The ROI of Enterprise AI Agents: 2026 Benchmarks
The short answer
Executives expect an average return of about 171% on agentic AI (192% among U.S. companies), according to a 2025 PagerDuty survey of 1,000 executives at companies with at least $500M in revenue. Only 51% of those respondents had deployed AI agents, so the figure is an expectation held across the whole sample. The same executives self-report an average return of about 152% on the generative AI they already run. Both numbers are useful as a direction. Neither was measured on a European mid-cap in the EUR 250M to 2B range, and neither belongs in a business case until it has a baseline, a measurement window and an owner. Below are the sourced figures, what each one measures, and three questions to ask before a headline ROI number goes in front of your board.
Key Takeaways
- 171% is the average return executives expect from agentic AI, rising to 192% in the U.S. It comes from all 1,000 respondents, of whom 51% had deployed agents. The 152% in the same survey is a self-reported return on generative AI. (PagerDuty, 2025 Agentic AI ROI Survey)
- No continental European company was in that sample. A DACH or Nordic board should read 171% as a direction of travel.
- Published 2026 customer-support resolution rates run 40% to 60% after launch and 50% to 76% when mature. The 80% to 84% figures are one vendor’s top performers. (Lorikeet, 2026; Fin.ai, 2026)
- Tools built by external vendors succeed about twice as often as internal builds, and back-office functions offer some of the highest returns, per MIT’s The GenAI Divide (July 2025). (Legal.io summary of the MIT report; Dataiku)
- Before any headline ROI figure enters your case, ask for its baseline, its measurement window and who is accountable for it after go-live.
What “ROI” actually means for an AI agent in 2026
For a board, agent ROI reduces to one formula:
Value released = (hours automated × loaded headcount cost) + error and rework avoided + revenue protected, minus (build + run + governance cost).
The headline 171% fills none of those terms. It is self-reported, it is an expectation, and it was given by executives of whom roughly half had agents live. The 152% is closer to a result, but it is also self-reported, and it describes generative AI in general rather than agents. A separate line of research looks at the other end: MIT’s GenAI Divide study found that 95% of generative AI pilots show zero financial return (Legal.io). The two studies ask different questions of different people, and a business case needs both answers:
- What does a system return once it is live? The best available survey answer is a self-reported 152% on generative AI and an expected 171% on agentic AI.
- What is the probability of getting one live? Low, the way most pilots are run. The reasons are covered in why enterprise AI pilots fail.
A business case that answers only the first question is a forecast. One that answers both is something a CFO can sign.
2026 ROI benchmarks by number
| Benchmark | Figure | What it measures | Source |
|---|---|---|---|
| Average expected agentic AI ROI | 171% (U.S.: 192%) | Executive expectation across all respondents | PagerDuty, 2025 |
| Average self-reported GenAI return | 152% | Self-reported return on generative AI | PagerDuty, 2025 |
| Respondents with AI agents deployed | 51% | Share of the 1,000 surveyed | PagerDuty, 2025 |
| Companies expecting more than 100% agentic ROI | 62% | Expectation | PagerDuty, 2025 |
| Enterprise GenAI investment studied | $30 to 40B | Total enterprise AI spend in the MIT report | MIT NANDA via Legal.io |
| External vendor tools vs internal builds | about twice as often successful | Delivery model | MIT NANDA via Legal.io; Dataiku |
| CX resolution, first months after launch | 40% to 60% | Problems solved end to end | Lorikeet, 2026 |
| CX resolution, mature deployments | roughly 50% to 76% | Problems solved end to end | Lorikeet, 2026 |
| CX resolution, vendor top performers | 80% to 84% | One vendor’s best customers | Fin.ai, 2026 |
Two cautions on reading this table. First, the resolution figures come from vendors: Lorikeet and Fin both sell customer-support AI. Lorikeet’s round-up is useful because it sets several vendors side by side, for example Ada’s reported average of about 52% across more than 550 deployments against Fin’s 76% across 12,000 customers. Second, resolution and deflection are different numbers. Deflection counts conversations no human touched. Resolution counts problems actually solved. When you underwrite ROI, underwrite the resolution number, measured on your own traffic.
Who the 171% survey asked
The PagerDuty survey covered 1,000 IT and business executives at companies with at least $500M in annual revenue, split across the U.S., the U.K., Australia and Japan, between late February and early March 2025. No continental European market was in the sample. For a mid-cap board in DACH or the Nordics, 171% says where executive expectations sit. The number that belongs in your case is the one you measure on your own baseline.
Named production results, and what they leave out
Klarna’s customer-service assistant is the most quoted case. In its first month live, by Klarna’s own account, it handled 2.3 million conversations, two-thirds of its customer-service chats, the equivalent work of 700 full-time agents. Errands that took 11 minutes took under 2, repeat inquiries fell 25%, and it ran in 23 markets in more than 35 languages. Klarna estimated a $40 million profit improvement for 2024 (Klarna press release, February 2024).
The press release is not the end of the story. In May 2025 Klarna’s CEO told Bloomberg the company was hiring human customer-service agents again, after an AI-led hiring freeze, and that the AI-focused strategy had not been the right path (Entrepreneur, May 2025). The first-month numbers were real. The profit figure was an estimate, and the operating model behind it changed within about fifteen months. Klarna is also far larger than most European mid-caps.
