In mid-2025 Gartner published a prediction that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls. The same release coined a useful phrase for part of the problem: agent washing, meaning the rebranding of existing products — assistants, robotic process automation, chatbots — as agentic without substantial agentic capability. Gartner estimated that only about 130 of the thousands of agentic AI vendors are real.
If you are buying, this matters more than any benchmark. A chatbot with an agent sticker will pass a polished demo and fail on the first real workflow. The good news is that the difference is testable in a single working session.
What a real agent does that a chatbot doesn't
A chatbot produces text for a person to act on. An agent takes actions in your systems: it reads a queue, calls an API, writes a record, checks that the write worked, and decides what to do next based on the result. The practical test is whether anything changes in a system of record without a human typing it.
Five questions to ask any vendor
- Show me an action, not an answer — can it create, update or close something in a real system (ticket, order, CRM record) during the demo, on data you choose rather than data they prepared?
- What happens when a step fails — does it retry, escalate, or stop? An agent without a failure path is a script with a language model attached. Ask to see the failure path live.
- Where is the approval gate — which actions need a human, who gets asked, and how long does the system wait? Mature systems make this configurable per action type.
- Can I see the log of what it did and why — every tool call, input, output and decision, in a form an auditor could read. If the answer is a chat transcript, the audit trail doesn't exist.
- What does it cost per completed task — not per seat or per token. Vendors who can't state a per-task cost haven't measured one.
Quick filter: ask the vendor to break something. Give it a malformed input, an unavailable API, or an ambiguous instruction. Real agents have designed behaviour for that moment. Relabelled chatbots improvise.
Why the cancellation forecast is mostly a scoping problem
Gartner's three reasons — cost, unclear value, weak risk controls — are all decided before a model is chosen. Projects get canceled when nobody defined the task precisely enough to price it, when success was never expressed as a number, or when the blast radius of a wrong action was never designed. Those are scoping and governance decisions, and they are cheap to get right on paper and very expensive to fix in production.
What to do with this
Before signing anything labelled agentic, run the five questions in a working session and write down the answers. Then pick one narrow workflow with a reversible action, a measurable outcome and a named owner, and start there. If a vendor can't do that with you, you have learned something important for the price of an afternoon.