Agent Washing: The AI Vendor Problem No One Talks About

Gartner published an estimate that I had to read twice. Of the thousands of vendors currently marketing themselves as agentic AI providers, only approximately 130 are actually building what the term implies. Thousands of companies. One hundred and thirty real ones. That is not a rounding error. That is a structural problem in the AI vendor market, and it has a name: agent washing.

The term parallels greenwashing and AI washing in investment contexts. In each case, the mechanism is the same. A real concept with real value gets attached to products that do not embody it, because the label carries credibility that the product alone would not. Greenwashing attaches environmental claims to products with questionable environmental credentials. AI washing, which the SEC has already taken enforcement action against, involves describing investment strategies as AI-driven when the actual methodology relies on conventional screening. Agent washing is the same pattern applied to AI agents. A vendor takes an existing product, adds a chatbot interface or some scripted automation, calls it an "AI agent," and markets it to enterprise buyers who do not yet have a crisp enough definition of agentic AI to push back.

The distinction between a real AI agent and a rebranded chatbot or workflow automation tool is not just semantic. Agentic AI systems are distinguished by their capacity to pursue goals autonomously, decompose complex tasks, make intermediate decisions without human approval at each step, and adapt their approach based on outcomes. A system that answers customer service questions from a fixed script is not an agent in that sense. A system that monitors an infrastructure environment, decides which alerts require escalation, drafts incident reports, and initiates remediation steps within defined parameters is much closer. The difference matters for what you can actually do with the system, for what governance you need around it, and for what risks it introduces.

Gartner's broader agentic AI data makes the timing of this problem clear. Only 17% of organizations have deployed AI agents, but more than 60% expect to within the next two years. That means most organizations are in the evaluation and procurement phase right now. They are reading vendor materials, sitting through demos, and trying to decide which AI agent platforms to bet on. If thousands of vendors are misrepresenting what their products can do, and only about 130 are building genuine agentic capabilities, the organizations doing procurement right now are walking through a market where most of what they see is theater.

Gartner also projects that more than 40% of agentic AI projects will be canceled by the end of 2027. I think those two numbers are related. The cancellation rate is not just about technology immaturity or organizational readiness. It is partly about organizations purchasing products that were never going to deliver on the agentic promise because the products were not actually agentic. When the pilot fails to produce autonomous, goal-directed behavior, the organization concludes that agentic AI does not work for their use case. Sometimes that conclusion is correct. Sometimes the conclusion should be that the vendor delivered a chatbot with an agent label on it.

The procurement problem is real, and it is harder to solve than it might seem. Enterprise buyers evaluating AI agent vendors face several compounding challenges. First, the underlying technology is genuinely complex, and most procurement teams do not have deep technical staff who can assess whether a system is architecturally capable of the autonomy and goal-directedness that the marketing claims. Second, vendor demos are designed to impress, and an impressive demo does not distinguish between a genuine agent and a well-scripted automation. The demo runs in a controlled environment with prepared inputs, and the limitations of scripted automation do not surface until the system encounters the messiness of real enterprise data and edge-case tasks. Third, buyers are under pressure to move. With 60% of organizations planning agent deployments in the next two years, procurement teams feel that falling behind creates competitive risk. That time pressure works against careful evaluation.

What would careful evaluation look like? I think it starts with asking about the architecture before looking at the demo. Specifically: how does the system decompose goals it has not seen before? What happens when it encounters a task that falls outside its training distribution? How does it handle ambiguous instructions? What does its failure mode look like, and who gets notified when it fails? A vendor with a genuine agentic system should be able to answer all of these questions in concrete, technical terms. A vendor that deflects to the demo or to capability claims without architectural specifics is a warning sign.

The ROI question also cuts differently for agent washing. Organizations implementing what they believe to be agentic AI are making investment decisions based on the productivity and capability claims that real agentic systems could plausibly deliver. When the deployed system turns out to be sophisticated automation rather than genuine agency, those returns do not materialize, not because the underlying idea was wrong but because the product was not what it claimed to be. This is how AI washing in investment creates reputational damage for legitimate AI approaches: the failed expectations from the washed version contaminate the perception of the real thing.

The IS research angle here is governance. How should organizations build procurement governance for a technology category where the boundary between "real" and "washed" is technically complex and actively obscured by vendor marketing? The standard make-or-buy analysis, capability assessment, and reference check process is not calibrated for this environment. I think there is a genuine research question about what governance mechanisms allow organizations to accurately classify AI vendor capabilities and make procurement decisions that match what the technology can actually deliver. The variation is going to be enormous, and the organizations that get procurement right will have systematically different AI outcomes from those that do not.

For now, the safest heuristic I can offer is this: if a vendor cannot explain how their agent handles a goal-directed task it has never encountered in a demo, in concrete architectural terms rather than in marketing language, it is probably not an agent in any meaningful sense.