Half of large US firms deploy AI agents; a quarter can see what their AI costs

What KPMG asked, and whom

KPMG's AI Quarterly Pulse for the second quarter of 2026, released on 24 June, surveyed 204 US-based C-suite and business leaders.¹ All worked at organisations with annual revenue of $1 billion or more, and the survey ran from 28 April to 25 May 2026. More than a third of those firms had revenue of $10 billion or more. This is a survey of very large US companies, and its findings should be read that way.

The release leads with agents. 53% of organisations were deploying AI agents, against 55% the previous quarter, so adoption held roughly level.¹ The bigger change was in how agents are used: the share orchestrating multiple agents across workflows doubled from 9% to 18%.¹

KPMG lists the uses as aligning shared goals and metrics across functions (64%), supporting joint decisions (49%) and automating cross-functional workflows (48%).¹ Those are coordination jobs, and coordination jobs multiply the number of model calls behind each task.

What "deploying" means here

"Deploying" in a pulse survey means a leader says the organisation uses AI agents. It does not say how many agents, in how many teams, at what volume or with what result. One agent drafting replies for a single support queue counts the same as fifty agents running finance operations.

Other surveys draw the line in a different place. Gartner's 2026 CIO and Technology Executive Survey, published in April 2026 with the data period not stated, found 17% of organisations had deployed AI agents.² The gap with KPMG is likely partly about who was asked: KPMG surveyed only $1 billion-plus US firms, while Gartner gives no sample details for its survey of CIOs and technology executives. "Deployed" is also a stricter bar than "deploying".

Our view: neither number tells a mid-sized company whether to deploy. They tell you agents are common in large firms, and that "we have agents" has become an easy thing to say. The harder question is what each one costs to run.

How many firms can see what their AI costs?

About a quarter of them, and that is the finding with the most practical weight. Two-thirds of organisations had monitoring dashboards (66%) and 61% had approval processes, but only 26% reported full, real-time visibility into what their AI systems cost to operate.¹ KPMG frames the economics of running AI at scale as the emerging challenge.

Controls in place vs cost visibility

Monitoring dashboards¹66%

Approval processes¹61%

Full, real-time visibility of AI operating costs¹26%

The gap makes sense once you look at how agent costs arise. A dashboard shows tokens, calls and errors. It rarely shows cost per completed task, because that needs three things joined together: model and tool spend, the number of tasks finished, and the minutes people spend on the exceptions the agent hands back. Multi-agent setups make the join harder, since one task can trigger several agents and many calls.

The risk is familiar. Gartner predicted in June 2025, more than a year before this note, that over 40% of agentic AI projects will be cancelled by the end of 2027, naming escalating costs among the causes.³ That is a forecast, not an outcome, but cost blindness is one of the ways it would come true.

The running-cost check before a second agent

Before approving a second agent, put a cost per completed task on the first one. It takes a month of logs and a spreadsheet, and it turns "the agent is live" into a number a finance lead can compare with the manual process.

1. Count completed tasks for the month, from the system of record, not the agent's own log.
2. Add model spend for the month, including retries and failed runs.
3. Add tool and API fees the agent triggers: document parsing, search, messaging.
4. Add human exception time: items the agent handed back, multiplied by the minutes each took and the loaded hourly cost.
5. Add a share of the owner's time for monitoring, prompt changes and model upgrades.
6. Divide by completed tasks, and compare with the cost per task before the agent.

The exception line is usually the largest and the most often left out. An agent with cheap model calls and a high exception rate can cost more per task than the process it replaced. Our guide to choosing which process to automate first scores exception rate before the build for that reason.

What this means for a mid-sized company

A company smaller than KPMG's sample has one advantage: fewer agents, so the cost join is easier. Build it in from the first deployment rather than retrofitting it at the tenth.

In practice that means three design choices. Log every agent run with a task ID that matches the system of record. Route exceptions through a queue that records who handled each item and for how long. Tag model and tool spend by agent, not by account. Our guide to the controls an invoice agent needs in the ERP shows where those logs sit in a finance workflow.

Our view: cost per completed task should be in the acceptance criteria of any agent build, signed before it starts. On our AI automation work, the sprint ends with hours saved signed off by the process owner, and managed AI-ops includes cost-per-task monitoring each month, at prices listed on our pricing page. Whoever runs your agents, ask for that number before you ask for the next agent.