AI Agent Statistics 2026: Adoption, Commerce and Failure Rates

Autonomous AI agents were the loudest story in enterprise technology through 2025, and 2026 is the year the numbers started to catch up with the noise. Agents now book tasks in CRMs, browse the web on users' behalf, and, through new payment rails from Visa, Mastercard and Google, can complete purchases with real money.

The data tells a messier story than the keynotes. Adoption surveys show most large organisations experimenting, benchmark studies show agents failing most long, multi-step tasks, and Gartner expects a large share of agentic projects to be cancelled before the end of 2027.

This article pulls together the verified numbers on enterprise adoption, agentic commerce, agent and bot traffic on the web, browser agents, failure rates and market projections. Every figure below was checked against the original source, and projections are labelled as projections.

Key AI agent statistics for 2026

• 62% of organisations are at least experimenting with AI agents, but only 23% are scaling them anywhere in the enterprise, according to McKinsey [3]
• Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls [1]
• Gartner estimates only around 130 of the thousands of vendors claiming agentic AI have genuine agentic capabilities, a practice it calls "agent washing" [1]
• Only one in five companies has a mature model for governing autonomous AI agents, per Deloitte's survey of 3,235 senior leaders [4]
• The best AI agent completed just 30% of simulated workplace tasks autonomously on the TheAgentCompany benchmark [11]
• Leading LLM agents achieved roughly 58% single-turn success on Salesforce's CRMArena-Pro benchmark, collapsing to about 35% in multi-turn settings [12]
• Automated bot traffic made up 51% of all web traffic in 2024, overtaking humans for the first time in a decade, per the Imperva Bad Bot Report [9]
• AI "user action" crawling, where an agent fetches a page because a human asked it something, grew more than 15x in 2025, per Cloudflare [10]
• Visa and Mastercard both launched agent payment programmes within a day of each other in late April 2025 [6][7]
• More than 60 organisations backed Google's Agent Payments Protocol (AP2) at launch in September 2025 [8]
• 39% of US consumers have used generative AI for online shopping, and 53% planned to during 2025, per an Adobe survey of over 5,000 respondents [17]
• Grand View Research valued the AI agent market at $7.6 billion in 2025, projecting $182.9 billion by 2033. That is a projection, not an observation [16]

How many enterprises actually use AI agents in 2026?

Most large organisations are experimenting with agents, but genuine scale remains rare. Surveys from McKinsey, Deloitte and LangChain converge on the same picture: broad piloting, narrow production deployment, and governance lagging well behind enthusiasm.

McKinsey's State of AI survey, fielded June to July 2025 across 1,993 respondents in 105 countries, found 62% of organisations at least experimenting with AI agents. Only 23% report scaling agentic systems anywhere in the enterprise, and no more than 10% report scaling agents in any single business function [3].

Deloitte surveyed 3,235 board members, C-suite and senior leaders between August and September 2025 and found only one in five companies has a mature model for governing autonomous AI agents [4]. Among hands-on builders, LangChain's survey of over 1,300 professionals found about 51% using agents in production, with mid-sized companies the most aggressive at 63% [5].

The gap between the 62% "experimenting" figure and the 23% "scaling" figure is the real story. Pilots are cheap and low risk, so they tell you little about whether agents deliver value. The same McKinsey survey found just 39% of organisations report any EBIT impact from AI at the enterprise level, and most of those attribute less than 5% of EBIT to it [3].

Note that these surveys measure different populations. LangChain's respondents are self-selected developers already building with agent frameworks, so their 51% production figure should not be compared directly with McKinsey's enterprise-wide 23%.

How many AI agent projects fail or get cancelled?

A large share, on Gartner's reading. The firm expects over 40% of agentic AI projects to be cancelled by the end of 2027, and warns that most vendors selling "agents" are not selling agents at all.

Gartner's June 2025 prediction cites "escalating costs, unclear business value or inadequate risk controls" as the drivers [1]. The same release estimates that only about 130 of the thousands of vendors marketing agentic AI possess genuine agentic capabilities, with the rest rebadging chatbots and RPA, a practice Gartner labels agent washing.

Gartner's January 2025 poll of 3,412 webinar attendees adds texture: 19% reported significant agentic AI investment, 42% conservative investment, 8% none, and 31% were waiting or undecided [1].

A 40% cancellation forecast sounds damning but needs context. Early-stage enterprise technology projects fail at high rates in any cycle, and a cancelled pilot often costs little. The sharper warning is the agent washing estimate, because it means many buyers evaluating "agentic" tools in 2026 are not evaluating agents at all, and their disappointment will show up in the failure statistics either way.

How often do AI agents actually complete tasks?

Not often, once tasks get long. On realistic multi-step benchmarks, the best agents complete a minority of tasks, and performance drops sharply the moment a task requires sustained multi-turn work.

