What the Best Business AI Users Are Doing Different

The clearest distinction between organizations reporting established AI returns and those still experimenting is not simply whether they use AI, but whether they have built the systems to deploy and govern it at scale. KPMG’s Q3 AI Pulse Survey of more than 2,100 senior leaders across 20 countries found that established-ROI organizations were more likely to report significant employee adoption of AI agents (48%, versus 15% among experimenters), AI-assisted cyber defense (58% versus 8%), and a formal AI harness layer (86% versus 31%). KPMG describes the harness as a management layer that helps AI operate within organizational constraints and context; it sees that layer as a key difference between experimentation and returns.

The same pattern appears in organizations’ approach to data and accountability. Enterprise-wide data sovereignty strategies were reported by 53% of organizations with established ROI, compared with 8% of experimenters, and more mature organizations were also more likely to assess value consistently against cost. Across respondents, 53% said accountability for AI-informed decisions rested at the C-suite level, while 23% said their organizations had built model-routing capability. These findings point to governance and operating discipline as part of AI adoption, rather than issues to address only after deployment.

Agent strategies are broadening, although organizations are at different stages. A quarter were developing multi-agent systems and a fifth were orchestrating multiple agents. From Q1 to Q3, the share naming productivity as an AI goal fell from 42% to 37%, while human-AI collaboration and each of responsible AI, governance, trust, and security gained four points in priority. Revenue-focused agent strategies increased as efficiency-focused strategies declined, but most organizations pursued both. KPMG says sovereignty, monitoring, and related management concerns have moved from largely off the radar to mission-critical decisions.

AI spending management is also becoming more comprehensive than simple cost cutting. KPMG calls the shift “AI economic management”: organizations with established ROI were more likely to have AI cost-monitoring dashboards (77%, versus 43% of experimenters) and to review costs at the approval stage (73% versus 50%). Yet tighter controls can involve a trade-off. The host cautioned that restrictive usage or token budgets may constrain experimentation; such budgets were reported by 46% of established-ROI organizations and 31% of experimenters. The figures suggest that cost visibility is common among more mature users, while limits on usage are not automatically evidence of stronger outcomes.

Ramp’s AI index showed token spending down 5.2% from the prior week even as token volume rose and companies’ AI-use costs fell. Eric Karazian attributed the cost decline almost entirely to competition between OpenAI and Anthropic, and said open-source models account for less than 5% of business AI spending. The host noted a possible consequence for infrastructure investment: if usage keeps rising while costs fall, investors may be less willing to finance further build-out. That outcome was presented as conditional, not as an established result.

Planned AI investment is rising despite the reported decline in per-use costs: average investment for the next 12 months increased from $186 million in Q1 to $210 million in Q3. The host expects organizations to spend substantially more on intelligence over time. Taken together, the survey’s strongest signal is that returns are associated with adoption supported by governance, monitoring, cost assessment, and organizational context—not merely with running more experiments or deploying agents.