spotting agent washing

Of the thousands of vendors calling themselves agentic AI, only about 130 actually deliver agentic capabilities, according to Gartner. The rest sell polished chatbots and repackaged process automation. Gartner gave the phenomenon a name in June 2025, agent washing, and that word has since made it into legal reference works, but not into the average vendor demo. Anyone buying software today cannot wait for a regulator to filter first: in the Netherlands there simply is none yet. You have to set the bar yourself, and you can, with a handful of questions no vendor can answer with a slide.

In the 95 percent myth I hold the failure numbers around AI projects against their method. This essay is the buying side of that same story: how do you keep your project from starting on a promise the product cannot keep?

what exactly is agent washing?

Relabeling existing software as agentic AI without anything materially agentic about it. Gartner describes vendors giving AI assistants, chatbots and RPA tooling a new label, and predicts in the same press release that over 40 percent of agentic AI projects will be canceled before the end of 2027, through rising costs, unclear value or inadequate risk controls (Gartner, 2025). The estimate that only around 130 of the thousands of self-declared vendors are the real thing comes from that same release. Gartner publishes no precise methodology or denominator alongside it, so treat it as an order of magnitude from an analyst firm, not a measurement. Secondary sources happily turn it into “95 percent is fake”; that percentage appears nowhere in the primary text.

That the phenomenon can exist at all has a simple cause: there is no fixed definition to test against. Even MIT Sloan states explicitly that a universally agreed definition of agentic AI does not exist (MIT Sloan, 2025). The common rule of thumb, generative AI responds to a prompt, an agent perceives, reasons and acts independently toward a goal, is workable, but every vendor draws its own maturity ladder next to it. A definitional vacuum is a free port for marketing. Which is exactly why a test of your own works better than a label.

how do you tell bolted-on ai from agent-native?

With two tests you can run without any technical knowledge. The first is the removal test: imagine the AI feature disappearing from the product tomorrow. Does anything essential change about what the system can do? If the answer is no, the AI was a chat layer on top of an otherwise unchanged package. The second is the equivalence test: can the agent perform the same actions as a human at the screen? Not just summarize and suggest, but create, change, schedule, and at best configure the system itself. Analyst firm HFS Research casts the same check more formally as a two-gate test, agency and scalability, and warns specifically about copilots relabeled as agents while they only offer text-to-action inside existing workflows (HFS Research, 2025).

The distinction is not an academic matter. AI bolted on means every new automation is custom work all over again: the chat layer can talk about your data, but cannot act in your process. Agent-native means the actions themselves are reachable for agents, so the second and third automation no longer require a new project. You will not see that difference in the demo, because demos always show the best side. You see it in the architecture, and therefore in the answers to your questions.

which questions do you ask a vendor?

Six, and you ask them in plain language. The answers tell you more than any product video.

1. Can the agent perform the action itself, or only advise? Ask concretely: can it create an order, move an appointment, add a field? In which systems does it have write access?
2. Can I see a live demo on my own scenario? No recorded video, no prepared script: your customer, your process, on the spot.
3. Where do I find what the agent has done? A serious product keeps a log per action. HFS explicitly advises buyers to demand evidence, test results, audit logs, run traces, instead of accepting claims.
4. What happens when the agent gets it wrong? Is the action reversible, are there permission boundaries, when does it bring in a human?
5. How do you measure the success rate? Customer service vendor fin.ai, itself a player in this market, describes two patterns to watch for: blended figures in which human and machine form one score, and customers who give up without help but are counted as “resolved by AI” (fin.ai, 2026). Ask for the number without human intervention.
6. Does the platform distinguish between human users and agents? Separate permissions for agents are a sign the system was designed for agents, not an afterthought (MindStudio, 2026, vendor-affiliated but concrete enough to ask).

A vendor who answers these questions with ease no longer needs any label. A vendor who falls back on the slide with the word agentic on it has given you your answer too.

where does it go wrong when nobody presses?

Then it eventually becomes a matter for the courts, and that is not a theoretical scenario. The American FTC has been running enforcement cases against misleading AI claims since September 2024 under the banner Operation AI Comply, from the “robot lawyer” DoNotPay to AI tools that wrote fake reviews (FTC, 2024). The SEC fined investment advisers who claimed AI but did not use it, and in January 2025 took on the first listed company with Presto Automation: the “proprietary” speech technology turned out to come from a third party, and most orders still needed human processing (SEC, 2025). The founder of shopping app Nate has even faced criminal prosecution since April 2025: the app promised neural networks, the orders were done by hand by contract workers (Holland & Knight, 2025). The pattern across all those cases is always the same, and it reads like the mirror image of the six questions above: technology claims that were not true, inflated automation percentages, concealed human intervention.

Two famous stories deserve the same skepticism you grant your vendors. The viral story that Builder.ai “had 700 engineers pretending to be AI” is not true: the most thorough reconstruction shows a small AI team, a thick layer of legitimate outsourced development marketed as AI speed, and a bankruptcy that ultimately came from revenue fraud, not from the AI fiction itself (Pragmatic Engineer, 2025). And Amazon’s checkout-free stores were not “just a thousand people instead of AI”: there was a genuinely working model, and the dispute was about the concealed scale of the human review (The Verge, 2024). The lesson is more precise than the meme: a human in the loop is normal in serious AI systems. The red flag is a vendor who will not tell you the ratio honestly.

and in the netherlands?

Not a single case on agentic claims here yet. The closest precedent is a ruling by the Advertising Code Committee about AI-generated songs sold as “personally composed” (Stichting Reclame Code, 2024): real, but narrow. Dutch AI supervision is spread across some ten bodies, with the ACM in position for consumer deception, and the transparency obligations from the AI Act only start applying on 2 August 2026 (ICTRecht, 2026). Against greenwashing the ACM has proven it can come down hard on vague claims; that playbook fits AI claims seamlessly in theory, but today that is an analogy, not practice.

The sober conclusion: anyone buying software in the Netherlands is wise to act as if no regulator is ever coming. The six questions cost half an hour in a sales conversation. A project that starts on a promise the product cannot keep costs a year. And for the record: I build agent-native software myself, so feel free to hold my work to the same bar. That is exactly what it is for.