What kind of leadership do enterprises need in the era of Agentic AI?

If your enterprise has been struggling with how to enable employees to use chatbots more efficiently in the past year, it's time to adjust the strategic focus. We are experiencing a fundamental leap from "automation" to "Agentic AI."

For enterprise managers, this is an exciting yet challenging moment. As change agents, we have the opportunity to lead our enterprises beyond mere efficiency improvements and into a new era of high autonomy. At the same time, this also requires us to completely rethink the operational logic of the enterprise.

Redefining "Agent"

In the current market, it seems that everything is labeled as an "Agent." However, looking beneath the surface and from the underlying logic, the development of Agents to date presents a clear maturity ladder.

The first level is traditional automation, which processes repetitive and predictable tasks through scripts and fixed workflows. Although it is efficient, it lacks the ability to think.

The second level is AI assistants, which are the currently popular chatbots. They have a powerful knowledge base, can answer queries, summarize documents, and even call data through RAG (Retrieval Augmented Generation). Their core lies in "responding," and they can provide assistance in conversations, but their ability to act independently is still limited.

The third level is goal - and task - based Agents. They no longer rely on instructions but on "intentions." When you give them a specific business goal, they can collaborate with humans and autonomously call tools to complete the task.

The last level is the Agentic system, which is also the highest form of autonomy. Multiple Agents collaborate like a team and can handle highly complex and even ambiguous tasks. They can break down goals into numerous specific steps and divide the work to complete them. Here, the core value of AI evolves from "responding" to "achieving."

Goal - driven: They don't wait for step - by - step instructions but receive and understand "high - level intentions," then autonomously plan the path to turn intentions into actions.

Resource integration ability: They can not only access data but also understand the context, organizational hierarchy, and roles. They know which tools to call and which APIs to connect to complete the task.

Memory ability: A real Agent has memory. It can remember the problems encountered when handling similar tasks last time, thus avoiding repeating the same mistakes.

Learning and adaptation ability: They are not static software but systems that can continuously optimize themselves from feedback loops.

Reporting mechanism: This is often misunderstood. When an Agent encounters a problem it cannot handle and seeks human intervention, it is not a failure of the system but an inherent trust mechanism. This means you can safely let it handle business because it knows when to stop and ask for your help instead of blindly creating risks.

Of course, many people will ask why we are emphasizing "Agents" now when the AI boom has lasted for several years. The answer lies in the intersection of two key curves. Nowadays, the complexity of tasks that AI can handle independently doubles approximately every seven months, which means AI can take on increasingly complex tasks. At the same time, the cost of intelligence is dropping precipitously. Taking the MMLU (Massive Multitask Language Understanding) benchmark test as an example, a score of 83 is equivalent to that of a doctoral - level expert, capable of handling nuances and ambiguities. A few years ago, the cost of obtaining such "doctoral - level" intelligence was extremely high, while now, the cost of processing one million Tokens has dropped from $98 in 2022 to about $0.48.

When the soaring ability meets the plummeting cost, we enter the "sweet spot" for large - scale application of Agents.

Embracing "Non - determinism"

In the new era of Agentic AI, one of the major differences from the past is that non - determinism is no longer a defect of the system but an inherent characteristic. Therefore, different from traditional management, which only focuses on rewarding "determinism," in the era of Agentic AI, value comes precisely from non - determinism, that is, the ability to make immediate adjustments and optimizations based on context changes and ambiguous information.

To harness this power, leaders need to make a mental model shift in four dimensions.

1. Governance model: From "turnstiles" to "board of directors"

Traditional governance is like a toll gate, achieving compliance by defining processes, controlling each step, and conducting checklist verification. However, when facing thousands of high - frequency interacting Agents, this model will completely collapse. Instead, we should manage Agents like a board of directors manages a CEO.

Just as a board of directors doesn't tell the CEO what to do every day but sets strategic intentions and boundaries, you also need to build a "strategy engine" to clearly tell the Agent "this is your goal, and this is the red line you must never cross." Then, manage it through continuous calibration and observation rather than manual approval before each transaction.

2. Risk control: From "factory assembly line" to "trading floor"

The traditional risk management model is similar to a factory, often setting fixed thresholds. For example, a purchase order over $2 million requires vice - president approval, and over $10 million must be submitted to the CFO. This rigid risk control will stifle the core advantages of Agents.

In the new era, enterprises should manage risks like managing a trading floor.

A financial trading floor is full of high risks and uncertainties, but it is effectively managed through real - time visibility and "circuit breakers." Every action of a trader is not intervened, but if their trading portfolio or a specific transaction violates the policy, the circuit breaker will automatically trigger to stop trading. Managing AI Agents should follow the same logic. AI Agents must operate within the risk control scope designed by the enterprise to gain their "liquidity" (i.e., operational authority). This includes not only monitoring the behavior of specific Agents but also monitoring systematic deviations in the Multi - Agent system. Once a deviation occurs, the circuit breaker should be triggered immediately.

