AI Agent ROI in 2026: We Traced the Most-Quoted Numbers to Their Source

We checked 11 AI agent ROI statistics that keep showing up in 2026 round-ups against the publisher's own page. Four hold up as quoted. Three are real numbers attached to the wrong company, the wrong unit or the wrong technology. Four have no primary source we could find. The most-shared one, "80% of organizations already see measurable returns from AI agents," is real. It comes from Anthropic's 2026 State of AI Agents report, and it is a self-reported survey answer.

This matters if you are writing an AI agent ROI case for a budget meeting. Quote "$117,558 a month per agent workflow" to a CFO and you are off by roughly 11x, because IDC's number is per company, not per workflow. Quote "171% average ROI" and someone will eventually notice it is an expected return from a vendor survey, not a measured one.

Below: the full verdict table, what each source actually says, the four numbers to stop using, and a short list of figures you can put in an ROI deck without getting caught out.

Which AI agent ROI statistics are real?

We read each primary source between October 29 and 30, 2026: the PDF, press release or research page put out by the organization that ran the study. A number is verified only when the publisher's own page states it. Misquoted means the number exists but the round-ups credit it to the wrong publisher or describe it wrongly. Untraceable means we found it only in secondary posts that don't link a primary source.

What we found: the 80% figure has been called unverified because some summaries of Anthropic's report don't show it. It is in the PDF three times: twice in the foreword (pages 3 and 4) and on the page titled "80% of leaders say AI agents are delivering financial value today." When a quote looks too good, read the primary document before deciding it is fake.

What does Anthropic's 2026 State of AI Agents report actually say?

The report is a 48-page PDF linked from Anthropic's report page. Anthropic ran it with research firm Material and surveyed "over 500 technical leaders" in the US: engineering leaders, IT executives and technical decision-makers at companies from startups to large enterprises. The methodology note says it was fielded in late 2025, and the file's own metadata dates it to December 2025. Treat the data as a late-2025 snapshot, even though round-ups often present it as fresh 2026 research.

The ROI numbers, as printed:

• 80% say their AI agent investments "are already delivering measurable economic impact today" (p. 17). The foreword adds that another 1 in 10 expect economic impact in the future.
• 88% expect continued or increased returns.
• 96% felt optimistic about the business impact of agents at their company, and 97% expect more efficiency gains over the next 12 months (both in the Economic Index section).

The adoption and barrier numbers people pair with them:

Three caveats belong next to any of these numbers:

1. It is a vendor survey. Anthropic sells the models these agents run on and ran the survey, and the report's second half is customer case studies. That doesn't make the numbers wrong. It does mean "80% see ROI" is a self-reported answer from a sample selected by a vendor, with no published question wording, segment sample sizes or margin of error.
2. "Measurable economic impact" is undefined. The report doesn't say how respondents measured it, over what period, or net of what costs. A team that saved 10 hours a week and a team that cut a headcount both count.
3. The report disagrees with itself in places. The foreword says "nearly 90%" of organizations use AI for coding; page 9 says "more than 9 in 10." Quote the page you are citing.

The case studies (Novo Nordisk's 10+ weeks to 10 minutes for clinical documentation, eSentire's 5 hours to 7 minutes for threat analysis, Doctolib shipping features 40% faster) are customer-reported results in a vendor document, not survey data. Some round-ups present them as "what organizations reported as representative." The report doesn't say that.

Where did the 171% ROI figure come from?

It comes from PagerDuty, not Deloitte. The PagerDuty survey page says 62% of companies expect more than 100% ROI from agentic AI, "with the average expected return at an astounding 171% ROI." US companies expect 192%. The same respondents report a 152% average return from generative AI.

Wakefield Research ran it for PagerDuty: 1,000 IT and business executives (Director level and up) at companies with $500 million or more in revenue, 250 each in the US, UK, Australia and Japan, fielded February 27 to March 6, 2025.

So the honest version is: executives at large companies expected agentic AI to return 171% on average, in a March 2025 survey run for an incident-management vendor. It measures expected returns, not realized ones, even though 51% of respondents said they had already deployed AI agents.

Is $117,558 a month the cost of one AI agent?

No. IDC's post (August 28, 2026, from its Future Enterprise Resiliency and Spending Survey, Wave 4, July 2026) says enterprises "with visibility into their own agent costs (most do) report average monthly spend on agent inference and related orchestration services of $117,558." That is company-wide spend. The same post says 95% of enterprises run at least one company-funded agent workflow in production, and "the average organization is running roughly 11 of them."

Dividing one average by another is rough, but it gives you the order of magnitude: about $10,700 a month per workflow in inference and orchestration (our arithmetic, not IDC's). That excludes the build, the people who review agent output, and integration work. For build costs, our AI agent development cost breakdown covers what a project costs before it runs.

The rest of IDC's numbers from that post are verified and worth quoting:

IDC also reports that 61.8% of organizations rate their own cost governance as "defined" or "optimizing." Read next to the 67% overrun rate, many teams think they have cost control and don't.

Which AI agent ROI numbers have no source?

Four numbers repeat across AI agent ROI round-ups and appear nowhere we could find on a publisher's own page. Until someone links the original report, leave them out of anything a finance team will read.

