Ask ten vendors this year whether their product is an AI agent, and eleven will say yes. Gartner actually sat down and counted. In a June 25, 2025 press release, the research and advisory firm estimated that of the thousands of vendors now marketing agentic AI, about 130 are the real thing. The rest are doing what Gartner calls agent washing: taking a chatbot, a workflow tool, or a robotic process automation script that already existed, changing the label, and shipping it as autonomous.

The same release carries a harder number, the one that should worry anyone about to sign a contract. Gartner predicts that more than 40 percent of agentic AI projects will be canceled by the end of 2027, done in by rising costs, unclear business value, and weak risk controls. Anushree Verma, the Gartner analyst quoted in the release, put it without much cushioning: "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." Hype is not really a technology problem. It is a labeling problem, and labeling problems are the kind you can catch before you sign anything, provided you know which questions to ask.

What agent washing actually looks like

Gartner's definition of the practice is unambiguous, which is rare for an industry that treats the word agentic as a mood rather than a specification. Vendors, in the firm's words, are "rebranding existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities" and calling the result an agent. The tell is almost always in the demo. A chatbot wearing an agent costume can hold a conversation and retrieve a document, then hand the result back to a human for the next step. That is genuinely useful software. It is also not agentic. It is a very polite lookup tool.

A real agent does something else. It plans a sequence of steps toward a goal you gave it, picks which tools to call and in what order on its own, and keeps working with limited supervision when the first attempt does not land. This outlet laid out the fuller version of that definition in its June 19, 2026 explainer on agentic AI, so the short version will do here: plans, acts, adjusts, and does not wait for a human to click next between every step. If a product needs someone to approve each individual action before it happens, that product is an assistant with good manners. It is not an agent, no matter what the pricing page calls it.

Same industry, same day, opposite directions

Here is a small coincidence worth sitting with. On July 9, 2026, Meta launched Muse Spark 1.1, a paid model it says handles multistep reasoning, manages digital workflows, and uses tools with limited hand holding, priced at $1.25 per million input tokens and $4.25 per million output tokens, according to Meta's announcement and TechCrunch's coverage that same day. On that identical day, OpenAI told TechCrunch it was shutting down Atlas, the standalone AI browser it had launched only in October 2025, folding whatever it did into a Chrome extension, a heavier desktop app, and a cloud agent that runs tasks remotely. One company spent the day betting a subscription price on autonomy holding up under real use. The other spent the day admitting that a browser with a chat window taped to the side was never quite that, and pulled it before its first birthday. Neither event is the Gartner statistic. Both are the weather the statistic was describing.

The buyer's checklist

Skip the roadmap slide. None of the following needs the vendor's five-year vision, and all of it can be answered in the first meeting by anyone who actually built the product, as opposed to anyone who sells it.

  • What actions can it take without a human approving each one, specifically, not in general.
  • What happens when a step fails. Does it retry, escalate, roll back, or just stop with the task half finished.
  • Where are the audit logs, and can you read them yourself without filing a support ticket.
  • What does it cost per completed task, not per seat, not per token, not per query that may or may not finish the job.
  • What systems and data can it reach, and who signed off on that access list.
  • What happens the day the underlying model changes. Does the vendor retest the workflow, or do you find out when it breaks.

The math for 2028

Gartner is not only predicting a wave of cancellations. The same research arm projects that 15 percent of day to day work decisions will be made autonomously through agentic AI by 2028, up from zero percent in 2024, and that roughly a third of enterprise software applications will carry genuine agentic features by that year, up from under 1 percent in 2024. Both numbers assume the market finishes sorting out which vendors, 130 or however many by then, are worth the contract. That sorting is not happening in a research report. It is happening in procurement meetings, one unanswered question about audit logs at a time.

This is general analysis, not procurement advice. Gartner's June 2025 press release is the primary source for the cancellation forecast; every other figure is sourced inline above. Questions go through our contact page.