3. Organizational structure: From "functional silos" to "immune system"

Organizational structures always adjust with major technological changes. Traditional organizational structures are often vertical, which, although stable, are slow in problem - solving.

When building an organizational structure for AI Agents, the thinking model to be adopted is the "immune system." When white blood cells detect a virus, they don't wait for instructions from the brain, nor do they call on the lungs to discuss. Instead, they quickly assemble and solve the problem through goal - oriented collaborative orchestration. Therefore, organizing cross - functional teams around business workflows is the real way to obtain value results from AI Agents.

4. Cultural gene: From "operational execution" to "continuous learning"

Most corporate cultures reward "precise execution" and punish "deviations." However, the laboratory culture is open to new discoveries, even if the results are not satisfactory. In the laboratory, if a mistake occurs, it is used as a learning mechanism, recorded, and widely shared. In such a culture, new discoveries and adjustments themselves become a characteristic rather than a defect.

Building the "Brain" and "ID Card" of Agents

So, what technical capabilities make AI Agents a reality? The answer lies in three core factors: intelligence, context understanding, and trust.

If we compare an AI Agent to a human, intelligence is its "brain." The technical models, including the thinking models for chain - of - thought reasoning and reflection, endow the "brain" with the ability to turn intentions into tasks. Ideally, we hope that AI Agents have the highest intelligence, the lowest price, and the fastest response speed. However, in reality, there must be a trade - off among the three dimensions. Each model has its unique advantages: some are good at reasoning, some at quick response, some at handling text tasks, and others at image and math tasks.

However, having a brain doesn't mean having the ability to act. Access rights and context understanding are the "hands" of an AI Agent. They endow the AI Agent with the ability to access the right data, take actions, and call various tools to achieve goals. When applying context understanding in an organization, especially for complex workflows, the AI Agent needs to understand roles, hierarchies, data, systems, and tools. This includes using knowledge graphs to enable the AI Agent to understand the relationships between different objects and domains in the data, and using vector databases to enable the AI Agent to understand semantics (i.e., the similarity between things, for example, "cat" and "dog" are synonyms, while "cat" is very different from "cloud computing"). It is also crucial for the AI Agent to have memory ability. It needs to remember role settings (who you are), procedural memory (how to do it), semantic memory (various knowledge), and episodic memory (what happened last time).

The last level is trust. Without trust, AI Agents cannot be applied on a large scale.

One way to ensure trust is to set up security protection mechanisms. For example, a sensitive information filter (i.e., PII, personal identity information). Enterprises don't want such information to be leaked, so relevant rules can be defined in the security protection mechanism. Context consistency checks can provide policy - level information guidance for all AI Agents. For example, clearly state: "Our return policy is 90 days." These rules must exist independently of the model, and the security protection mechanism is always effective no matter which large - scale model is called.

"Automated reasoning checks" are also one of the core tools that enable enterprises to trust AI Agents and ensure their effective expansion in the company. This is a mathematical verification method. Instead of drawing countless triangles like a tester to verify the Pythagorean theorem, use mathematical theorems to prove its correctness. Amazon Web Services has reduced the hallucination rate by 99% using automated reasoning.

It must be emphasized that AI Agents are still in the early stage of development and cannot be regarded as a "magic wand." In fact, if an enterprise has a predictable workflow with fixed steps and highly limited tool usage, basic automation or simple generative AI assistants are sufficient. When an enterprise needs to dynamically select tools or hopes to utilize the adaptability and learning ability of Agents, Agents will show their real strengths. At this time, enterprises can start from three specific areas with the greatest value: software development, customer support, and knowledge - based work.

For example, Thomson Reuters uses Agentic AI for code modernization, increasing the migration speed of legacy.NET code by four times. This is not only an improvement in efficiency but also frees engineers from technical debt and allows them to focus on creating new value. Amazon's shopping assistant, Rufus, can answer highly context - specific questions such as "What size batteries does this toy need?" or "Is this compatible with the accessories I bought before?" Data shows that the purchase conversion rate of users using Rufus has increased by 60%. The accounts payable case is a typical scenario of knowledge - based work and exception handling. The traditional goal of accounts payable (AP) processing is "timely payment," while the goal of Agent AP can be "optimizing cash flow." It can autonomously choose to pay in advance according to exchange rate fluctuations or delay payment using foreign exchange hedging strategies based on exchange rate fluctuations, vendor historical credit, and contract terms. This is the value leap from "step - by - step" to "high - initiative."

Ishit Vachhrajani | By

Ishit Vachhrajani is the Global Head of Technology, AI Analytics, and Enterprise Strategy at Amazon Web Services.

This article is from the WeChat official account "Harvard Business Review" (ID: hbrchinese). Author: HBR - China. Republished by 36Kr with permission.