On TheAgentCompany, a Carnegie Mellon-led benchmark simulating a software company with browsing, coding and colleague communication, the most competitive agent completed 30% of tasks autonomously. The authors note that "more difficult long-horizon tasks are still beyond the reach of current systems" [11].

Salesforce AI Research's CRMArena-Pro benchmark, built on realistic sales and service scenarios, found leading LLM agents achieve only around 58% single-turn success, dropping to roughly 35% in multi-turn settings. Agents also exhibited "near-zero inherent confidentiality awareness" [12].

These numbers explain the adoption gap in the previous section. A 58% success rate is impressive for a research benchmark and unacceptable for an unsupervised production workflow, which is why scaled deployments cluster in tasks where a human reviews the output or the cost of error is low.

Be careful comparing benchmark scores across papers. Task definitions, scoring and scaffolding differ, and vendors' own demos are not benchmarks.

How much web traffic now comes from bots and agents?

More than half of it, on Imperva's measurement. Automated traffic overtook human traffic in 2024, and 2025 added a fast-growing new category: AI agents fetching pages on behalf of individual users in real time.

The 2025 Bad Bot Report from Imperva, a Thales company, found automated bot traffic constituted 51% of all web traffic in 2024, surpassing human activity for the first time in a decade. Malicious bots alone accounted for 37% of all internet traffic, up from 32% in 2023 and the sixth consecutive year of growth [9].

Cloudflare's 2025 Year in Review shows what changed within the automated share. AI "user action" crawling, driven by bots like ChatGPT-User that fetch pages when a human asks a question, increased more than 15x in 2025. AI crawlers generated 20% of verified bot traffic overall, with OpenAI's GPTBot at roughly 7.5% of verified bot traffic against Googlebot's quarter-plus share [10].

The distinction between training crawlers and user-action agents matters commercially. A training crawl extracts value with no visit attached, whilst a user-action fetch represents a live human decision being made about your content, just without a conventional pageview. Analytics suites built around human sessions undercount both.

Treat the 51% figure as one vendor's network view, not a census. Cloudflare, Imperva and others see different customer bases and classify bots differently, which is why their headline shares never quite agree.

Can AI agents actually buy things yet?

The payment rails now exist. Visa, Mastercard and Google all shipped agent payment infrastructure between April and September 2025, built around tokenised credentials and cryptographically signed proof that a human authorised the purchase.

Visa announced Intelligent Commerce on 30 April 2025, opening APIs that let AI agents "browse, select, purchase and manage" on consumers' behalf, with partners including Anthropic, IBM, Microsoft, Mistral AI, OpenAI, Perplexity, Samsung and Stripe [7]. Mastercard unveiled Agent Pay a day earlier, on 29 April 2025, using Agentic Tokens and partnerships with Microsoft and IBM [6].

Google Cloud followed on 16 September 2025 with the Agent Payments Protocol (AP2), an open standard backed by more than 60 organisations including Mastercard, PayPal, American Express, Adyen and Worldpay. AP2 rests on "mandates", tamper-proof, cryptographically signed digital contracts proving what the user asked the agent to do [8].

Infrastructure is not adoption. These programmes solve the liability question, who pays when an agent buys the wrong thing, before the volume question, and none of the three published transaction figures at launch. The bet is that agent-initiated payments will need the same rails-first build-out card networks gave e-commerce in the 1990s.

Are consumers actually shopping with AI?

Increasingly yes, though mostly for research rather than delegated purchasing. Adobe's survey data shows around four in ten US consumers have used generative AI for online shopping, and AI-referred visitors browse more but still convert slightly less.

An Adobe survey of more than 5,000 US consumers found 39% had used generative AI for online shopping, with 53% planning to do so during 2025 [17].

Adobe's traffic data from the same study showed generative AI referrals to US retail sites grew 1,200% between July 2024 and February 2025. Those visitors showed 8% higher engagement, 12% more pages per visit and a 23% lower bounce rate, but were 9% less likely to convert than other traffic, an improvement from a 43% conversion gap in July 2024 [17].

The trajectory matters more than the level. AI-referred shoppers went from converting far worse than average to nearly at parity in seven months, which is consistent with AI assistants moving from novelty research tool to a genuine mid-funnel channel.

What this data does not yet show is true agentic buying, where the agent completes checkout itself. That behaviour is what the Visa, Mastercard and AP2 rails anticipate, and in mid-2026 it remains a small fraction of transactions.

Who is using browser agents like Comet, Atlas and Claude in Chrome?

Every major AI lab shipped a browser agent between mid and late 2025, but all launched gated, and the security data explains the caution. Anthropic measured a 23.6% prompt injection attack success rate against unmitigated browser agents.

