KPMG's AI Pulse: Governance Confidence Jumped. Built-In Agent Controls Fell.
In KPMG's Q3 2026 survey, 73% of large-company leaders now say they can govern AI at scale, up from 57%. Multi-agent projects jumped from 6% to 25%. But fewer companies are building controls into their agents. Here's why that matters.
Worky ClawsonHead of Growth at Workmate
Large companies are feeling much better about their ability to govern AI agents. In a single quarter, the share of US leaders who say they have the governance and capabilities to manage AI risk at scale jumped from 57% to 73%.
Over the same stretch, something else moved the other way. The share of organizations building controls directly into their agents, alongside monitoring and evaluation, fell. That combination is worth a closer look, especially as companies shift from single assistants to systems of agents working together.
What KPMG's Q3 2026 Pulse Found
KPMG's Quarterly AI Pulse Survey for Q3 2026 captured responses from 314 US-based C-suite and business leaders at organizations with $1 billion or more in annual revenue between July 24 and August 25, 2026. More than a third have revenues of $10 billion or more.
The headline is about value. 58% of leaders report measurable business value from AI. Productivity gains are still the most common (55%), but leaders also report faster decision-making (49%), better customer and employee experiences (38%), and stronger financial performance (37%).
The agent numbers moved fast:
- 62% of organizations say they are building, deploying, or developing AI agents, up from 53% the previous quarter.
- 25% are actively developing or implementing multi-agent systems, compared with only 6% in each of the prior two quarters.
- 44% report significant workforce adoption of agents, up from 23% last quarter and 10% a year ago.
Confidence Is Rising Faster Than Controls
Here is the tension. KPMG reports that 73% of leaders now express confidence in the governance and capabilities they have in place, "a steep jump from 57% last quarter." At the same time, the share of organizations building controls into their agents alongside monitoring and evaluation procedures fell to 30%, from 43% two quarters ago, "as many are relying on guardrails already in place."
That is not necessarily a bad sign. Reusing guardrails you already trust is efficient, and many companies have spent the past year putting real infrastructure in place. KPMG's Q3 report lists what organizations now include in their "AI harness layer":
- 59% data access controls
- 53% security and identity controls
- 47% monitoring of AI outputs
- 42% policy and guardrail enforcement
- 40% model access management
Financial controls have tightened too. 74% now include cost reviews in AI approval processes, up from 61% last quarter. 70% use AI monitoring dashboards, and 43% have implemented usage or token budgets. And 49% have defined high-risk use cases where autonomous decision-making by agents is not allowed.
Why Multi-Agent Systems Change the Math
The catch is timing. Most existing guardrails were designed around a person using an AI tool, or a single agent doing a single job. A multi-agent system is different: one agent hands work to another, which calls a tool, which triggers a third agent. Every handoff is a place where data can move further than intended or an action can happen that no single agent was supposed to take.
Multi-agent development roughly quadrupled in one quarter in KPMG's data, from 6% to 25%. Controls built into agents fell from 43% to 30% over the same half year. Those two lines are moving in opposite directions at exactly the moment the architecture gets more complex.
KPMG's own AI Enterprise Transformation Leader, Rahsaan Shears, framed it this way in the release: "As organizations get better at managing cost and risk, they are scaling what works. But governance is never finished. The challenge is enabling oversight to keep pace with the technology."
Data, Not Models, Is Still the Top Barrier
One more finding cuts across all of this. KPMG reports that data readiness and access is the top barrier to AI agent deployment. In the banking cut of the same survey, 80% of bank leaders named it the top challenge, up from 63% last quarter.
Data access is also the most common control in the harness layer (59%). That makes sense: the question "which data can this agent see, and on whose behalf?" is both the thing that slows deployment and the thing that most needs to be right before agents start handing work to each other.
Questions to Ask Before Your Next Agent Goes Live
If your organization is in the 73% that feels confident, a few practical checks help confirm that confidence holds up as agent systems grow:
- Do your existing guardrails see agent-to-agent handoffs? Monitoring that watches what a person types into an AI tool may not watch what one agent passes to another.
- Does each agent have its own identity and permissions? Shared service accounts make it hard to know which agent did what, and hard to cut off one without breaking the rest.
- Is there a hard stop for high-risk actions? KPMG found 49% have defined no-autonomy zones. Make sure those rules are enforced in the system, not just written in a policy.
- Can you trace a result back through every step? When a multi-agent workflow produces a wrong answer, you need to see which agent, which data, and which tool call caused it.
- Is someone accountable for each workflow? KPMG's data shows half of organizations are now establishing clear protocols for when and how to use agents, up from 23% a year earlier. Protocols work best when a named owner reviews them.
What This Means for Companies Scaling AI Agents
KPMG's Q3 data is mostly good news. Value is showing up in more places, costs are being managed, and employees are actually using agents. Confidence in governance has real foundations behind it.
The risk is treating a quarter's worth of progress as finished. As companies move from one agent to many, the controls that worked for a single assistant need to stretch to cover handoffs, identities, and actions across the whole system. The companies that keep building controls into their agents, not just around them, will be the ones whose confidence is still justified a year from now.
Frequently Asked Questions
What is the KPMG AI Quarterly Pulse Survey? It is a recurring KPMG survey of US business leaders at companies with $1 billion or more in revenue. The Q3 2026 wave surveyed 314 leaders between July 24 and August 25, 2026.
How many companies are building multi-agent systems? 25% of organizations in KPMG's Q3 2026 survey were actively developing or implementing multi-agent systems, up from 6% in the prior two quarters.
Are companies confident in their AI governance? 73% of leaders said they have the governance and capabilities to manage AI risks at scale, up from 57% the previous quarter.
What is the biggest barrier to deploying AI agents? KPMG reports that data readiness and access is the top barrier to AI agent deployment.