AI agent or rebranded chatbot? How to spot agentwashing before you invest

Open almost any SaaS website and there is a good chance the phrase 'AI agent' appears somewhere on the homepage. Within a year, the word has shifted from a technical term into a marketing label stuck onto chatbots, assistants, and even old automation scripts. Analyst firm Gartner has a name for this: agentwashing.

Of the thousands of software vendors now claiming to sell an agentic AI product, Gartner estimates only a small fraction actually deliver agentic functionality. The rest is a repackaged chatbot, a rule-based workflow, or a loose AI feature with a new label. For anyone deciding which tool is worth the budget, that is a practical problem.

What Gartner means by agentwashing

A real AI agent makes its own decisions within a defined task: it plans intermediate steps, calls tools and APIs, retains context across multiple steps, and adjusts its approach as the situation changes. A chatbot wearing a friendly label answers questions but plans nothing and takes no action. An RPA script runs a fixed sequence of steps but improvises nothing once anything deviates from the script.

The difference looks subtle in a product demo, where everything already runs smoothly. It becomes painfully clear once a process in practice runs slightly differently than expected and the 'agent' gets stuck, because there was never a decision-making layer underneath it.

Why forty percent of projects are still scrapped

Gartner predicted in 2025 that over forty percent of agentic AI projects would be canceled by the end of 2027. Not because the technology fails, but for three recurring reasons: costs escalate once a pilot has to move into production, business value stays unclear because nobody set a measurable goal beforehand, and risk management falls short once an agent is allowed to take independent action on company data or with customers.

That last point is often underestimated. A chatbot giving a wrong answer is annoying. An agent that independently sends an email, places an order, or changes a customer record based on a wrong assumption hits the business directly. Letting an agent run without oversight is a risk that, in most organisations, nobody has explicitly weighed.

Five questions that separate a real agent from a label

Before buying or building a tool labeled an agent, five questions help you see past the marketing. Does the system plan intermediate steps itself, or does it only perform a single action per instruction? Does it actually call other systems through an API, or does it stop at generating text? Does it retain context of an ongoing task across multiple steps, or does every request start from zero?

Can it adjust its approach if a first attempt fails, or does it simply get stuck? And is there a measurable outcome defined, such as time saved, an error rate, or a turnaround time, that lets you confirm in three months whether it actually works? A vendor with no concrete answer to any of these is probably selling a chatbot with a new label.

The label is everywhere, from customer service to marketing automation

The word agent now shows up in almost every software category: customer service tools claiming to independently resolve tickets, marketing platforms promising 'agentic' campaigns, and workflow tools that renamed their existing automations. That does not make all those products worthless, often they still do exactly what they always did, and that can be perfectly useful. The problem starts when an organisation sets a higher expectation and a bigger budget on the label than the product delivers.

How to set up a pilot that does not end up in the forty percent

The projects that do survive share a few traits. They start with one narrow, well-defined process instead of an organisation-wide rollout, for example pre-sorting incoming support tickets rather than fully automating customer service. They set a measurable success criterion upfront, so that after the pilot period an actual go or no-go decision is possible instead of a gut feeling.

They also build in an escalation path: the moment the agent moves outside its defined task or is uncertain, the task goes to a person instead of proceeding on a guess. And they budget not only for licence costs but also for the time needed to describe the process, open up the data, and adjust during the first few months. That last cost item is rarely mentioned in demos, and it is exactly where many projects run into trouble in practice.

What this means for your next decision

Agentwashing is not a reason to avoid AI agents; the projects that do succeed deliver demonstrable time and quality gains. It is a reason to look past the word agent on a website and ask about the five points above: planning, tool calls, context, adaptability, and a measurable goal. Having that conversation upfront keeps the budget from going to a rebranded chatbot and gets you a process that actually does independent work.

Not sure whether a tool you are considering is a real agent, or want to set up a pilot with a defined scope and a measurable outcome? Zarioh can help work out which step fits your process, from a first narrow pilot to a fully built AI agent. Contact us for a no-obligation conversation.