,

McKinsey’s The State of AI in 2025: from “everyone is using it” to the few who actually know how

—— A rereading of a survey of 1,993 companies across 105 countries

INTRO|Three years on, we can finally see AI’s real progress bar

Since the first wave of generative AI tools appeared in 2022, the business world’s imagination of AI has moved from “interesting” to “necessary.” McKinsey’s latest report, The State of AI in 2025: Agents, innovation, and transformation, is based on a global survey of 1,993 respondents in 105 countries, 38% of them from companies with annual revenue of more than $1 billion. It is, for now, one of the most representative pictures we have of enterprise AI as it actually stands.

The report sends two signals that look as if they contradict each other. On one side, 88% of the companies surveyed already “use AI regularly” in at least one business function, a sharp rise from 78% a year earlier. On the other, nearly two-thirds of companies say they are still in an “experiment or pilot” stage; only about one-third have truly entered cross-functional scaling. The heat around AI agents is similar: 62% of companies are at least experimenting with an agent architecture, and 23% have begun to scale in a given function, but in any single business function, the share that is actually “expanding the deployment of agents” does not exceed 10%.

One sentence from McKinsey names the odd state of 2025: “AI tools are already everywhere, but organizations that have truly embedded AI deeply into a process, and produced an enterprise-level financial effect, are still a minority.” This piece follows the main line of McKinsey’s survey, and cuts in through six questions: Where has the spread of AI got stuck? Where has the agent wave actually landed so far? Why have only about 6% of “high-performing AI leaders” truly produced EBIT? What did those companies get right in organizational design, talent, and risk governance? In the end, I come back to one core conclusion: the AI contest from here is no longer “whose model is stronger.” It is “whose organization is more plastic.”

I. AI is already “in use across the company,” but most are still stuck in a pilot loop

McKinsey’s survey shows that the range of AI use inside companies continues to widen: more than two-thirds of respondents say the organization uses AI in “more than one business function,” and more than half already use it as a matter of course in “three or more functions.” AI is no longer a toy for an innovation group. It has entered the everyday work of IT, marketing, service operations, and knowledge management. If you only ask “whether AI is being used,” 2025 can already be called a year of AI’s spread.

The problem is that this spread still sits more at the “function layer” than at the “process layer.” Among companies already using AI, about two-thirds are still in an “experiment” or “pilot” stage; only about one-third have truly entered scaling (that is, expanding deployment across functions and across processes). The gap is especially clear among large companies: among firms with revenue of more than $5 billion, nearly half have entered a scaling stage; among those with revenue below $100 million, only 29% have crossed that threshold. This tells us that, inside most organizations, AI is still a string of “handsome use cases,” not an end-to-end process that has been redesigned.

This state can be summed up in one phrase: the pilot loop — a company keeps running proofs of concept, PoCs, and pilot projects in different functions; each one looks as if it “has some effect,” but a repeatable, expandable capability of productization and platformization never forms. Behind McKinsey’s data sit three structural bottlenecks. First, data and technical debt — without a unified data foundation and a modern architecture, AI can only ever stop at the demo. Second, processes have not been redesigned — the work is only “adding an AI tool to an old process,” not redesigning the workflow for AI. Third, there is no clear scaling strategy or order of priority — every function is running its own AI project, and no one is responsible for “lifting them to the enterprise level.”

In other words, “being able to run an AI project” is no longer unusual. What is scarce is the ability to “turn an AI project into enterprise infrastructure.”

II. The state of agentic AI: from “can answer” to “can act,” but still a single-point breakthrough

One of the most watched parts of the report is the actual adoption of AI agents. McKinsey defines agents as “systems based on a foundation model, able to act in the real world, and able to plan and execute a multi-step workflow on their own.” The survey shows that 23% of companies have already entered an “agent scaling stage” in at least one business function, and another 39% are experimenting. It looks as if agents have taken a first step into enterprise deployment.

