Agent washing
Riskagentwashing
Relabelling as an "AI agent" a product that isn't one: assistants, RPA or chatbots sold under a new name without real ability to plan, act and adapt. The term is Gartner's, which in June 2025 estimated only around 130 of the thousands of vendors marketing themselves as agentic actually are.
Gartner coined "agent washing" in the same note where it predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. The pattern is recognisable: something that only runs a fixed flow, decides nothing off-script and cannot recover from an error gets presented as an agent. Spotting it is a procurement job, not a marketing one: ask for the trace of a real run, look at which tools it invokes on its own initiative, and check what happens when the scenario leaves the happy path. In AI Operations the cost of agent washing isn't just the contract — it's a project that burns six months of your team to land exactly where you started.
How it differs from
- Agentic workflow
- A real agentic workflow lets the model choose the sequence of steps and tools within bounds. In agent washing the sequence is fixed up front and the model only writes the prose.
Examples
- A fixed decision-tree support chatbot resold as an "autonomous agent"
- An RPA flow with one LLM call in the middle presented as a multi-agent system
FAQ
- How do you spot it in a demo?
- Take the demo off-script. Ask for a case the vendor hasn't prepared, ask to see the execution trace (which tools it called and why) and ask what the system does when a tool fails. A real agent replans; a disguised flow either breaks or makes something up.
- Does this mean agentic AI is vapour?
- No. It means the market is ahead of the product. Gartner's prediction that over 40% of these projects will be cancelled before 2028 points at cost and controls, not at the technology being useless.