62% Are Building Agents. Nearly Half Already Ban Autonomy Somewhere.

KPMG published its Q3 2026 AI Pulse on September 24. Among large U.S. organizations, 62% are now building, deploying, or developing AI agents, up from 53% last quarter. Nearly half of leaders (49%) have already defined high-risk use cases where they do not allow autonomous decision-making by agents. That second number is the useful one. Deploying agents is easy to announce. Naming where they must stop is the start of a real workflow.

Source: KPMG Quarterly AI Pulse Survey, Q3 2026 (314 U.S. C-suite and business leaders at firms with $1B+ revenue, fielded July 24 to August 25, 2026).

What the Pulse actually says

Beyond the agent headline, a few signals matter for operators:

• 58% report measurable business value from AI initiatives
• 74% now include cost reviews in AI approval processes, up from 61% last quarter
• 25% are developing or implementing multi-agent systems, up from 6% across the prior two quarters
• 44% report significant workforce adoption, up from 23% last quarter

Productivity is still the most common value claim at 55%, with faster decisions at 49%. The survey also shows confidence in governance rising to 73% from 57%. That confidence is part of why agent deployment is accelerating. It does not mean every SMB should copy a billion-dollar control stack. It does mean the mature pattern is clear: agents expand only after someone owns the boundaries.

Why the 49% matters more than the 62%

Turning on an agent is a software decision. Banning autonomy in a high-risk case is a process decision. The second one answers: who may change price, credit, inventory commitment, or customer status without a human in the loop?

Without that line, teams treat the agent like a junior hire with no manager. It drafts, suggests, and sometimes acts. When something stalls, Slack becomes the exception system again. That is the same failure mode we see when a pilot works and the business does not change, or when a hosted harness still leaves the path unowned.

The before: agents everywhere, stops nowhere

A typical mid-market stack now has chat assistants in the CRM, draft helpers in email, and maybe a shop-floor or support bot. None of those tools know which steps are allowed to complete alone. So people invent informal rules:

• Refunds under a threshold are fine until they are not
• Inventory swaps are fine until a kit ships wrong
• Quote discounts are fine until margin disappears

Multi-agent experiments amplify this. KPMG says 25% are already building multi-agent systems. More agents without a stop list multiplies unowned handoffs.

The first build: one stop list on one painful path

We still start the way we always start: one path that hurts on the tools you already run. For agent work, that path needs an explicit stop list before autonomy expands:

• Name the trigger and the system of record for each step
• List the actions the agent may complete alone, with write-back to that record
• List the high-risk actions that require a named human, a timeout, and a single failure channel
• Keep cost visibility in the loop: if you cannot see token or tool spend on the path, you cannot manage it

That matches the Pulse direction without copying enterprise bureaucracy. Cost reviews and monitoring dashboards showed up in 74% and 70% of large orgs for a reason. On a first build, the light version is: one path, one budget signal, one owner when the agent is not allowed to decide.

What we leave alone on purpose

We do not tell you to freeze agent projects until governance is perfect. Use agents for drafts, lookups, and low-risk updates. Keep humans on money, inventory commitment, and anything that changes a customer promise. Scale autonomy only after the stop list is real.

62% are building agents. Nearly half already ban autonomy somewhere. Draw that line on one painful path before you add a second agent.