AI Agents at Work: Where Companies Are Actually Seeing ROI

Imagine a claims supervisor at a regional insurer in Hartford on a Tuesday morning. An AI agent has already handled 400 routine address changes and policy questions before 10 a.m., work that used to fill two people’s days.

Her CFO’s email asks a simpler question: what did it actually save us, and what did it cost?

That question is now landing on desks across corporate America. Fresh survey data says more companies can answer it than a year ago. It also shows how many still cannot.

What does the newest data say about AI agent ROI?

KPMG’s third-quarter AI Pulse Survey, released Sept. 24, 2026, is the most recent large read. It polled 314 C-suite and business leaders at U.S. companies with at least $1 billion in revenue between July 24 and Aug. 25.

Nearly six in 10 leaders reported measurable business value from AI. Adoption of agents is climbing fast:

• 62% are building, deploying or developing AI agents, up from 53% the prior quarter.
• 25% are developing or running multi-agent systems, up from 6% in each of the two quarters before.
• 44% report significant workforce adoption, up from 23% last quarter and 10% a year ago.

“AI’s value story is getting sharper,” said Todd Lohr, who leads client technology and innovation at KPMG.

Where are companies actually seeing returns?

The value is real, but it is concentrated. In KPMG’s survey, productivity was the most cited benefit at 55%, followed by faster decision-making at 49%. Stronger financial performance came last among the main categories, at 37%.

That order matters. Most companies are seeing time saved before they see it on the income statement.

Two vendor surveys point to the same functions. Google Cloud’s study of 3,466 executives in 24 countries, published Sept. 4, 2025, found that among early adopters of agents, the most common areas reporting ROI were:

Use case	Agentic early adopters	All organizations
Customer service and experience	43%	36%
Marketing effectiveness	41%	33%
Security operations	40%	30%
Software development	37%	27%

Anthropic’s 2026 State of AI Agents report lists a similar top tier: software development (57%), customer service (55%) and marketing and sales (46%). In that survey, 80% of respondents agreed their agent investments had already delivered measurable financial impact.

Treat both with care. They are self-reported surveys published by companies that sell AI products, and agreement on a survey is not an audited return.

Why are so many agent projects still struggling?

Because deployment is easier than value. Gartner predicted in June 2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls.

Gartner also warned about “agent washing,” where vendors rebrand chatbots or automation tools as agents. It estimated only about 130 of the thousands of vendors claiming agentic products were the real thing.

Finance teams are feeling the gap. In a Gartner survey of 204 finance leaders conducted in March 2026, 63% said their AI implementation in 2025 was slower than expected. Gartner analyst Marco Steecker put it plainly: “They must not mistake activity for impact.”

KPMG’s data adds a warning sign. The share of organizations building controls directly into their agents, alongside monitoring and evaluation, fell to 30% from 43% two quarters earlier. Agents are spreading faster than the guardrails.

What separates the companies getting real ROI?

The surveys point to a few habits that show up again and again.

They review the cost before they approve the project

In KPMG’s survey, 74% now include cost reviews in AI approvals, up from 61% a quarter earlier. And 43% have usage or token budgets, which cap what an agent can spend.

They draw clear lines on autonomy

Nearly half, 49%, have defined high-risk use cases where agents cannot make decisions on their own. In KPMG’s banking cut, 75% do not let agents touch sensitive data without human oversight.

They move from pilot to production

Among asset managers and private equity firms in KPMG’s industry survey, agent deployment rose to 30% from 19%, while the share still piloting fell to 34% from 52%. Value tends to show up after that shift, not before it.

They fix the data first

Among banks, 80% named data readiness and access as the top barrier to deploying agents, up from 63%. An agent is only as useful as the information it can reach.

How should executives measure AI agent ROI?

A workable scorecard keeps it simple and keeps it honest.

1. Pick one baseline metric per agent. Cost per ticket, hours per close, time to quote. Measure it for a month before launch.
2. Count the full cost. Licenses, usage or token fees, integration work, human review time and the vendor’s support tier.
3. Separate time saved from money saved. Hours freed only become savings if the work is redeployed or the budget changes.
4. Set a 90-day review. Scale it, fix it or shut it down. No pilot runs forever.
5. Read the contract. Data rights, pricing changes and liability terms can erase the return. Our guide to AI vendor contract provisions executives should read twice covers the clauses that matter most.

Leaders weighing more personal agents for their own work can also see our profile of Zinley’s bet on an AI that represents you, which looks at where individual assistants fit.

What happens next?

Expect sharper questions in 2027 budget reviews. With most large companies now building or deploying agents in some form, the debate is shifting from whether to use them to which ones earn their keep.

“But governance is never finished,” KPMG’s Rahsaan Shears added, a reminder that controls have to grow with deployment.

For the claims supervisor in Hartford, the answer to her CFO will come down to a few lines on one page: the baseline before the agent, the cost to run it and the hours her team got back to spend on harder claims. If she can show all three, the agent stays. If she cannot, it joins the projects Gartner expects to be canceled.

Frequently asked questions

What is an AI agent?

An AI agent is software that can take actions toward a goal, such as answering a customer, updating a record or drafting a report, with limited human input. It differs from a basic chatbot because it can complete multi-step tasks and use other tools or systems.

How long does it take for AI agents to show ROI?

It varies widely by use case and data readiness. Many companies see time savings within months, while measurable financial returns can take longer. Setting a baseline and a 90-day review helps show whether a project is on track.

What are the hidden costs of AI agents?

Beyond licenses, costs can include usage or token fees, integration and data cleanup, human review of agent output, security controls and staff training. These can change the return significantly, so include them in the business case.

Which departments benefit most from AI agents?

Recent surveys most often cite customer service, software development, marketing and sales, and security operations. Results depend on the quality of the data and processes the agent works with.

Will AI agents replace employees?

Some companies cite AI when announcing job cuts, but many deployments shift people toward harder or higher-value work rather than eliminating roles. The effect depends on how a company chooses to redeploy the time saved.

Sources

• KPMG: Q3 2026 AI Pulse Survey
• Google Cloud: Study reveals 52% of executives say their organizations have deployed AI agents
• Anthropic: The 2026 State of AI Agents Report
• Gartner: Over 40% of agentic AI projects will be canceled by end of 2027
• Gartner: CFOs must stop mistaking finance AI deployment for value creation