Agent Washing and Why the AI Vendor Market Is a Research Problem

Gartner made a striking claim recently: of the thousands of vendors currently marketing their products as "agentic AI," only around 130 are offering something that is genuinely agentic. The rest are rebadging existing tools. Gartner called this "agent washing," and their 2027 prediction follows logically: more than 40% of agentic AI projects will be canceled by end of 2027. That is a lot of canceled projects in a market projected to reach $376.3 billion in the same year.

I have been sitting with the agent washing claim because it is not really a surprise from an IS theory standpoint, but it is a significant practical problem for how the field studies this phenomenon. The term "agentic AI" is doing enormous definitional work right now, and most of it is not definitional at all. It is marketing.

DiMaggio and Powell (1983) described what they called mimetic isomorphism: under conditions of uncertainty, organizations copy perceived leaders and adopt the same forms, labels, and structures, not because of independent analysis but because conformity to the institutional category signals legitimacy. The same pressure applies to vendors. When "agentic AI" became the dominant category in the enterprise software market in late 2024 and early 2025, vendors faced a simple choice: claim the category or risk appearing behind the curve. Most chose to claim it. The technology did not change. The label did. This is precisely what isomorphism predicts, and it is exactly what happened with "cloud" in 2010 and "big data" in 2013. The market created a category with high legitimacy value, and existing products migrated into it without fundamentally changing.

Scott (1995) would point to the cultural-cognitive pillar here. Once "agentic AI" became the taken-for-granted baseline for what enterprise software vendors are supposed to offer, the question shifted from "is this actually agentic?" to "how do we position our product as agentic?" That shift is not cynical calculation in most cases. It is institutional pressure at work. When the category is what generates resources and attention, you adapt to fit the category. The 130 vendors that Gartner considers genuinely agentic are the ones whose technology actually changed. The rest are responding rationally to institutional pressure in a market where the category label matters more than the technical substance behind it.

This is familiar enough from prior technology cycles that I should not be surprised. And yet I keep noticing that IS researchers studying "agentic AI adoption" have a real methodological problem on their hands. If the unit of analysis is "organizations that have deployed AI agents," and if most of what gets labeled an AI agent is actually a triggered automation or a scripted workflow with a chatbot interface, then studies that survey organizations about their agent deployments are not measuring a coherent phenomenon. They are measuring the adoption of a category label by organizations that may have deployed very different kinds of technology.

Gartner's own data illustrates the gap. Only 17% of organizations have actually deployed agents as of early 2026, but more than 60% expect to within two years. That 60% expectation sits alongside the prediction that 40% of agentic projects will be canceled by 2027. My read is that the 60% expectation reflects institutional pressure, organizations feeling they should be deploying agents because the category has become strategically mandatory, while the 40% cancellation rate reflects the organizational reality that most of what gets labeled agentic does not deliver what was promised in procurement conversations. The gap between mimetic adoption and realized value is where the canceled projects will come from.

From a research standpoint, this makes it very hard to draw conclusions from studies that do not carefully define what "agentic" means in the specific organizational context being studied. The difference between a genuinely autonomous agent that can receive a high-level goal, decompose it into tasks, call external tools, monitor its own output, and report back, and a workflow automation tool that routes tickets based on keyword matching, is not a minor operationalization detail. It is a fundamental distinction that determines what the technology can plausibly do, what governance structures it requires, and what organizational changes are needed to exploit it. Treating both as "AI agent deployment" in the same study is a construct validity problem, and it is one the field will struggle with as long as the vendor market is as definitionally promiscuous as it currently is.

I wrote about this kind of definitional slippage before when discussing how institutional isomorphism shapes AI adoption. The coercive, mimetic, and normative pressures that push organizations toward AI adoption also push vendors toward AI labeling. Both processes are rational responses to institutional pressure. Both produce outcomes that look like adoption and capability without necessarily being adoption and capability. The difference is that vendor isomorphism creates a measurement problem for researchers on top of the strategy problem for buyers.

What would help is a taxonomy that IS researchers could use consistently, one that distinguishes between triggered automation (if-then logic), supervised AI assistance (human in the loop for all non-trivial decisions), and genuinely agentic systems (capable of goal decomposition, tool use, and autonomous task execution with limited human intervention). Gartner's 130-vendor estimate suggests such a taxonomy would map most of what is currently called agentic onto the first two categories. That is not a criticism of the organizations buying these tools. The first two categories can be valuable. It is a call for the research community to be precise about what it is studying before drawing conclusions about what "agentic AI" does or does not produce.

The agent washing problem will not resolve itself. As long as "agentic" carries institutional value, vendors will claim it. The question is whether IS researchers can build the methodological infrastructure to cut through the label and study the phenomenon that actually matters: what happens when organizations deploy systems with varying degrees of autonomy, and what organizational capabilities determine whether that deployment produces value.