AI agents: what is real and what is hype

admin
AI agents: what is real and what is hype cover graphic

“Agent” is now attached to almost every AI product. Some of these tools really do plan steps, use other software and finish tasks with little supervision. Others are chatbots with a new label. For a buyer, the hard part is telling them apart and judging how much of the adoption talk is evidence and how much is marketing.

This article collects what dated, public sources said about AI agents in 2026, with attention to who ran each survey and what they sell. It does not rank products. We did not run any agent ourselves for this piece, and the October news section relies on a news aggregator whose items we could not check against original sources.

Key takeaways

  • Surveys disagree sharply on adoption. A Mayfield survey of 266 senior technology leaders found 42 percent with agentic AI in production, while a vendor survey from CrewAI reported 65 percent of 500 executives using agents.
  • Gartner said in June 2025 that over 40 percent of agentic AI projects would be cancelled by the end of 2027 because of costs and unclear value, and warned about “agent washing”. That was a forecast, not a measured result.
  • Governance lags behind. Mayfield reported that 60 percent had early-stage or no formal AI governance, and Deloitte’s report was described as finding only a small minority with mature agent governance.
  • Many October 2026 headlines concern long-running agents, sandboxes and security reports, but we could not verify them against original sources.
  • The sensible buyer move is a narrow pilot with human review, not a broad rollout.

Summary

The evidence on AI agents is mixed in a particular way. Almost every survey we read finds strong interest and rising budgets. Fewer give hard evidence of measured savings, and the ones with the largest adoption numbers tend to come from companies that sell agent software. Analysts and the surveys themselves also point to the same blockers: data quality, integration and governance.

So the honest summary is that agents are real enough to pilot and not yet proven enough to trust unsupervised. Our explainer on AI agents for non-technical readers covers the concepts, and this article focuses on the evidence.

What people mean by agent

A chatbot answers when you ask. An agent is software that takes a goal, decides on steps, uses tools such as email, calendars, databases or browsers, and acts. The line is blurry, and vendors use the word loosely. Gartner’s concern, reported by Reuters in June 2025, is that vendors rebrand assistants and chatbots as agentic without meaningful agentic capability. Gartner called this agent washing and estimated that only about 130 of the thousands of agentic AI vendors are real. We could not check that estimate independently. For help with the choice, read how to choose between a chatbot and an AI agent.

What the surveys say

The table lists the 2026 studies we found, with who ran each one. Compare the sample and the sponsor, not just the headline percentage.

Study Date reported Sample Headline finding Sponsor note
Mayfield, The Agentic Enterprise in 2026 19 January 2026 266 CIOs, CTOs, CAIOs, CISOs and CDOs at large firms 42% have agentic AI in production, 72% including pilots Venture firm survey of its own network
Deloitte, State of AI in the Enterprise 2026 (via Techstrong.ai) 27 January 2026 More than 3,000 director to C-suite leaders in 24 countries Nearly three-quarters plan to deploy autonomous agents within two years Reported secondhand
CrewAI, 2026 State of Agentic AI (via Digital Commerce 360) 11 February 2026 500 executives at firms with over $100 million revenue and 5,000 or more staff 65% say their enterprises already use agents CrewAI sells agent software
OutSystems, 2026 State of AI Development (via iTWire) 13 April 2026 1,900 global IT leaders 96% use agents in some capacity, 94% worry about sprawl OutSystems sells development software

How to read the gaps

The 42 percent and 65 percent figures are not contradictory once you note the different samples and wording, but they show how definitions drive results. “Using agents in some capacity” is a low bar. “In production” is higher. The Digital Commerce 360 article itself notes the CrewAI results are self-reported, vendor-sponsored and lack full methodology.

The blockers are consistent

Mayfield reported that 58 percent named data readiness or quality as the top obstacle, and that 84 percent call security and compliance non-negotiable even though 60 percent have an early-stage or no formal governance framework. CrewAI’s respondents named data readiness and integration (35 percent) and skills (33 percent) as top barriers. OutSystems reported that 94 percent worry about sprawl, and only 12 percent have a centralised platform to manage it. Mayfield also found 70 percent want to test tools in their own environment before committing.

The sceptical forecast

Reuters, via The Star, reported on 25 June 2025 that Gartner expected over 40 percent of agentic AI projects to be cancelled by the end of 2027 because of escalating costs and unclear business value. A Gartner analyst was quoted as saying most agentic AI propositions lack significant value or return on investment. The same report gave Gartner forecasts for 2028: at least 15 percent of day-to-day work decisions made autonomously, and 33 percent of enterprise software applications including agentic AI. These are predictions made in 2025. We found no 2026 measurement that confirms or refutes them, and we are not claiming either outcome.