The 41% payback rate and the 5.1-month median

"41% of agent deployments reach positive ROI within 12 months" and "5.1-month median time to value" usually travel together, along with function-level payback figures (3.4 months for outbound sales, 11.2 for legal). Depending on the post, they are credited to Gartner, to BCG and Forrester, or to Bain. When several firms get credit for one number and nobody links a report, treat the number as unsourced.

The $0.46 vs $4.18 cost per ticket

We found no source for either number, and no post that cites one. Real per-resolution prices exist, because support vendors publish them on their pricing pages, but those are list prices, not measured cost per ticket. For your own support numbers, our ROI calculator works from your volume and wage inputs instead.

The 80% of enterprise apps figure

Credited to Gartner, but Gartner's own August 2025 release predicts "up to 40%" of enterprise apps will include task-specific agents by 2026, up from under 5%. That is a forecast, and half the size.

One more misattribution shows up next to these: the line "40% of AI agent projects risk shutdown," credited to Deloitte on the same benchmark page that credits Deloitte with the 171%. The number is Gartner's, and it's a prediction: "over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls." In the same release, Gartner estimates only about 130 of the thousands of agentic AI vendors are real.

Watch out: a stat repeated across 30 sites is not 30 confirmations. Most AI statistics pages copy each other. Count how many link a primary document, not how many repeat the number.

Which numbers can you safely put in an AI agent ROI case?

If you are building the internal case for an agent project, these hold up and come with a source a CFO can open:

Notice what is missing: a benchmark for your payback. No public dataset we found gives a trustworthy payback period by function or use case. The useful number is the one you calculate from your own baseline: hours spent on the task today, error and rework rate, volume, and the run cost from your provider's pricing. The automation payback calculator does that math from your own inputs.

Two things the verified studies agree on are worth building into the plan from day one:

1. Meter cost per workflow before launch. Under half of IDC's respondents have real-time cost dashboards, and two-thirds went over budget.
2. Budget for oversight. The ROI surveys above don't net out the time people spend reviewing agent work. October 2026 studies on that review cost are in our piece on the AI agent distrust tax.

If you want a second opinion on an agent business case before you commit budget, that is part of how we scope AI agent development projects: baseline first, then the build. A free consultation is the quickest way to get that.

How we checked each number

For every claim we searched for the exact figure, collected the round-ups repeating it, and followed their citations back. A claim counted as verified only when we read the number on the publisher's own page: Anthropic's PDF, PagerDuty's survey page, Gartner's press releases, IDC's research post, Google Cloud's press release, and PwC's survey release. We read Anthropic's report in full rather than relying on summaries.

We left out figures we couldn't open at the source. McKinsey's 2026 State of AI numbers, which circulate widely, are one example: its site blocked our reader during this check, so they aren't in the table either way. We'll update this page when we confirm them. If you have the primary source for any number we marked untraceable, send it and we'll re-check it.

FAQ

What is the average ROI of AI agents in 2026?

There is no trustworthy average. The most-quoted figure, 171%, is the return executives expected in PagerDuty's March 2025 survey, not a measured result. Anthropic's late-2025 survey found 80% of US technical leaders say agents already deliver measurable economic impact, but it doesn't put a percentage on that return.

Did Anthropic's 2026 State of AI Agents report say 80% of companies see ROI?

Yes. Page 17 of the PDF says 80% of organizations report their AI agent investments are already delivering measurable economic impact, and 88% expect continued or increased returns. It surveyed over 500 US technical leaders with research firm Material in late 2025. The answers are self-reported, and the report doesn't define "measurable."

How much do AI agents cost per month?

IDC found enterprises with visibility into agent costs spend an average of $117,558 a month on agent inference and orchestration across all their agent workflows, about 11 per organization. That works out to roughly $10,700 per workflow by our arithmetic, before build and review costs.

What percentage of AI agent projects fail?

Gartner forecasts that over 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and weak risk controls. It is a prediction from June 2025, not a measured failure rate. We found no primary study that measures agent project failure rates directly.

How long does it take for an AI agent to pay back?

No public study we could trace gives a reliable payback period for AI agents. The widely shared "5.1-month median" and "41% positive within 12 months" figures have no primary source we could find. Calculate payback from your own baseline: current hours and error rate on the task, volume, and the run cost from your provider's pricing.

Is the $0.46 per ticket AI agent cost real?

We couldn't trace it. Neither $0.46 per AI-resolved ticket nor $4.18 per human-resolved ticket appears on any primary source we found. Support vendors do publish per-resolution list prices, but those are prices, not measured costs.

Which AI agent statistics should I use in a business case?

Use numbers a reader can open at the source: Anthropic's 80% (self-reported, late 2025), IDC's 67% budget overrun rate and $117,558 monthly enterprise spend, Gartner's 40% cancellation forecast, and PwC's 56% of CEOs with no significant AI benefit yet. Then put your own baseline next to them, because none of these surveys measure your process.

---

Sources: Anthropic, The 2026 State of AI Agents Report (with Material), PagerDuty agentic AI survey (Wakefield Research, March 2025), IDC, The agent economy is scaling faster than it can be metered (Aug 28, 2026), Gartner, over 40% of agentic AI projects will be canceled by end of 2027 (June 25, 2025), Gartner, 40% of enterprise apps will feature task-specific AI agents by 2026 (Aug 26, 2025), Google Cloud ROI of AI study (Sep 4, 2025), PwC 29th Global CEO Survey (Jan 20, 2026)