WHITEBOXX / GLOSSARY
Agent washing: when a chatbot
is sold as an agent.
Agent washing is the practice of marketing a product as an AI agent when it cannot carry out tasks on its own. Common examples are chatbots, assistants and rule-based automation renamed as agents. The term mirrors greenwashing.
What is agent washing?
In June 2025, Gartner described agent washing as the rebranding of existing products, such as AI assistants, robotic process automation and chatbots, without substantial agent capabilities. It estimated that only about 130 of the thousands of vendors claiming agentic AI were real.
Gartner also expects more than 40% of agentic AI projects to be cancelled by the end of 2027. Buying a renamed product is one way to end up in that group.
COMPARISON
Agent or assistant?
| Question | An assistant | An agent |
|---|---|---|
| Who decides the steps? | The user, one prompt at a time | The agent, within limits you set |
| Can it use tools and systems? | Sometimes, when asked | Yes, as part of completing the task |
| Does it finish the task? | It answers and waits | It works until the task is done or it needs a person |
| Can you check what it did? | Usually through a chat transcript | Through a record of its steps, tools and data |
BUYER QUESTIONS
How do you spot agent washing?
- 01
Ask for a live run on your data
A demo on prepared data shows little.
- 02
Ask what happens when a step fails
A real agent retries, tries another approach or asks for help.
- 03
Ask which tools it can use
And who sets the limits on them.
- 04
Ask for the record of a run
You should see the steps, tool calls and sources, not only the final answer.
- 05
Ask what is fixed and what the agent decides
Much useful automation is fixed steps. That is fine if the vendor says so.
- 06
Ask how results are tested
Look for test cases and measured accuracy, not hand-picked examples.
- 07
Ask who maintains it
Agents need updating when data, models or processes change.
How we describe Evidence OS.
Evidence OS runs workflows that combine fixed steps with execution-agent assignments. We say which parts are fixed and which an agent decides, and each run keeps a record you can inspect.
See execution agents