October 2026 headlines

An aggregator page called AI Agent News lists many October 2026 items. It gives no item-level source links, notes that many entries rely on secondhand reporting, and its text was truncated. We therefore treat everything below as reported, not verified. Check original sources before acting on any item.

Date (October 2026) Item as reported by the aggregator Type
9 October Anthropic published an internal report on unintended model actions on real websites Safety report
14 October Workday positioned its Sana assistant as the front door for HR and finance agents Product
16 to 17 October India’s CERT-In reported AI agents reaching live systems at Indian organisations, with exposed credentials in code and configuration as root causes Security report
20 October Notion rebuilt its agents to run tasks for more than 24 hours Product
24 October The Agentic AI Foundation added four projects, including Agent Skills from Anthropic and Open Responses from OpenAI Standards
26 October Microsoft described a long-running agent called Autopilot Product

Two themes stand out even with that caveat: agents that run for hours or days, and security incidents tied to agents holding credentials. Both are practical risks for a buyer. Under the CERT-In item, the reported root cause was ordinary credential hygiene, not exotic AI behaviour.

What looks real

  • Budgets and pilots are growing. Mayfield found 91 percent plan to increase agentic AI budgets in 2026.
  • Developer productivity is a leading use case. Mayfield reported 70 percent rank it in their top three priorities.
  • Mixed build and buy is common. Mayfield found about 65 percent mix in-house builds with vendor tools.
  • Governance and data readiness are recognised problems across all the sources.

What looks like hype

  • Unqualified adoption percentages from vendors, especially without sample, wording or field dates.
  • Claims of “fully autonomous” work with no description of human review or error handling.
  • Savings figures that are self-reported, such as CrewAI’s time savings, which are perceptions rather than measured results, as the article itself cautions.
  • The agent label applied to any chatbot.

What it means for tool buyers

Ask what the agent can actually touch

List the systems an agent could read or change: email, files, payment tools, customer records. Give it the least access that does the job. Agents with broad credentials turn small mistakes into big ones. Read how to protect company data when using AI tools before connecting any agent to real data.

Plan for errors

Agents inherit the weaknesses of the models behind them. AI hallucinations and how to reduce them explains why a confident wrong step can compound across a multi-step task. Require review for anything customers see or money touches.

Write down the rules

Mayfield’s finding that 60 percent have early-stage or no governance is a reminder to set rules early. Writing an AI use policy for your team is a start.

How to run a small pilot

  1. Pick one narrow, repetitive task with a clear definition of done.
  2. Run the agent beside your current process for two weeks, with a person approving outputs.
  3. Record time spent, errors caught and cost, so you have your own measured numbers.
  4. Check logs: can you see what it did and why?
  5. Decide in advance what result would make you expand, and what would make you stop.

Our guide on how to test an AI tool before you pay for it gives more detail.

What to watch next

  • Independent measurements of agent success rates and costs, rather than self-reported surveys.
  • Whether Gartner’s 2027 cancellation forecast looks right as projects mature. We make no prediction.
  • Standards work such as the foundation projects reported on 24 October, once confirmed by primary sources.
  • Security guidance and incident reports involving agent credentials.

Frequently asked questions

Are AI agents real or just marketing?

Both. Surveys report real pilots and production use, but the label is also applied to ordinary chatbots, and several of the loudest adoption numbers come from vendors.

What is agent washing?

Gartner’s term, reported in June 2025, for vendors rebranding assistants and chatbots as agentic without meaningful agentic capability.

Will most agent projects fail?

Gartner forecast in 2025 that over 40 percent would be cancelled by the end of 2027. That is a forecast. We found no 2026 data that settles it.

Can I trust the survey percentages?

Check who ran the survey, the sample and the wording. Vendor surveys can be useful but may overstate adoption, and the article we read on CrewAI’s survey notes missing methodology.

What is a safe first agent project?

A narrow, reversible task with human approval and minimal access, such as sorting inbound support emails into categories.

Conclusion

Agents are worth a careful look, not a leap. Use evidence from your own pilot over any vendor’s adoption statistic. If you are new to the idea, start with AI agents explained, then decide whether you need an agent or a chatbot with this decision guide.

Sources

Keep reading