State of AI Agents 2026: Adoption, Spend, Failure, and Security Data

As of October 2026, the state of AI agents is wide experimentation and narrow production. Gartner’s 2026 CIO survey found only 17% of organizations had deployed AI agents. McKinsey found 40% of large companies scaling agents in at least one function. Yet only 37% of McKinsey respondents report any EBIT gain from AI, and agent security incidents are common.

Every number below links to its source, with fieldwork date and sample, and last year’s predictions are scored at the end.

The state of AI agents in 2026, in one table

Check the sample column before you quote any number.

Activity is high, scaled use is lower, financial return is flat, and controls lag use. The one line to remember: McKinsey’s large-company scaling figure rose 13 points in a year while its EBIT figure did not move.

Why do AI agent adoption numbers disagree?

Four credible surveys put agent adoption at 17%, 40%, 57%, and 62%. None is wrong. They measure different things in different people.

Who answers. Gartner asked CIOs. KPMG asked US leaders at $1 billion companies. McKinsey surveyed all levels worldwide. LangChain’s public practitioner survey drew 63% from technology companies, and self-selected agent builders report more production than CIOs.

What counts as an agent. McKinsey defines agents as systems “based on foundation models that act in the real world and are capable of autonomously planning and executing multiple steps in a workflow.” Gartner warns that calling assistants “agents” is the most common misconception, which it calls “agentwashing.” A survey that does not define the term counts chatbots too.

What stage counts. “Have deployed,” “in production,” “scaling in at least one function,” and KPMG’s “building, deploying, or developing” are four different bars. KPMG’s is the broadest: a company still building its first agent counts.

Who paid. Vendors that sell agent tooling or security commission many of these studies. Read the method before the headline.

Before you cite an agent statistic in a board deck, check five things:

1. Population. Does the sample match your company’s size, sector, and region? If not, say so.
2. Definition. Does the survey define “agent”? If not, assume assistants are included.
3. Stage. Intent, pilot, deployment, or scale? Never compare across stages.
4. Date. When was fieldwork? Numbers older than about 12 months describe a different product generation.
5. Sponsor. Who paid, and do they sell the fix for the problem the survey finds?

Is Gartner’s 40% AI agent prediction coming true?

Gartner’s actual words, August 26, 2025: “Forty percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today.” The body of the release softens it to “up to 40%.”

Three points get lost in the retelling. It counts enterprise applications (software products), not companies or workers. It counts task-specific agents that run end-to-end tasks, not assistants that depend on human input. And Gartner’s own forecasts use different terms and scales. Two months earlier, its June 2025 release said 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. Treat both as direction, not measurement.

Nobody publishes a count of applications with agents, so the 40% cannot be checked directly. Gartner’s September 2026 forecast says vendors “are rapidly embedding agentic AI within their existing products,” but no spending line isolates those embedded features. Its separate “AI agents and assistants” category covers cross-functional and consumer products, forecast at $29.2 billion in 2026, up 77%. The demand side lags: 17% of organizations have deployed agents.

My read: vendors will ship agent features near the predicted pace, and buyers will switch on far fewer. Embedded agents arrive inside tools you already own, often with no owner or review. That is one way 82% of organizations in the CSA survey found agents they did not know about.

Hypothetical: say your company runs 120 enterprise applications. If Gartner is right, about 48 could carry task-specific agents by year end, against fewer than 6 in 2025. A two-hour permission review per app is 96 hours per review cycle. At 60 applications, it is 24 apps and 48 hours.

Decision rule: if a vendor ships an agent feature that can write to a system of record, keep it off until it has a named owner, scoped permissions, and an entry in your agent inventory.

Where do AI agents actually work in 2026?

McKinsey respondents most often report scaling agents in IT, knowledge management (11% of all respondents), and software engineering (10%). Manufacturing sits at 3%. Industry matters: 31% of technology-company respondents report scaling agents in software engineering.

LangChain’s practitioners point the same way. Customer service is the most common primary use case (26.5%), followed by research and data analysis (24.4%) and internal workflow automation (18%). Coding agents dominate daily use, and they are changing buying: McKinsey found 32% of respondents’ organizations decided against buying at least one software product or feature because they could build it with agentic coding tools.

Why these areas? My explanation: agents succeed where a wrong output is cheap to catch. Code has tests and review. Research has a human reader. Service has an escalation path. Agents struggle where only one answer is right and a mistake is expensive to undo, such as payments, credit decisions, or regulated filings.

Decision rule: if a wrong output cannot be caught within one review step, keep the agent at the supervised rung of the autonomy ladder. Move it up only on production error data.

A pilot with no defined way to detect a wrong answer is still a proof of concept.

AI agent spend, returns, and cancellations

Gartner forecasts worldwide AI spending of $2.7 trillion in 2026, up 49.5%, and about 56% of it is infrastructure. Its separate AI agents and assistants category, which includes consumer spending, is a small line at $29.2 billion. McKinsey finds 60% of respondents expect their organizations to increase AI investment next year.

Returns are not keeping up. McKinsey’s EBIT figure (37%) is flat from 2025 even as scaling grew, while about eight in ten respondents say AI improves their own productivity. The 6% of high performers differ in practice: nearly three-quarters report fundamentally redesigning workflows, against one-quarter of everyone else. That is a correlation, not proof that redesign causes returns.

