Large US enterprises plan to invest a weighted average of $202 million in AI over the next twelve months, according to KPMG's Q2 2026 AI Pulse Survey of 204 C-suite leaders at $1 billion-plus organisations. In the same survey, 26 percent said they have real-time visibility of what their AI actually costs.
Those two numbers describe the state of enterprise agentic AI more precisely than any benchmark. The money is committed; the instrumentation is not. And Gartner's June 2025 prediction — that over 40 percent of agentic AI projects will be cancelled by the end of 2027 on cost, unclear value, and inadequate risk controls — reads less like a forecast now and more like arithmetic.
What the Q2 Data Says
Agent deployment has stopped climbing: 53 percent of organisations, down marginally from 55 percent the previous quarter. Beneath the flat headline, three things moved sharply. Organisations orchestrating multiple agents doubled, from 9 to 18 percent. Banking-sector deployment jumped from 23 to 56 percent in a single quarter. And employee resistance quadrupled, from 5 to 20 percent.
The governance picture explains the plateau better than any technology argument. Monitoring dashboards exist in 66 percent of organisations and approval processes in 61 percent — the watching layer is half-built. The controlling layer is not: 36 percent have direct token or usage controls, and 26 percent have real-time cost visibility. Organisations can increasingly see what their agents do; far fewer can constrain what they spend.
The resistance data carries the most information per percentage point. Of the 20 percent now resisting, 53 percent cite trust and ethics and 51 percent cite workload and complexity — both up sharply — while the fear-based drivers fell: skills concerns dropped from 76 to 57 percent, job-security concerns from 67 to 55 percent. Employees are not afraid of the agents. They distrust how the agents are being rolled out, which is a governance verdict delivered from below.
What the Data Does Not Say
The sample is 204 US-based leaders at organisations with $1 billion-plus revenue, a third above $10 billion. It says little directly about UK mid-market organisations, where budgets are smaller but the visibility problem is, in our experience, identical. One flat quarter is not a trend reversal — 53 percent against 55 could be noise. And self-reported deployment tells you an agent exists somewhere in the estate, not that it does economically meaningful work; the survey cannot distinguish a production system from an expensive pilot wearing production clothes.
What Separates the Leaders From the Stalled
The concentration pattern is the story. While overall deployment went sideways, multi-agent orchestration doubled and banking went from 23 to 56 percent — the organisations already in production are compounding, while the rest circle. PwC's 2026 AI performance study found three-quarters of AI's economic gains are being captured by just 20 percent of companies. This matches the pattern we documented in our enterprise adoption reality check: the gap between leaders and laggards is not widening because leaders have better models — everyone has the same models — but because they built the operational scaffolding that turns model capability into shipped work.
The orchestration signal deserves particular attention from teams still on single-agent deployments: coordinated multi-agent systems are where the leaders went next, and the framework choices involved are the subject of our comparison of orchestration tooling.
A Governance Checklist Drawn From the Numbers
- Instrument cost before scaling spend. Real-time cost visibility is the rarest governance layer (26 percent) and the one that decides whether the CFO conversation in Q4 is about value or about surprise. Token-level controls (36 percent) are the enforcement half of the same discipline.
- Do not confuse dashboards with control. Two-thirds of organisations can watch their agents; barely a third can constrain them. If an agent can spend without limits, your monitoring dashboard is a record of the incident, not a defence against it.
- Treat resistance as a signal, not an obstacle. When 53 percent of resistance is trust-shaped, the fix is transparency about what agents do and where their output goes — not another enablement webinar.
- Isolate before you orchestrate. Multi-agent systems multiply the blast radius of a single misbehaving agent. The isolation options are covered in our guide to sandboxing agent code; orchestrating agents without them is how Gartner's 40 percent cancellation cohort gets populated.
If this analysis is useful, the easiest way to get the next one is by email. Sign up for new-post updates.
Frequently asked questions
How many enterprises have deployed AI agents in 2026?
KPMG's Quarterly AI Pulse Survey of 204 US C-suite leaders at $1 billion-plus organisations, fielded April to May 2026, puts AI agent deployment at 53 percent — marginally down from 55 percent the previous quarter. The headline number has plateaued, but the composition is shifting: organisations orchestrating multiple agents doubled from 9 to 18 percent in a quarter, and banking-sector deployment jumped from 23 to 56 percent. Deployment is concentrating and deepening rather than broadening.
Why do agentic AI projects fail?
Gartner predicted in June 2025 that over 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. The 2026 governance data supports that diagnosis: in KPMG's Q2 pulse, only 26 percent of large enterprises have real-time visibility of their AI costs and only 36 percent have implemented direct token or usage controls — while planning an average $202 million of AI investment over twelve months. Projects rarely fail because the model is weak; they fail because nobody can see what they cost or prove what they return.
What does AI agent governance actually require?
Four layers, in order of how often they are missing: real-time cost visibility (present in only 26 percent of large enterprises per KPMG's Q2 2026 pulse), direct token and usage controls (36 percent), approval processes for agent deployment (61 percent), and monitoring dashboards (66 percent). Most organisations have built the watching layer and skipped the controlling layer. Add workload isolation for agents that execute code, and a trust programme for the workforce — employee resistance quadrupled from 5 to 20 percent in a single quarter, and 53 percent of it is driven by trust and ethics concerns.
Is employee resistance to AI agents growing?
Sharply. KPMG's Q2 2026 pulse recorded employee resistance at 20 percent, quadruple the 5 percent of the previous quarter. The composition matters more than the level: 53 percent of resistance is driven by trust and ethics concerns and 51 percent by workload and complexity — while fear-based drivers actually fell, with skills concerns down from 76 to 57 percent and job-security concerns down from 67 to 55 percent. Employees are not resisting AI because they fear it; they are resisting how it is being deployed.
← All posts