The McKinsey 2025 survey shows a paradox: adoption has become routine, but enterprise “scaling” and economic impact remain the exception. Technology doesn’t make the difference: the organization does.

Everyone is talking about it, almost everyone is using it-but very few are turning it into a measurable management advantage. This is the less obvious message that emerges from McKinsey’s The State of AI 2025: AI has entered everyday work (88 percent of organizations report regular use in at least one function), yet the majority remain stuck in the middle ground between enthusiasm and transformation (nearly two-thirds have yet to initiate enterprise-wide scaling). In other words: AI is becoming a departmental habit rather than an enterprise capability.

“Island” adoption: many use cases, little architecture

McKinsey’s survey, updated to November 2025, suggests diffusion by contagion, not by strategy: more than two-thirds use AI in more than one function and about half in three or more functions. Translated: use cases accumulate, but there is often a lack of a common “operating system,” rules, roles, processes, metrics. This is where the gap between experimentation and transformation plays out: improving individual tasks is easy; redesigning end-to-end is what really moves the needle.

AI agents: high curiosity, low scale (and no accident)

2025 marks the entry of AI agents into the lexicon of decision makers: 62 percent say they are at least experimenting. But when looking at scale, the music changes: only 23 percent say they are scaling agent systems somewhere in the enterprise, and within individual functions, the share of those scaling is no more than 10 percent.

Economic value: lots of innovation, little EBIT ((Earnings Before Interest and Taxes)

According to the report’s data, 64 percent of companies see AI as an innovation lever, but only 39 percent attribute it an impact on enterprise-level EBIT (often stated to be under 5 percent).AI generates micro-improvements easily; but moving an economic indicator requires other choices (priorities, budgets, governance, and especially rethought processes).

The real benchmark: not who “uses,” but who “redesigns”

McKinsey identifies a small group of AI high performers (6 percent): they stand out because they are about 3 times more likely to redesign workflows and because they make net investment choices-more than a third allocate more than 20 percent of their digital budget to AI. The real advantage, however, comes not from prompts but from organizational decisions.

Risks: not rare accidents, but management still immature

Finally, 51 percent of AI users report at least one negative consequence; inaccuracies (nearly one-third) stand out among the most cited. On average, companies go from managing 2 risks (2022) to 4 today: governance and risk management become competitive accelerators, not just compliance.

If AI is now widespread but scaling is rare, then the key competency is not “knowing what an LLM is,” but knowing how to design work with AI in it:

• Choose a few high-impact use cases and measure them;
• Define responsibilities and controls (who is accountable for what);
• Redesign end-to-end workflows (not isolated tasks);
• Set governance and risk management as accelerators, not brakes.

The real question for management, then, is not “what tool do we adopt?” but: what part of our work are we willing to really redesign, with accountability, controls, and metrics, to turn AI from an initiative into a capability?