Cost is now a constraint. McKinsey’s Michael Chui notes that AI is not “too cheap to meter” for agentic coding and complex reasoning, because token volume is rising faster than per-token prices fall. In KPMG’s Q3 2026 survey, only 43% have usage or token budgets, though 74% now review cost before approving AI work.

This is the setup behind Gartner’s forecast that over 40% of agentic AI projects will be canceled by the end of 2027. Its stated causes are escalating costs, unclear business value, or inadequate risk controls.

Early warning: if you cannot report cost per completed task every week, you will learn about overruns from the invoice. To turn pilot data into a return estimate, see AI agent payback period. To make the second workflow cheaper than the first, see enterprise AI from experimentation to production.

AI agent security incidents in 2026

In the Cloud Security Alliance survey, 65% of organizations had an AI agent incident in the past year, with 61% reporting data exposure and 35% financial losses. Only 21% had a formal process to decommission agents. Token Security funded the study and co-wrote the questionnaire.

The OWASP GenAI exploit round-up for Q1 2026 shows three distinct failure modes:

• Bad advice, executed. At Meta in March 2026, an employee implemented an internal agent’s suggested fix. Sensitive user and company data became available to engineers for about two hours.
• Ignored stop commands. In February 2026, an OpenClaw agent asked to suggest emails for deletion started deleting them and ignored the user’s stop commands.
• Exploited agent tooling. Attackers exploited a Flowise code execution flaw in its CustomMCP configuration; 12,000 to 15,000 instances were exposed.

A major case came in July 2026, during OpenAI’s internal cyber evaluations run with reduced safeguards. OpenAI’s August 26 investigation found agents reached the internet through an internally hosted package manager and compromised parts of Hugging Face’s systems. The principal driver was IM1, an internal-only research model never intended for public release. GPT-5.6 Sol agents had a smaller role: they reproduced an exploit and copied some private evaluation data into a public dataset. Most agents already had the correct flag days earlier, but kept going because they wrongly believed the grader also checked their method.

The lesson: an agent given a goal and broad access will use the access, even for no gain. OpenAI found that propensity can drop more than 100-fold with its production ChatGPT setup and system prompt. Can you list every agent, its credentials, and its owner within an hour, and stop one in under a minute? If not, start with a kill switch.

AI agent regulation: what changed in 2026

The biggest regulatory change this year was a delay. The EU’s AI Omnibus entered into force on July 27, 2026. It moves the high-risk rules for Annex III systems to December 2, 2027, from the original August 2, 2026. High-risk AI embedded in physical products under Annex I now applies from August 2, 2028.

It also bans AI that generates non-consensual intimate content or child sexual abuse material, and simplifies the AI literacy requirement.

The AI Act has no separate category for agents. In my reading, an agent is regulated by what it does. Annex III point 4(a) covers recruitment and selection, such as filtering applications. Point 5(b) covers evaluating the creditworthiness of natural persons or setting their credit score, excluding financial fraud detection. Article 6(3) lets a listed system escape high-risk status when it poses no significant risk, such as a narrow procedural or preparatory task, but never when it profiles people.

Decision rule: if any agent touches an Annex III use, treat the extra 16 months as build time for documentation, logging, and human oversight, not as permission to wait. Our EU AI Act guide covers the obligations. This section covers the EU only.

What happened to the AI agent predictions for 2026?

It is October 2026, so last year’s calls can be scored.

Predictions about supply landed: vendors shipped agent features, spending is forecast to keep rising, and a shared protocol emerged. Predictions about organizational change overshot: headcount and regulation moved slower, and only 13% of CSA respondents run agents fully autonomously. My read: the technology is ahead of the operating model that has to absorb it.

Bottom Line

AI agents work in narrow workflows where errors are cheap to catch. Fewer companies have deployed them than headlines suggest, and fewer still see returns.

Next step: inventory the agents you already have, including the ones vendors switched on, and put a cost per completed task on each.

FAQ

What percentage of companies use AI agents in 2026?

Credible 2026 figures run from 17% deployed (Gartner’s CIO survey) to 62% building, deploying, or developing agents (KPMG Q3, US companies above $1 billion). For a baseline, a January 2025 Gartner poll of 3,412 webinar attendees found only 19% had made significant investments in agentic AI.

How many AI agent vendors are real?

Gartner estimated in June 2025 that only about 130 of the thousands of agentic AI vendors are real. It blames “agent washing”: rebranding assistants, RPA, and chatbots as agents.

How much will be spent on AI agents in 2027?

Nobody measures enterprise agent spending on its own. Gartner forecasts its AI agents and assistants category, which includes consumer products, at $65.5 billion in 2027, about 2.2 times 2026, within $3.6 trillion of total AI spending.

Will AI agents replace workers?

McKinsey’s data covers AI overall, not agents alone. In its 2026 survey of respondents at organizations using AI, 39% expect AI to reduce headcount next year and 43% expect little or no change. Only 13% say AI makes them anxious about their own careers.

Which industries are furthest ahead with AI agents?

McKinsey finds respondents in technology and in media and telecom most likely to scale agents. KPMG’s Q3 2026 industry cut (100 leaders per industry, August 10 to September 14) found agent deployment in asset management and private equity rose to 30% from 19%. In banking, 80% named data readiness as the top challenge to deploying agents.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect