KPMG Q3: more large firms now build agents, and most add cost reviews. What changed in a quarter

What KPMG found in Q3

KPMG's third-quarter AI pulse, released on 24 September, shows agent work spreading and governance tightening around cost. The survey covers 314 US C-suite and business leaders at organisations with revenue of $1 billion or more, interviewed between 24 July and 25 August 2026.¹

In that sample, 62% were building, deploying or developing AI agents, up from 53% last quarter by KPMG's comparison.¹ The Q2 release itself described its 53% as organisations using AI agents, so the two quarters' wording is close but not identical.² The share reporting significant workforce adoption of AI rose to 44%, from 23% in Q2.¹ And 73% said they were confident in their governance and capabilities, against 57% a quarter earlier.¹

The governance figures are the more useful part for anyone planning a first agent. They show what large firms now put around agents, and much of it costs little to copy.

Agents and cost controls, Q2 to Q3 2026

Building, deploying or developing agents, Q3¹62%

Q2 figure as KPMG compares it¹53%

Cost reviews in AI approvals, Q3¹74%

Q2 figure as KPMG compares it¹61%

Cost reviews moved into the approval step

The clearest change is financial. In Q3, 74% of organisations included cost reviews in their AI approval processes, which KPMG compares with 61% last quarter.¹ In the same survey, 70% used AI monitoring dashboards and 43% had usage or token budgets.¹

Read the comparison with care. The Q2 release reported 61% for approval processes in general, and 36% for direct token or usage controls, in a smaller sample of 204 leaders surveyed 28 April–25 May 2026.² KPMG treats the two as the same series. The direction is plausible, but the wording moved between releases.

The more telling Q2 number is one Q3 did not repeat. Only 26% of Q2 respondents had full, real-time visibility of what their AI systems cost to run.² Reviewing cost before approval and seeing cost while an agent runs are different controls. Our view: the second is the one that catches a runaway agent, and most firms still lack it. Our note on the Q2 pulse covers that gap.

Where firms now draw the line on autonomy

Nearly half of Q3 respondents, 49%, have defined high-risk uses in which agents are not allowed to make decisions on their own.¹ That is a written list of decisions reserved for people, and it is the cheapest control in this survey.

One figure moved the other way. The share building controls into their agents alongside monitoring and evaluation fell to 30%, from 43% two quarters earlier, and KPMG attributes this to firms relying on guardrails already in place.¹ That may be fine for a firm with a platform team. For a first agent on an ERP, it is the wrong lesson: controls inside the workflow, such as drafts that a person submits, are what stop a bad action reaching a customer or a ledger.

Read the multi-agent figure carefully

KPMG reports that 25% of Q3 respondents were actively developing or implementing multi-agent systems, against 6% in each of the previous two quarters.¹ The Q2 release used different wording: 18% were orchestrating multiple agents across workflows, up from 9%.²

Those are two questions, not one trend. Quoting them as one series would mix them. The safe reading is that multi-agent work is growing in large US firms, with the size of the jump uncertain.

It also helps to set the pulse beside a broader sample. Gartner surveyed 1,303 leaders at organisations with revenue of $50 million or more between January and April 2026 (geography not stated) and found that 22% had successfully scaled AI across multiple business units or adopted an AI-first approach.³ About 11% did not know what their own function spent on AI in 2025.³ Building agents and scaling them are still far apart.

What should a mid-sized firm copy from the Q3 pulse?

The controls large firms are adding are mostly paperwork and configuration, not platforms. A firm with one agent planned can put all four in place before launch.

1. A cost line in the approval. Estimate cost per task and monthly volume before sign-off, as in our guide to monthly running costs.
2. A usage budget. Set a monthly cap on model usage, with an alert well before it is reached.
3. A no-autonomy list. Name the decisions the agent may draft but never take: payments, price changes, anything sent to a regulator.
4. A running-cost view. Track cost per completed task weekly, so a change in exception rate shows up as money.

The pilot-to-production checklist sets these inside a fuller operating model. At Sigzen AI, our AI automation work builds them into every workflow, because a sprint's result is only credible if its cost is visible.

Our view: the pulse is a survey of large US firms, and its absolute numbers will not transfer to a 200-person company. The direction will. Agent work is moving from pilots into budgets, and the firms doing it are writing down what each agent costs and what it may never decide.