Pull the view down to the granularity of a function, though, and the picture changes: in any one business function (IT, knowledge management, software engineering, customer service, marketing, and the rest), the share that has truly reached “scaling or fully scaled” does not exceed 10%. Even in IT and knowledge management, where agent use is most active, companies that can say they are already scaling are still a minority. McKinsey’s team, in its commentary, points directly at this gap: the narrative around agents is already very hot, but doing the work correctly, and doing it deeply, is in fact an extremely laborious piece of engineering.

From an architectural vantage, the key shift with agents is this: AI is moving from “an assistant that can answer a question” to “a workflow executor that can take a task, know how to break it down, and call a tool.” That shift asks three new things of a company. First, process standardization — without clear process nodes and a definition of inputs and outputs, an agent will only “busy itself blindly” in the confusion. Second, modern system interfaces — a legacy system that lacks APIs, permissions, and an audit design will freeze the agent in the middle layer. Third, risk and boundary governance — an agent that can “execute automatically,” if it lacks observability and a stop-loss mechanism, may simply run faster in the wrong direction.

So what we see for now is companies first testing the water in relatively closed, risk-controllable, clearly processed settings: an IT service desk, internal knowledge retrieval, assistance in software engineering, parts of a product R&D process, and the like. These domains already have enough data and toolchains, and a safety boundary is relatively easy to set, so they make a suitable “first battlefield” for agent adoption. The next step, if a company truly wants to enter a stage of “several agents linked, driving an end-to-end process,” is to design agents as a new generation of “dynamic intermediary layer,” and not merely as a higher-spec chatbot.

III. Only about 6% of “AI high performers” have truly turned AI into EBIT

McKinsey pulls out one group in particular in the report: AI high performers — companies for which two conditions hold at the same time: (1) more than 5% of EBIT can be attributed to AI; and (2) an internal assessment that they have “obtained significant value from AI.” Such companies are about 6% of the sample. They are, for now, the small group that has truly “turned AI into a business.”

The most visible trait of this group is not technical. It is “ambition.” High performers are 3.6 times as likely as other companies to say they intend to use AI to bring “transformative change at the enterprise level,” and not only incremental optimization. Besides almost all of them treating efficiency as an AI goal, they also significantly more often list “growth” and “innovation” as equally important goals — simply put, they are not only using AI to “save money”; they are also chasing “earning more” and “doing something new.”

More decisive still, high performers go very hard at “the workflow.” 55% of high performers say that, in the course of introducing AI, they “fundamentally redesigned the workflow,” 2.8 times the share among other companies. The difference here is almost the watershed between “treating AI as a plug-in” and “reconstructing the system for AI”: the former is having employees use one more tool inside the original process; the latter is asking, from the beginning — if we assume we have an intelligent system whose capability boundary is completely different, would we still use the same process?

The attitude of leadership among high performers is also entirely different. 48% of high-performer respondents “strongly agree” that senior management has shown clear ownership of, and a long-term commitment to, AI programs, including personally role-modeling the use of AI and continually putting AI first in the budget; among other companies, only 16% have the same sense. McKinsey further uses the Rewired framework to break down the shared behavior of these companies: they are more likely to have a clear AI vision and roadmap, a more mature technology and data infrastructure, a more complete human-in-the-loop mechanism, and more agile product delivery and KPI tracking.

To put it in one sentence: high performers win on the organization, not on the model. The people who truly produce EBIT from AI are not the ones who “ran the best demo.” They are the ones willing to rewrite the process, rewrite the organization, and rewrite leadership behavior for AI.

IV. AI’s real ROI: gains at the function level, enterprise EBIT still scarce

If you only look at “financial indicators,” AI’s results seem less dazzling than imagined. Only 39% of respondents say that, in the past year, AI had any degree of effect on the company’s overall EBIT, and among those people, most believe that “the share of EBIT attributable to AI is below 5%.” That is why many senior leaders ask: we have put so much money into AI, why is there no very obvious linear return on the financial statements?

Pull the lens in to the “business-function layer,” though, and the picture is much richer. McKinsey’s data show that, in software engineering, manufacturing, and IT, respondents most often see a 10–20% cost reduction, or even more than 20%; and in marketing and sales, strategy and corporate finance, and product and service development, more people mention revenue growth of more than 10% from the use of AI. That is, function-level ROI is already very real, while enterprise-level EBIT is still “being integrated.”

McKinsey also asked a more interesting set of questions: besides the money, what “qualitative effects” do you think AI has had on the company as a whole? 64% of respondents believe AI has raised the organization’s capacity to innovate, 45% see a positive effect on employee and customer satisfaction, and 36% feel that competitive differentiation has been strengthened. This in fact reveals another fact — a good deal of AI’s value still exists, for now, in the form of “an improved speed of innovation and an improved experience,” and is not immediately converted into EBIT.

From FIL’s vantage, two points can be added here. First, most companies’ AI value is still a stack of “point use cases,” not an accumulation of “platform capability,” so the benefit is visible on a department’s books and is diluted on the enterprise P&L. Second, many of the benefits AI brings are “multiplicative”; they interweave with other changes (product strategy, organizational redesign, the market environment) and are hard to attribute simply. The real question is whether you have a clear path that can turn these scattered functional results into a sustainable, enterprise-level competitive advantage.

V. The talent map: from “can write code” to “can design human–AI collaboration”

On AI’s effect on employment, McKinsey’s data present a “split expectation.” In the next year, 32% of respondents expect AI to reduce the company’s total headcount by more than 3%, 13% expect an increase of more than 3%, and most believe that “overall headcount will not change much.” Pull the view to the function layer, and McKinsey finds that, over the past year, the median share of functions that actually saw a reduction of more than 3% was about 17%, but the median expecting “a headcount reduction next year because of AI” has already reached 30%. That is, people’s expectations of adjustment in the future are clearly more severe than the actual change over the past year.

Interestingly, even so, most companies are also hiring a large number of AI-related people, especially large companies. Firms with annual revenue of more than $1 billion hire data engineers, software engineers, ML engineers, data architects, and AI product owners / managers more often than smaller companies do. At the same time, roles such as prompt engineer, AI ethics specialist, AI compliance specialist, and “translator” — that is, someone who can turn a business problem into a data / AI problem — have already appeared on hiring lists, though they are still early.

The structure behind this is clear. On one side, entry-level “repetitive technical work” will be eroded quickly by LLMs: basic SQL writing, simple ETL, generation of boilerplate code, and the like. On the other, people who can design a structure of human–AI collaboration are becoming more and more decisive. In its commentary, McKinsey uses “hybrid intelligence” to describe this new division of labor — what truly creates value is no longer simply “more AI,” but “how a person with domain knowledge uses AI correctly, and steps in to judge at the right node.”

For a company, this means the talent strategy has to be rebuilt in at least three directions. First, from “stacking technical people” to a mixed complement of “technology + business + AI translator.” Second, designing an AI upskilling path for different roles, instead of leaving “learn AI” for employees to figure out on their own. Third, in workforce planning, facing squarely AI’s effect on the structure of roles — which work will be rewritten, and which will become a more leveraged core role under AI. High performers are already clearly ahead on these three points.

VI. Risk and governance: the real moat is “the ability to scale safely”

As AI soaks into more critical business, risk is no longer an abstract discussion; it is beginning to become a concrete incident. In this survey, 51% of respondents say the organization experienced at least one negative consequence from AI in the past year. The most common is “inaccuracy” (about 30% of the companies surveyed had encountered it), followed by risks such as compliance, reputation, personal privacy, and unauthorized action.

The good news is that companies’ action on risk governance is also accelerating. In the 2022 survey, the average company actively managed only about two kinds of AI risk; by 2025, that number had risen to four. The risks most often brought onto the governance list are inaccuracy, information security, personal privacy, and compliance. Another finding, though, is worth sitting with: although “explainability” has already become the second most frequently mentioned actual risk, the share of companies that have truly taken a governance measure against it still lags other risk items.

