Tonight in Palo Alto, ten teams will put their AI agents on stage at the AI Agent Arena and let a room full of builders vote on them. Before the first demo, I get five minutes to answer a simple question: where are AI agents, really, at the end of 2026? This is the long version of that talk, with the numbers and the sources behind them.
The short answer: almost every company is trying agents, very few have scaled them, and the gap between the two is not about model quality any more. It is about trust, permissions and measuring the right thing.
From answering to doing
It helps to look at the last four years as a steady handover of work from people to software.
- 2023, chatbots. AI answered questions. You asked, it replied, you did the work.
- 2024, copilots. AI suggested the next step inside the tools you already used: the next line of code, the next sentence of an email. You still pressed the button.
- 2025, agents. AI started taking actions through tools: searching, filing tickets, calling APIs. Mostly short tasks, mostly watched closely.
- 2026, coworkers. Agents now run for minutes or hours on their own, use the same screens and apps we do, and hand back finished work for a person to check.
Each step moved the line between "AI helps" and "AI does" a little further. The question for 2027 is not whether agents can act. It is which actions we let them take without asking.
Adoption: almost everyone is trying
The survey numbers this year are striking. Zapier's 2026 AI Agents survey of 513 enterprise leaders found that 88% say their organizations are using or testing AI agents, up from 72% a year earlier. Seventy-one percent say agents are already in production, and 90% expect to increase their agent spending over the next year.
CrewAI's survey of 500 executives at companies with more than $100 million in revenue tells a similar story: 65% already use AI agents.
A word of caution before we celebrate. Both surveys come from companies that sell agent software, and the definition of "using an agent" is generous. KPMG's AI Quarterly Pulse found that about 26% of organizations were actively using AI agents by the end of 2025, up from 11% at the start of that year. That is still fast growth, but it is a very different headline. The truth is probably that nearly everyone is experimenting and a meaningful minority is relying on agents for real work.
The gap: trying is easy, scaling is not
This is the part of the story that gets less airtime. A Gartner survey of 1,303 enterprise respondents, reported by TechTarget, found that only 22% have successfully scaled AI across multiple business units. TechTarget also reports that profitability gains from AI have stayed flat across sectors. And CrewAI's own data shows the average enterprise has automated only about 31% of its workflows.
The number that still frames the year is Gartner's prediction from June 2025: more than 40% of agentic AI projects will be canceled by the end of 2027, because of rising costs, unclear business value or weak risk controls. Gartner also called out "agent washing", vendors relabeling chatbots, assistants and old robotic process automation as agents. Of the thousands of vendors claiming agentic products, Gartner estimated only about 130 offer real agent capabilities.
What changed in 2026
Four shifts made this the year agents stopped being a demo category.
Where agents earn their keep today
Across the companies getting real value, the same five jobs come up again and again:
- Software. Writing, testing and fixing code, and reviewing pull requests.
- Customer support. Resolving common tickets end to end and handing the rest to a person with the context already gathered.
- Back office. Invoices, data entry, reports and reconciliations.
- Scheduling and admin. Finding times, booking, moving meetings and following up, the work we focus on at ANCI.
- Research. Reading widely, comparing options and summarizing with sources.
What these jobs share is more important than the list. Each has clear inputs, a clear finish line, and a person who can check the result quickly. That combination is what makes an agent safe to hand real work.
What still breaks
If you are building an agent, these are the four problems that decide whether you ship or join Gartner's 40%.
- Reliability over long tasks. Small errors multiply. An agent that is 95% right at each step is only about 77% right across five chained steps. A wrong guess early in a task quietly poisons everything after it.
- Permissions. There is still no shared standard across MCP, A2A and the other agent frameworks for what an agent is allowed to do, and that matters most for actions you can't undo: sending an email, paying an invoice, booking a meeting, deleting a file.
- Hidden clean-up. Time spent checking and fixing agent output eats into the savings. One study this month reported that engineers spend more than 17 working days a year fixing AI-generated code.
- Agent washing. Buyers are tired of "agents" that turn out to be chatbots with a new label. Every real agent pays a trust tax for every fake one.
Security is the shadow over all four. Gartner predicts that by 2028, 25% of enterprise security breaches will trace back to AI agent abuse. An agent with broad permissions is a new kind of attack surface.
The pattern behind agents that ship
The teams moving from pilot to production share a pattern. It is also how we build ANCI, an AI scheduling agent that works for executive assistants: the agent does the work, and the person stays in charge.
Five questions to ask every agent demo
Whether you are voting tonight or evaluating a vendor next week, these five questions separate a great demo from an agent you would trust with your own work:
- What real job does it finish, start to end?
- What happens when it gets something wrong?
- Who approves before it acts?
- Does it work in the tools I already use?
- How much time would it save me this week?
The flashiest demo is rarely the right answer. The right one is the agent whose failure mode you understand and whose savings you can count.
The state of AI agents, in one line
Agents are ready for real work, when people stay in charge. The technology crossed the line this year: agents can use computers, run for hours and plug into almost anything. What has not caught up is the boring part, which is permissions, measurement and earning trust one approved action at a time. The companies that get that part right in 2027 will be the 22% everyone else is trying to catch.
See you at the Arena.
References
- Zapier, "AI Agents Survey 2026" (n=513 enterprise leaders, fielded by Centiment). zapier.com/blog/ai-agents-survey
- CrewAI, "The State of Agentic AI in 2026" (500 executives at $100M+ revenue companies). crewai.com/blog/the-state-of-agentic-ai-in-2026
- KPMG, "AI Quarterly Pulse Survey". kpmg.com/us/en/articles/2025/ai-quarterly-pulse-survey
- TechTarget, "Deep dive: the state of AI adoption in 2026" (Gartner survey, n=1,303). techtarget.com
- Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", June 25, 2025. gartner.com
- AI Agents Directory, "AI agents evolve from workflows to desktop control and enhanced security", November 2, 2026 (computer use, AI-generated code clean-up study, Gartner 2028 security prediction). aiagentsdirectory.com
- RSM, analysis of permission standards across MCP, A2A and agent frameworks. rsm.global
- Anthropic, "The 2026 State of AI Agents Report". claude.com
Further reading: meet the 10 teams on stage at the AI Agent Arena, why we took the LLM out of our scheduling agent's control loop, and why 95% isn't good enough. We build ANCI, the AI scheduling agent for executive assistants.