A new KPMG survey of 204 U.S. business leaders — all from companies with over $1 billion in annual revenue — has landed a finding that should make every executive deploying AI agents sit up: only 26% of organisations have full, real-time visibility into what their AI systems actually cost to run.

That number is striking when you set it against the deployment reality. The same survey found that the share of organisations orchestrating multiple AI agents across workflows doubled in the past quarter, rising from 9% to 18%. More agents, more complexity, more spend — and most companies are flying with the instruments off.

The Q2 2026 KPMG Global AI Pulse survey, which captured perspectives from April 28 through May 25, is the clearest signal yet that enterprise AI has outpaced enterprise AI governance.

The Dashboard Problem

The survey data reveals a governance gap that is easy to miss at first glance. Two-thirds of organisations (66%) say they have AI monitoring dashboards in place. Nearly as many (61%) have approval processes for new AI deployments. On the surface, that sounds like a managed environment.

But dashboards and real-time cost visibility are not the same thing. Only 36% of organisations have implemented direct token or usage controls — the mechanisms that actually let you intervene before a bill arrives. The rest are watching dials without the ability to turn them.

This matters because agentic AI has a fundamentally different cost profile from the chatbot tools most enterprises deployed in 2023 and 2024. A multi-agent workflow that plans, retrieves information, calls external tools, and iterates across a task can consume orders of magnitude more compute than a single prompt-response exchange. If you are not tracking token-level consumption in real time, you have no early warning system.

35% of surveyed leaders explicitly named AI cost management and economic literacy as a barrier to realising value from their AI investments. Understanding usage-based pricing models — where costs are tied to tokens and inference volume rather than seat licences — remains a genuine operational gap in most finance and technology teams.

The Governance Gap Is Growing Faster Than the Deployment Gap

The doubling of multi-agent adoption — from 9% to 18% in a single quarter — is one of the more important data points in recent enterprise AI research. Orchestrated agent systems are not just more expensive per task. They introduce new governance challenges that single-model deployments do not.

When one AI agent hands off a task to another, and that agent calls an external API, and a third agent processes the result, the cost attribution becomes genuinely difficult to track. Which workflow consumed which tokens? Which department is responsible for which spend? Which agent orchestration is delivering ROI and which is burning budget without measurable output?

These are not edge cases. They are the operational reality of coordinated multi-agent deployment, and most organisations lack the infrastructure to answer them cleanly.

The banking sector shows how pronounced the gap can be. Among banking leaders surveyed, just 31% say AI operating costs are fully visible today. More than half (58%) say costs are only “somewhat visible” — a description that in practice means finance teams are working from estimates, not actuals.

What This Means for Business

The headline story is not that AI is too expensive. The headline story is that AI cost opacity is now a business risk.

Organisations that cannot see their AI costs clearly cannot make sound decisions about where to invest, which workflows to scale, and which deployments to shut down. They also cannot defend those decisions to boards and investors who are increasingly asking hard questions about AI ROI.

There are three practical implications for business leaders right now.

First, dashboards are not governance. If your organisation has an AI monitoring tool but cannot pull real-time token consumption by workflow, department, or use case, you have visibility theatre rather than visibility. The question to ask your team: can we see what we spent on AI agents yesterday, broken down by workflow?

Second, multi-agent rollouts need cost architecture from the start. If you are deploying coordinated agent systems — or evaluating vendors who want to sell them to you — cost attribution needs to be designed into the implementation, not bolted on afterward. Getting independent expert input on that architecture before you commit is considerably cheaper than unwinding a deployment that is costing three times what was projected.

Third, economic literacy around AI is now a leadership competency, not a technical detail. Understanding what tokens are, how inference pricing works, and how agentic workflows compound costs is no longer something you can safely delegate to your technical team. The decisions made in those conversations have material budget consequences.

The KPMG data confirms what many CFOs are learning the hard way this year: AI is not a software subscription. It is an operational cost that scales with usage, and usage in an agentic world is no longer predictable from past behaviour.

---

Want the practical version of this? The free Working With Claude field guide covers the full Claude ecosystem, Claude Code, and how to roll it out across a real business. Download it here.