变革推动者：智能代理时代CEO的目标、决策与影响

QuantumBlack, AI by McKinseyThe change agent: Goals, decisions, and implications Companies are feeling agentic AI growing pains. Here’s what CEOs can do tomove past them and position their companies to succeed. This article is a collaborative effort by Alex Singla, Alexander Sukharevsky, Lari Hämäläinen, Oana Cheta, OlliSalo, Pallav Jain, Raghav Raghunathan, Sandra Durth, Stéphane Bout, and Vito Di Leo, representing views fromQuantumBlack, AI by McKinsey; McKinsey Technology; and McKinsey’s People & Organizational Performance and Executives are fond of quotinghockey great Wayne Gretzky, who is credited with saying: “Iskate to where the puck is going to be, not where it has been.” This is sound business advice atone level. But that puck is moving a whole lot faster than it used to as agentic AI rapidly evolves. The call to move faster may seem tone deaf as CEOs and their senior teams struggle to seebottom-line value from early gen AI investments. Developing and scaling gen AI use cases haveproven frustratingly challenging. Some executives remain unconvinced that AI agents will have a As CEOs navigate the uncertainty, it is worth acknowledging both the pace and potential scopeof the change that is happening.AI agents—software systems built with gen AI that have theability to plan, act, remember, and learn to achieve predefined outcomes autonomously—areevolving quickly and, as they mature, could completely change how companies are run and howtheygenerate value(see sidebar “Key trends shaping gen AI and agents”). In fact, this “trough How CEOs manage this changewill determine how well they can capture the benefits. AlthoughAI agents are in their infancy, earlylessons and experienceshighlight four mindsets and actions —Reimagine what’s possible.Much of the thinking around agentic AI today is still focused onautomating basic tasks or augmenting knowledge. The real win, however, will come frommuch bolder aspirations of rearchitected workflows and organizations built around agent- —Act with urgency and start the learning.The rapid rate of improvement of gen AIagents means that a wait-and-see approach is potentially a high-risk move (see sidebar“Breakthroughs in gen AI and agents”). Early practical learnings are invaluable in quickly —Tackle scale and long-term competitiveness issues now.Critical decisions aroundtechnology, trust, governance, what to buy versus what to build, capabilities, and talent areimportant to drive a wider transformation. While you experiment, start forming your strategy —Turn everyone into an agent leader.As agents and agentic systems take over more of theexecutional work, everyone in the organization will need to develop agent leadership andsupervision skills. The executive team especially needs to role model and champion learning While much is still unknown, building a business for the agentic age will require afundamentalrewiringof how the business operates, innovates, and protects sources of value creation. Thisarticle, however, will focus on a few of the most important elements anenterprise CEOshouldaddress related to value, scale, and talent. We will outline what a hypothetical two-year agentic Key trends shaping gen AI and agents AI agents are becomingmore human-like in the kinds of tasks they can do andthe way people interact with them. Thesefeatures democratize AI in a way priortechnologies haven’t and underscoreagents’ potential to affect a broad set of $36.00 in March 2023 to about $3.50in August 2024.10For some models, thecost is less than $1.00.11 system outperformed a single-agent Claude Opus 4 by more than —Breakthroughs in model and systemcapabilities.New reasoning modelsdeploy “test time compute” thinkingduring inference (“system-2 thinking”);standardized tool-calling interfaces,such as Anthropic’s Model ContextProtocol (MCP), let models invokeenterprise APIs safely; vastly largerand more precise short- and long-term memory structures improve —Large growth in spend andinvestments.The compute used totrain state-of-the-art models has beengrowing roughly four to five timesper year.6The top three hyperscalerscollectively plan to invest more than —An acceleration in innovation pace. Only two new-frontier large languagemodels (LLMs) were announced in2020;1by 2025, the number is in thedozens, even hundreds, depending oncounting methodologies.2Similarly,the number of new large-scale AImodels has grown by 167 percentper year since 2020.3The length of —Sharp gains in model training andinferencing efficiency.Breakthroughsin architecture and optimization havedriven training costs down significantlyfor a given capability. The inference Are agents worth it? Claims about the value of AI agents permeate the internet. But since the technology is still sonew, those claims are hard to verify. Early implementations, however, suggest there is significant value at stake. Our experience withmodernizing technology estatesindicates that harnessing AI agents can accelerate tim