More interesting still, high performers in fact “more often” encounter negative consequences from AI, especially risks related to intellectual property and regulation — the reason is simple: they use AI in more critical, more complex settings, and naturally hit more mines. At the same time, they are also more active than other companies in laying defenses across several risk items, including a more rigorous human-in-the-loop design, a clearer process for verifying outputs, and a more centralized AI governance architecture.

McKinsey organizes these practices into a “management-practice list,” and uses the six dimensions of Rewired (strategy, talent, operating model, technology, data, and adoption & scaling) to see which behaviors most distinguish high performers. Particularly high in the ranking are: clearly defining when a person needs to verify a model’s output; embedding AI in an actual business process rather than stopping at the prototype stage; building a centralized governance team; and having senior leaders take an active part and demonstrate the use of AI. The shared ground of these practices is this: the real moat is not “whether you can make a stronger AI,” but “whether you can let AI operate inside a frame that has a boundary, an audit, and a mechanism for correction.”

From FIL’s vantage, this in fact points to an important trend: when we talk about AI competitive advantage in future, a large part of it will be “a contest of governance capability.” Whether you can do two things at once — try boldly, and keep error inside a boundary — will decide how fast, and how far, you can go.

Closing|The next fight in AI: from “using AI” to “becoming an AI-native organization”

This McKinsey survey, based on nearly 2,000 companies, throws a basin of cold water on the AI heat of 2025, and at the same time points to a clearer path forward. On one side, we can be sure that AI has already crossed the discussion of “whether to use it” — 88% of companies use AI as a matter of course in at least one function, and agents have begun to test the water in domains such as IT and knowledge management. On the other, we also have to admit that most organizations are still in a pilot loop, more handsome than useful, and the high performers who have truly turned AI into EBIT are still a minority.

These high performers have shown, by what they do, one thing: the core of the next stage of competition is not the technology, but organizational plasticity. Are you willing to treat AI as an engine for reconstructing the business, and not as a plug-in added to an existing process? Are you willing to treat agents as a new generation of “digital employees,” and to design a work contract, a boundary of responsibility, and a performance metric for them? Are you willing to rewrite the talent structure and the governance mechanism for AI, and not push all the risk onto some data-science group?

From FIL’s angle, we would compress this transformation into three questions, and leave them with every leader and decision-maker who is thinking about AI:

  • If you do not rewrite the process, how much EBIT can your AI project bring, at most?
  • If you do not rewrite the organization, how far outside the sandbox can your agents actually go?
  • If you do not rewrite the talent structure, who will you rely on to design the human–AI collaboration of the future?

A lead in technology is no longer a sufficient condition. The real moat is whether, in a short enough time, you can let the organization complete an “AI-native” rewiring — turning AI from a tool into a structure.

Link to the original

https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai#

The author

David Yang (楊閎), founder of Future Insights Lab

David Yang (楊閎) is the founder of Future Insights Lab. He writes AI Future Reflections (《AI 未來反思錄》), with a focus on Agentic Commerce, AIGC, data methods, and organizational renewal.

Learn more → About the author

More deep-dive pieces

The judgment, restated

My judgment is that AI has already moved from “everyone is using it” to a threshold — only the few who have trained agents, innovation, and transformation into a system actually know how to use it.

Three things this week

  • Take stock: where are you stopped — experiment, pilot, or scale?
  • Pick one high-value process, and define the agent’s boundary and the point at which a person must sign.
  • Set a group of transformation metrics (value / risk / depth of adoption), and play them back once every two weeks.

Further reading: Google’s brand factory | the multi-agent white paper | a close read of State of AI 2025. Subscribe to Future Insights Lab, or reply and tell us where you are stuck.

訂閱 Future Insights Lab

不定期寄送 AI × Business 短篇筆記,協助你掌握實務重點。隨時可取消訂閱。