The Oliver Wyman Forum publishes the second edition of "The AI Agenda," titled "The reality check." The opening finding: AI is everywhere in the company except the P&L. Budgets are rising, pilots are multiplying, but proven gains remain delayed, diffuse, or dependent on changes external to the technology. The report draws on more than 25,000 responses across 15 markets, a survey of 415 CEOs conducted with the NYSE, and a sovereignty index.
On return on investment, the authors distinguish four investment types with different trajectories: targeted applications, employee tooling, foundations, transformation. The calculation often omits costs, such as the 22 minutes of verification per hour of AI-assisted work, and conflates task-level gains with returns. A return thesis must be formulated before committing capital, the process must be redesigned when AI shifts the bottleneck, and a "ROI harvester" must be named to convert freed-up capacity into results. Only 12% of CEOs report a gain above 10%.
On the workforce, the report describes a leadership misalignment: 37% of senior executives and 11% of non-managers see a clear AI ambition. Lacking a narrative, employees write a darker one of their own: 62% of daily users hide their AI proficiency, 68% would trade salary for a job-security guarantee. Management is unbundling, junior learning is eroding, and 27% of users report being twice as productive, especially where the employer provides framework, tools, and training.
On commerce, AI is becoming the second most common starting point for purchases and the most followed source of advice. Brands must practice "shopping agent optimization": agent-readable product sheets, independent evidence, and checkout flows that accept authorized agents. The seller must verify the agent, the customer, and the agent's mandate.
On governance, a simple approved path reduces shadow AI; a risk lab allows new uses to be tested; human in the loop can no longer be the supervision model, replaced by layered automatic controls and a living governance that re-examines approval with every change.
On sovereignty, no country or company controls the entire stack. The goal is controlled interdependence: owning the control layer (data, context, memory, permissions), keeping the model replaceable, paying the sovereignty premium only where the stakes justify it, and favoring rules and partnerships over public spending.