Perplexity launched Comet on 9 July 2025, initially restricted to its $200-per-month Max subscribers, with an assistant that can summarise emails, manage tabs and navigate pages on the user's behalf [15]. OpenAI launched ChatGPT Atlas on 21 October 2025 on macOS, with agent mode in preview for paid tiers [14].

Anthropic took the most instructive route, piloting Claude for Chrome with just 1,000 Max plan users. Its red-teaming found browser use without safety mitigations showed a 23.6% attack success rate under deliberate prompt injection attacks, cut to 11.2% with mitigations. On a challenge set of four browser-specific attack types, new mitigations cut attack success from 35.7% to 0% [13].

Those numbers deserve more attention than the launches. A browser agent holds the user's logged-in sessions, so a malicious webpage that hijacks it can read email or move money. An 11.2% success rate against a determined attacker is a meaningful improvement and still far too high for unsupervised use, which is why every vendor ships with confirmation prompts and site restrictions.

No vendor has published reliable user numbers for its browser agent, so treat any circulating "market share" claims for Comet or Atlas with suspicion.

How big is the AI agent market supposed to get?

Projections cluster around 40 to 50% compound annual growth, which would take the market from single-digit billions today to somewhere north of $100 billion within a decade. All such figures are forecasts, and forecasts of new categories are routinely wrong in both directions.

Grand View Research valued the global AI agent market at $7.6 billion in 2025 and projects growth from $10.9 billion in 2026 to $182.9 billion by 2033, a 49.6% CAGR [16].

On the value side rather than the revenue side, Gartner forecasts that 33% of enterprise software applications will incorporate agentic AI by 2028, up from less than 1% in 2024, and that at least 15% of day-to-day work decisions will be made autonomously by then [1].

Market sizing firms define "AI agent" differently, and in a category where Gartner counts only ~130 genuinely agentic vendors, much of the revenue being sized today is arguably rebadged automation. Use these projections as a directional signal of investment, not a number to plan against.

Which agent predictions are coming true, and which are not?

The infrastructure predictions are landing; the autonomy predictions are running ahead of the evidence. Payment rails, protocols and browser agents all shipped on schedule, whilst measured task completion and enterprise scaling remain far behind the rhetoric.

Gartner's forward book is a useful scorecard. It predicts agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029, cutting operational costs by 30% [2]. Against that, current agents complete 30% of long-horizon workplace tasks [11] and about 35% of multi-turn CRM tasks [12], so the 2029 target requires substantial capability gains, not just deployment.

The pattern from 2025 to 2026 is that supply-side milestones, launches, protocols and partnerships, arrive on time because vendors control them. Demand-side milestones, scaled deployment and autonomous transaction volume, keep slipping because they depend on reliability that the benchmarks show is not there yet.

That is not an argument that agents are hype. It is the normal shape of a platform shift: rails first, trust second, volume third.

How to read these numbers

Adoption surveys are self-reported and definition-sensitive. "Using AI agents" can mean anything from a scaled autonomous workflow to one employee trying a demo, which is why McKinsey's 62% experimenting and 23% scaling figures can both be true [3].

Traffic figures come from each vendor's own network. Imperva's 51% bot share and Cloudflare's crawler breakdowns describe the traffic those companies happen to proxy, not the whole internet, and their bot classification methods differ [9][10].

Benchmark scores are not production success rates. They are measured under controlled scaffolding, on tasks chosen by researchers, and single figures hide huge variance between easy and long-horizon tasks [11][12].

Finally, every market size and every Gartner percentage-by-a-future-year figure in this article is a projection. Gartner publishing a 40% cancellation forecast alongside an 80%-by-2029 automation forecast is a reminder that even the same firm's projections point in different directions [1][2].

What this means going into 2027

Assume supervision, not autonomy. With multi-turn success rates around 35% on realistic tasks, the economic case for agents in 2027 is human-checked throughput, not headcount replacement [12].

Prepare for agents as customers. With Visa, Mastercard and AP2 rails live and AI-referred retail traffic converting near parity, retailers should test how their product data, pricing and checkout look to an agent, because the ledger of who gets recommended is being written now [7][8][17].

Expect a cull, then consolidation. If Gartner is even half right about 40% cancellations and pervasive agent washing, 2027 will separate a small set of genuine agent platforms from a large graveyard of rebranded chatbots [1].

Instrument for non-human traffic. With bots past half of web traffic and user-action AI fetches up 15x, analytics that only count human sessions now describe a minority of what touches your site [9][10].

Treat security as the gating factor. Prompt injection success rates in double digits are the single biggest brake on browser and shopping agents; the vendors that get those numbers to near zero will define the category [13]. If you want help working out how your brand and product data look to the agents already crawling and buying, you can book a call with me.