A vendor demo opened with the word “agentic” and used it nineteen more times in forty minutes. What it actually showed was a chat box in front of a decision tree the company had been selling for years, now with a language model writing the prompts the user used to click. It was a perfectly decent product. It was not an agent. The buyer in the room nearly signed for the word rather than the thing — which is exactly what agent washing is designed to make happen.
What agent washing is
Agent washing is the practice of relabelling existing software — a chatbot, a robotic-process-automation script, a rules engine, or a single language-model call — as “agentic AI” to ride the category’s hype. It is the agentic-era successor to “AI washing”. The scale is not trivial: industry analysts have estimated that of the thousands of vendors now claiming to sell agents, only a small fraction are doing anything that meets a serious technical definition, and have warned that a large share of agentic projects will be cancelled as the gap between claim and capability becomes clear.
This matters because the word is doing real work in budgets and roadmaps. If you buy, build or restructure around “an agent” that is actually a workflow with a chat interface, you inherit the limitations of the old thing while paying for — and governing against — the risks of the new one.
What actually makes something an agent
Strip away the marketing and a useful working definition is narrow. An AI agent pursues a goal with some autonomy: it decides its own steps rather than following a fixed script, it acts on the world through tools, it observes the results of those actions, and it adapts — including recovering from failure — without a human in every loop. Take away the autonomy and you have automation. Take away the tool use and you have a chatbot. Take away the adaptation and you have a workflow.
| Capability | Real agent | Rebadged workflow / chatbot |
|---|---|---|
| Decides its own steps | Yes — plans toward a goal | No — fixed, pre-authored path |
| Acts through tools | Yes — calls systems, takes actions | Limited or none — it talks |
| Observes and adapts | Yes — reacts to results, recovers from failure | No — same path regardless |
| Autonomy | Operates without a human in every loop | A human drives each step |
None of this is a moral hierarchy. A deterministic workflow is often the right tool — more predictable, cheaper, easier to govern. The problem is not using automation; it is buying automation believing it is an agent, or building an agent where a workflow would have been safer and cheaper.
Half the “agents” we are asked to assess are last year’s workflow with a language model writing the labels. That can be a fine product — but you should buy it for what it does, not for the word on the slide. The word is the most expensive part.
Why it is worth the pedantry
Getting the category right changes three concrete things.
- Value: a real agent can handle variability a scripted workflow cannot — and also fail in ways one never could. Mistaking one for the other mis-sets every expectation that follows.
- Governance: an agent that acts through tools with autonomy needs controls a chatbot does not — scoped identity, action-level authorisation, an audit trail, a kill switch. Govern a relabelled workflow as an agent and you waste effort; govern a real agent as a workflow and you have a breach waiting to happen.
- Cost: agentic workloads can cost an order of magnitude more per task than a single model call. Paying agent prices for workflow capability is a quiet, recurring tax.
We treat the controls a genuine agent needs in our piece on access and the agent gateway — the point here is upstream of that: be sure you are governing the thing you actually have.
A buyer’s test
Before you believe the label, run four questions — out loud, in the demo.
- Does it decide its own steps, or follow a path someone pre-authored?
- Does it act through tools to change something, or only generate text?
- Does it observe results and adapt — including recovering when an action fails?
- Can it operate without a human in every loop, and should it?
Four yeses is an agent, with the value and the governance burden that implies. Mostly noes is automation — often the better choice, and one you should buy and price as exactly that. Either can be the right answer. Paying for the word instead of the capability never is.
Frequently asked
- What is agent washing?
- The practice of rebranding existing software — chatbots, RPA scripts, rule engines or a single model call — as “agentic AI” to capitalise on the hype. It is the agentic-era version of “AI washing”. Analysts estimate only a small fraction of self-described agentic vendors meet a serious definition of an agent.
- What makes something a real AI agent?
- A working definition: it pursues a goal with autonomy — deciding its own steps rather than following a fixed script, acting on the world through tools, observing the results, and adapting (including recovering from failure) without a human in every loop. Remove the autonomy and it is automation; remove the tool use and it is a chatbot; remove the adaptation and it is a workflow.
- Is a workflow worse than an agent?
- No. A deterministic workflow is often the better choice — more predictable, cheaper and easier to govern. The problem is not using automation; it is buying automation that has been relabelled as an agent, or deploying an agent where a workflow would have been safer and cheaper. Match the tool to the task, not to the trend.
- How do I avoid being sold agent washing?
- Run four questions in the demo: does it decide its own steps, act through tools, observe and adapt, and operate without a human in every loop? Four yeses is an agent — with the value and the governance burden that implies. Mostly noes is automation, which you should buy and price as exactly that.