Agent Washing: Spotting a Real AI Agent From a Fake

There is a word for what is happening in your inbox right now, and Gartner coined it: agent washing. It is the practice of slapping "AI agent" on a product that is really a scripted chatbot, a rules-based workflow or last year's RPA bot with a new label. The demand for agents is real, so the marketing raced ahead of the engineering. Gartner's own estimate is blunt: of the thousands of vendors claiming agentic capabilities, only around 130 are actually delivering them. The rest are selling you a costume.

This matters because the gap is expensive. Gartner expects over 40% of agentic AI projects to be canceled by the end of 2027, and a big share of those failures start at procurement, when a buyer pays agent prices for chatbot capability and only finds out in production. You do not need to become an AI researcher to avoid that. You need a working definition and a handful of questions that a fake cannot answer well.

What actually makes something an agent

Strip away the marketing and an AI agent has one defining property: agency. It can take a goal, reason about how to reach it, choose actions, use tools, and adapt when the situation changes, without a human scripting each step. That is the line.

A chatbot answers. An agent acts. A chatbot follows a decision tree someone drew in advance, so it can only handle the branches that were drawn. An agent decides what to do next based on the state it finds, which means it can handle situations nobody anticipated, and, importantly, can also fail in ways nobody anticipated. If a product only does exactly what its flowchart allows, it is automation with good copywriting. That is fine, plenty of problems want a reliable script, but it is not an agent and should not cost like one.

The questions that expose a fake

Vendors have polished demos. Demos are the easy path, the vendor picks the input. Your job is to ask what happens off the happy path, because that is where washing shows.

• "Show me a task the system completed that you did not explicitly script." A real agent has examples of handling novel situations. A washed one will pivot to describing its "flows", which are scripts by another name.
• "What tools can it call, and how does it decide which one?" Genuine agents use tools, they call APIs, query systems, take actions, and choose between them based on context. If tool use is a fixed sequence, the reasoning is yours, not the system's.
• "What does it do when it is unsure?" A serious agent has a concept of confidence and an escalation path to a human. A fake either barrels ahead or dead-ends, because it has no notion of not knowing.
• "Walk me through a failure in production." If a vendor claims their agent never fails, they are either not in production or not telling you the truth. Real deployments have failure stories and, more tellingly, the guardrails they added afterward.
• "How do you observe and evaluate it?" Ask about logging, tracing and evals. Teams running actual agents obsess over observability because they have to. Teams running scripts do not need it, so they have not built it.

The pattern across all five: specifics from a real vendor, deflection to marketing language from a washed one. When the answers slide toward "intelligent", "autonomous" and "next-generation" without a concrete mechanism underneath, you are hearing a brochure.

Why the label matters to your budget

This is not pedantry about definitions. The three failure modes of agent washing all cost money.

You overpay. Agentic products carry agentic pricing, often outcome-based or premium per-seat, and a rebranded chatbot delivering chatbot value at that price is a bad trade. You also over-scope. Sold an "autonomous" system, teams hand it problems that genuinely need reasoning, and a scripted tool fails at them quietly, eroding trust in the whole initiative. And you under-plan, because a real agent needs governance, guardrails, observability and a human-escalation design that nobody budgeted for when they thought they were buying a smarter FAQ bot.

The irony is that the honest version of each product is often the right buy. A well-built rules engine is a great fit for a well-understood, high-volume, low-variance task. The problem is not that scripts exist. The problem is paying agent prices, and building agent expectations, on top of one.

Build, buy, or buy the real thing

Once you can tell the difference, the build-versus-buy question gets clearer. Some tasks are so standard that a mature off-the-shelf agent, a genuine one, is the fastest path, and you should buy it. Some are so specific to how your business works that no vendor will ever model them well, and a custom agent built against your own systems and data is the only thing that actually works. The mistake is buying a generic "agent" for a specific problem and discovering it was a chatbot the whole time.

The way through is unglamorous. Define the task before you shop. Know whether it needs real reasoning or just reliable automation, because the honest answer is sometimes "just automation", and that is cheaper and more reliable anyway. Then make vendors prove agency against your questions, not their demo. We wrote a companion piece on how to evaluate an AI agent before you trust it that goes deeper on the testing itself.

If you want a second opinion on whether a vendor's "agent" is the real thing, or whether your problem is better served by a custom build than a subscription, that is a conversation we are happy to have. The vendors doing genuine agentic work will welcome hard questions. The ones doing agent washing will change the subject, and now you will notice.