on Claude and Claude Code can call them for reference in real time. Now a team of two can complete the work that originally required ten people. When the agent processes each task, it will implement Mapp's judgment criteria, just like a senior accountant who controls all details throughout the process, covering the links that junior employees tend to overlook in the past.
What Riazi and Mapp are doing is to systematize the judgment criteria in their work. The core work of managers has been transformed into implementing professional capabilities, and empowering both human employees and digital agents. This is a brand-new skill system, and the vast majority of enterprises have not yet established corresponding training and incentive mechanisms.
3. "Thinking Executors" Become the Core Talents
The boundary between traditional strategic decision-makers and front-line executors is disappearing. Top employees have dual characteristics: they can carry out strategic thinking, and can use agents to implement their ideas. They design work processes, solidify judgment standards, build and continuously iterate systems, and keep moving towards higher-value work.
Ramp, a leading financial platform, serves 30,000 enterprises, and the company's strategic focus is to cultivate such talents. All employees can use enterprise versions of ChatGPT, Notion and Perplexity; during the onboarding training phase, the company guides employees to independently build AI tools, rather than simply operate systems developed by others. Employees can standardize their own professional experience and implement it through agents. We define such talents as thinking executors: they do not only regard AI as a tool, but can take the lead in deciding how AI represents themselves to carry out work. Enterprises that cultivate such talents on a large scale will learn and adapt much faster than their peers who only focus on cultivating a single ability.
Implementation Entry Point
Most enterprises adopt a method that has misunderstandings: directly ask senior employees to write down their own experience. This method yields little result. Experts often find it difficult to describe tacit knowledge abstractly, and when asked to sort out documents directly, the content they can write out is far less than the experience they have in their minds.
A more effective way is: do not directly ask experts to sort out judgment criteria, but create scenarios for experience to emerge naturally.
Gather several senior practitioners in the same position, arrange a professional host, and hold discussions around various real scenarios and edge cases encountered by the enterprise. If the team can reach a consensus quickly, it means that this part of the content can form clear rules; if there are divergent views, the judgment logic contained in it is the core content worthy of extraction. The records of the whole discussion are the first draft of the systematic judgment criteria.
The claims team of an insurance company only held a two-hour seminar, sorting out rich details such as risk tolerance standards, customer empathy standards, and reporting mechanisms, with far more information than the written rules accumulated over the years. The reason is that debate of views can externalize the internal thinking logic, which simple document records cannot achieve. This discussion record will also become a crucial background document for the subsequent implementation of all agents.
New Competitive Barrier
Once the judgment criteria are successfully systematized, enterprises will gain significant strategic advantages: professional experience can be migrated. Best practices are no longer limited to a small number of senior personnel. Organizational knowledge can be deployed on a large scale across departments, regions and product lines. Enterprises that master the tacit knowledge coding method will form structural advantages: improved decision-making efficiency, stable execution quality, enhanced innovation capability, and a positive cycle of continuous learning and optimization.
The development trajectory of ITA Group intuitively shows this compound interest effect. The first six or seven months of the project progressed slowly, and the enterprise needed to figure out how to convert the experience in experts' minds into background documents available for agents. But after this operation model was formed, the advancement speed increased significantly, especially in the field of software development. After preliminary guidance, developers no longer only use agents to generate code, but quickly turn ideas into usable prototypes. A large number of tasks are carried out independently, and the project cycle is shortened from several months to several weeks; the idea of reconstructing work processes with the help of agents has also been gradually extended to other functional departments.
The rule behind this is: systematic judgment criteria have compound interest effect. The first batch of pilots progressed slowly, because the enterprise was learning how to make tacit experience explicit; the subsequent projects were implemented at a greatly accelerated speed, as the enterprise had established supporting governance mechanisms, trust and standardized work rhythms, and could replicate the mature model.
This is the natural evolution of the hybrid workforce strategy we proposed in the previous article discussing how autonomous agents reshape the workplace. Sorting out work tasks and building human-agent collaboration teams is a necessary first step, but far from enough. Enterprises that build human-agent collaboration teams will soon realize that the upper limit of team capabilities depends entirely on the quality of the guidelines given to agents. Deploying AI agents is only the basic entry ticket; the judgment infrastructure is the moat and strategic differentiation advantage of leading enterprises.
The focus of competition in the first phase of AI implementation is who can access top-tier large language models. This phase has basically come to an end, and the threshold for using large language models is becoming increasingly accessible to all. The core of competition in the next phase lies in who has completed the more difficult task: encoding and solidifying the real thinking patterns and working methods of enterprises.
The vast majority of enterprises have not yet started this work. Pioneers will shape the future work patterns of their industries; lagging enterprises will fall into passivity, just like companies that delayed building talent strategies in the early years, bearing the ever-expanding structural disadvantages.
Keywords: #AI
Jen Stave, Ryan Kurt, John Winsor | Text
Jen Stave is the first dean of the AI Institute at Harvard Business School. Ryan Kurt is the founder and CEO of the AI Lab, a strategic consulting agency, helping CEOs drive AI transformation. John Winsor is a senior researcher at the AI Institute at Harvard Business School, whose research focuses on AI, organizational design and the future of work, providing transformation consulting for enterprise executives.
Zhou Qiang | Editor
This article is from the WeChat Official Account "Harvard Business Review" (ID: hbrchinese), author: HBR-China, published with authorization from 36Kr.