From 2024 into 2025, the surveys rhyme. More organizations use AI. A thin slice captures outsized value. Enterprise EBIT impact remains uncommon. The through-line is not a new model name—it is whether teams redesign workflows, track KPIs, and start with one measurable process instead of a catalog of pilots.
Adoption rose; scaled value did not keep pace
McKinsey’s November 2025 State of AI survey put regular use in at least one function at 88% (up from 78%), with only about one-third beginning to scale and 39% reporting any enterprise EBIT impact. High performers—about 6%—combine ≥5% EBIT impact with significant value and are more likely to redesign workflows.
BCG’s 2024 “Where’s the Value in AI?” found only a small share creating substantial value; the 2025 Widening AI Value Gap reports about 5% achieving AI value at scale and 60% seeing little material impact despite investment.
Investment and capability kept climbing
Stanford’s AI Index 2025 documents record corporate AI investment—on the order of $250B+ in 2024—alongside falling inference costs and faster model progress. Spending and model quality alone do not explain the ROI gap inside companies.
The constraint is organizational: which processes change, who owns outcomes, and whether success is measured beyond demo metrics.
Workflow redesign and KPIs separate leaders
Across McKinsey and BCG, leaders push beyond isolated pilots. They redesign work, scale fewer initiatives further, and invest in the operating model—not only in licenses. Tracking well-defined KPIs for AI solutions shows up repeatedly as a practice tied to impact.
Agent hype runs ahead of deployment: McKinsey still finds agents scaled in under 10% of respondents’ organizations within any given function.
A practical synthesis
Start with one measurable workflow. Write acceptance criteria. Include review, logging, and handover. Price total cost—including retries and delay. Expand only when the KPI moves.
That is the same advice as the field notes in this series, now backed by two years of converging survey evidence.