Do not call a product an agent unless it completes a multi-step task on its own and makes decisions along the way without a person approving each one. That is the working test most skeptical reporters and analysts now apply before they will repeat the word in a headline, and it exists because so few products actually clear it. Gartner estimates that only around 130 of the thousands of vendors now marketing "agentic AI" offer something that genuinely meets the definition. Everyone else is, in Gartner's own term, agent washing. Here is what that means for your next announcement.
What is "agent washing," and who coined the term?
Gartner named it in a June 25, 2025 research note: vendors rebranding existing technology, chatbots, assistants, robotic process automation, as "agents" without adding the capabilities that make an agent different from those older categories. Gartner senior director analyst Anushree Verma put it directly: most current agentic AI offerings are "early stage experiments or proofs of concept that are mostly driven by hype" and "often misapplied," adding that "most agentic AI propositions lack significant value or return on investment" because the underlying models still lack the maturity to reliably pursue a complex goal on their own.
The term caught on because it described something reporters covering AI were already noticing without having a word for it: the same chatbot, rebuilt with a new label, pitched as a breakthrough.
How big is the gap between real agents and relabeled products?
Gartner's own estimate is that roughly 130 vendors, out of the thousands marketing agentic AI products, actually meet the bar. That is a gap of orders of magnitude between the category's marketing volume and the number of products Gartner considers genuine. The same research predicts that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls as the main reasons. Verma's explanation is blunt: companies underestimate the real cost and complexity of scaling an agent into production, so most of these projects never get there.
That prediction is now a live backdrop to how AI coverage gets written. A reporter who has read it, or just absorbed the skepticism it created, approaches an "agent" pitch assuming it probably is not one until shown otherwise.
What actually makes something an agent rather than a chatbot with extra steps?
The distinction that matters to a skeptical reader is autonomy over a sequence of decisions, not a single response. A chatbot answers a question. An assistant completes one task when asked. An agent takes a goal, breaks it into steps, decides what to do at each step based on what happened at the step before, and keeps going without a person approving every move. Robotic process automation follows a fixed script. An agent adapts when the script does not fit what it encounters.
That is a meaningfully higher bar than most products marketed as agents actually clear. A tool that drafts an email for a person to send is not an agent in this sense. A tool that decides which of ten possible emails to send, to whom, and adjusts its next action based on whether the first one got a reply, is closer to one. The honest version of most pitches lands somewhere between those two, and saying exactly where is more useful to a reporter than claiming the second one by default.
What will a reporter or analyst actually ask you?
Expect three questions, because they are the ones that separate a real agent story from a relabeled one.
How many steps does it take without a human approving each one? A specific number, even a small one, is more credible than "it's fully autonomous." If the honest answer is one or two steps before a human has to sign off, say that.
What happens when it encounters something it was not built to handle? Every agent fails at the edges of its training. A team that can describe its failure modes and what happens when one occurs sounds like it has actually run the thing in production. A team that says it does not fail sounds like it has not.
What is the exception rate in real use, not in a demo? A demo shows the case that was chosen to work. Production data shows what happens the other times. If you do not have that number yet because the product is new, say that directly rather than letting the question go unanswered.
What should actually go in the announcement instead of "agentic AI"?
Replace the category label with the specific workflow. Instead of "an agentic AI platform for customer operations," describe what it does: it reads an incoming support ticket, checks three systems for the customer's order status, and either resolves the ticket or routes it to a person with a summary of what it found. That sentence tells a reporter something checkable. "Agentic AI platform" does not.
Pair that with a number that reflects real use rather than a lab result: tickets resolved without escalation, average steps completed per task, time saved per resolved case, measured over a stated period with a stated sample size. The same discipline that applies to announcing any AI model or product launch applies here, benchmarks with full conditions, not cherry-picked wins, and a stated limitations section written before anyone has to ask for one.
And state plainly where a human is still in the loop. If your product escalates to a person for anything unusual, say so. A pitch that admits a boundary reads as more credible than one that claims none, because every reporter covering this space has now read enough overclaimed launches to recognize the pattern.
If the product really is early-stage, is honesty a liability?
No, and treating it as one is the mistake that produces agent-washing coverage in the first place. An early-stage agent that automates three real steps of a workflow, reliably, with a clear boundary where a human takes over, is a smaller but genuinely defensible story. A press release exists to be a factual, checkable record, and a modest claim that holds up under a reporter's three questions earns more durable coverage than a sweeping one that does not survive the first follow-up email. The pitch that gets covered twice, once at launch and again months later when the product has grown, is usually the one that told the truth about where it started.
The direct answer, for anyone who read this far
Call your product an agent only if it makes a sequence of decisions on its own, and be ready to say exactly how many steps it takes before a human has to step in, what it does when it fails, and what your real exception rate is. Gartner's own data says only a small fraction of the vendors using the word can back it up, and reporters covering this beat now know that. A pitch built around a specific workflow and an honest boundary will survive scrutiny that a pitch built around the word "agentic" will not. If you want help shaping how your product's capabilities actually translate into a story a skeptical reporter will run, talk to us.
