Walk the floor of any technology conference this autumn and count the stands that do not mention agents. It will not take long. Chatbots have become agents. Workflow tools have become agents. Last year's robotic process automation suite has, with a fresh logo and a new landing page, become an "agentic platform".
Gartner has a name for this: agent washing. In June it estimated that only about 130 of the thousands of vendors claiming agentic AI actually offer it, and predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls (Outlook Business, reporting Gartner).
Put those two numbers side by side and the risk for buyers is obvious. If you cannot tell a real agent from a relabelled script, you cannot price it, govern it or know whether it will survive contact with your operations.
What makes something an AI agent
Strip away the marketing and an agent does three things a traditional automation tool does not. It decides which steps to take towards a goal rather than following a fixed path. It uses tools, such as APIs, databases and other systems, to act on those decisions. And it notices when a step fails and adapts, instead of stopping at the first exception.
A chatbot answers. An RPA bot repeats. An agent works out what to do next. That distinction matters because each comes with a very different cost profile, failure mode and governance need. A rules-based bot that breaks is annoying. An agent that improvises badly with write access to your order system is a different conversation.
None of this means the older tools are bad. RPA is excellent at stable, high-volume, rules-based work. Sometimes a "dumb" automation is exactly the right answer. The problem is paying agent prices for script behaviour, or worse, assuming you have script-level predictability when you have actually bought agent-level autonomy.
Five questions to ask any agentic AI vendor
These are the questions we would put to any vendor. The answers tell you far more than any demo.
- What does it decide on its own? Ask for a specific task and the decisions the system makes inside it. If every branch was designed in advance by a human, you are looking at a workflow, however clever the interface.
- Show me it recovering from a failure. Break something live: an API timeout, a missing field, an ambiguous request. Real agents re-plan. Washed ones throw an error or hand over to a person.
- What can it touch, and with what permissions? You want a precise list of tools, scopes and approval points. Vague answers here are the biggest warning sign of all.
- How do you know it works? Ask for their evaluation method, the test set and the failure rate. "Our customers love it" is not an evaluation.
- What does a completed task cost? Agents are typically priced on usage, and usage grows with every retry and loop. A vendor that cannot estimate cost per task has not run it at scale.
Why the label matters less than the outcome
It is tempting to treat this as a branding argument. It is not. The label decides how you buy, and how you buy decides whether the project ends up in Gartner's 40%.
If you believe you are buying an agent, you will design for autonomy: looser process maps, fewer approval steps, bigger ambitions. If what arrives is a brittle script, the gap between expectation and reality becomes the reason the project gets cancelled. The reverse is just as costly. Buy a real agent while thinking it is a predictable bot and you will under-invest in guardrails, monitoring and ownership.
So start from the problem, not the product category. Is the work variable enough to need judgement? Does it cross several systems? Is the cost of a wrong action recoverable? If the answers are no, a well-built conventional automation may beat an agent on cost, reliability and auditability. If the answers are yes, you need an agent, and you need to evaluate it properly before it goes anywhere near production.
A practical next step
Before your next vendor meeting, write down one real task you want automated, the systems it touches and what a wrong action would cost. Then run the five questions against it. You will learn in an hour what a three-month pilot might otherwise teach you the expensive way.
That is the kind of decision our AI Opportunity Assessment is designed to make quickly: where AI agents earn their keep, where simpler automation wins, and where neither is worth it yet.
Frequently asked questions
What is agent washing?
Agent washing is the practice of rebranding existing products, such as chatbots, AI assistants or RPA tools, as AI agents without adding real agentic capability. The term was coined by Gartner.
What is the difference between RPA and an AI agent?
RPA follows a fixed, pre-designed sequence of steps and stops when something unexpected happens. An AI agent decides which steps to take towards a goal, uses tools to act, and adapts when a step fails.
How many agentic AI vendors are real?
In June 2025 Gartner estimated that only about 130 of the thousands of vendors claiming agentic AI capabilities offer real agentic technology.