A Nordic example shows the same pattern at a different scale. SpareBank 1 SR-Bank, which its vendor describes as Norway’s second-largest bank, runs an email automation system for a customer service team that received more than 420,000 email inquiries a year. The vendor’s case study says the system is live and that the bank has projected an ROI of more than 350% over the next three years (Simplifai case study). That is a production deployment with a projected, vendor-published return, and it should be read as one.
MIT’s case studies point the same way for back-office work: $2M to $10M in annual savings from replacing outsourced support and document review, a 30% reduction in external agency spend for marketing and content work, and $1M a year saved in financial risk monitoring (MIT NANDA via Legal.io).
Three questions to ask before a headline ROI number goes into your case
Most ROI figures in vendor decks travel further than their sources. A number from a survey of expectations becomes a result two decks later, and a projection becomes a benchmark. Before any of these figures goes in front of your board, ask three questions of whoever is quoting it.
- What was the baseline before the system went live, and who measured it? A return is a difference between two states. If nobody recorded handling time, error rates or cost per case before the build, the return is an estimate of what the old process probably cost. The PagerDuty figures are self-reported by executives. The SR-Bank figure is published by the vendor.
- Over what window was the return measured, or is it a projection? First-month volume and a three-year projection are different kinds of evidence. Klarna’s figures describe one month of volume. Fifteen months later its operating model had changed, and by then the headline had already circulated.
- Who is contractually accountable for the number after go-live? If the vendor’s fee does not depend on the figure, the figure is marketing. If someone has agreed to be measured on it, against your baseline and in a window set in advance, it is a commitment.
A figure that cannot answer all three can stay in the appendix as context. It should not be the number the CFO signs.
Where Nucleo stands
Nucleo is an AI-native transformation company for European mid-caps. We do not sell an agent platform. We re-architect one workflow where value is stuck, put it into production on infrastructure you own, model-agnostic and governed to EU rules, and agree the number it has to move before the build starts. The intelligence stays yours.
One thing we will not do is quote a benchmark as our result. We have no published engagement data, and the figures in this post are other companies’ numbers. The ROI we would put in front of your board is the one measured against your own baseline, in a window agreed in advance, with us accountable for it. For the delivery-model choice behind that, see build vs buy vs embed and what a forward-deployed engineer does.
FAQ
What is a realistic ROI for an AI agent in 2026? Executives surveyed by PagerDuty in 2025 expect an average 171% return on agentic AI (192% in the U.S.). That expectation comes from all 1,000 respondents, of whom 51% had deployed AI agents. The same executives self-report an average 152% return on generative AI. Neither figure is a measured outcome, so the number for your business case is the one you measure on your own baseline.
Which use cases have the best ROI? MIT’s GenAI Divide report, as summarised by Legal.io, says back-office functions offer some of the highest returns, even though AI budgets overwhelmingly favour sales and marketing. Successful organisations tend to deploy first in document automation, procurement and risk review. The report’s case studies include $2M to $10M in annual savings from replacing outsourced support and document review.
What resolution rate should I expect from a customer-service agent? Published 2026 benchmarks put AI customer-support resolution at roughly 40% to 60% in the first months after launch and 50% to 76% for mature deployments. Vendor top performers report 80% to 84%. Ask for resolution (problems actually solved) rather than deflection (conversations merely avoided), and measure it on your own traffic.
Should we build agents internally or partner? MIT’s data favours buying or partnering: tools built by external vendors succeed about twice as often as internal builds. The question we would add is whether the people delivering stay accountable for a number once the system is live. The decision rule is in build vs buy vs embed.
What should I ask a vendor who quotes a headline ROI number? Three things. What was the baseline before the system went live, and who measured it? Over what window was the return measured, or is it a projection? Who is contractually accountable for the number after go-live? A figure that cannot answer all three belongs in a footnote, not in the business case.
Does the 171% ROI figure apply to European mid-caps? Not directly. The PagerDuty survey behind it covered 1,000 executives at companies with at least $500M in revenue in the U.S., the U.K., Australia and Japan. No continental European market was in the sample, so a DACH or Nordic board should read it as a direction of travel.
Related reading
- Why enterprise AI pilots fail (and what the ones that ship do)
- Build vs buy vs embed: choosing an enterprise AI delivery model
- Salesforce’s own partners are not seeing Agentforce ROI
- What is a forward-deployed engineer, and why enterprise AI needs one
- What is an AI-native enterprise?
Nucleo starts with one workflow and a number agreed before the build. See how an engagement works, or write to us with the workflow you would start on.
Sources: PagerDuty, 2025 Agentic AI ROI Survey; MIT Project NANDA, The GenAI Divide (2025), via Legal.io; Dataiku on the MIT findings; Lorikeet resolution-rate benchmarks (2026); Fin.ai CX ROI benchmarks (2026); Klarna press release (2024); Entrepreneur on Klarna rehiring human agents (2025); Simplifai, SpareBank 1 SR-Bank case study.