Agent Washing: How to Tell a Real AI Agent From a Glorified Workflow
Everything is an 'agent' now, which means nothing is. Here's the test that separates products that actually decide and act from chatbots with a thesaurus.
Sometime in late 2025, every product in enterprise software became an "agent." The chatbot that answers FAQs? Now an agent. The if-this-then-that automation you've run for years? Agentic. The macro that fills in a form? An autonomous AI agent, according to the new landing page.
This is agent washing, and it's the defining marketing disease of 2026. When everything is an agent, the word means nothing — which is exactly the point for vendors who'd rather rebrand than rebuild. If you're buying, building, or betting on this category, you need a test that cuts through it.
The word lost its meaning on purpose
"Agent" became the magic word because it signals the frontier. It implies autonomy, intelligence, the future. So naturally, every product manager under pressure to look AI-forward reached for it, regardless of whether the underlying product did anything agentic.
The result is a market where the label tells you nothing. A genuinely autonomous system that plans and adapts gets called an agent. A decade-old rules engine gets called an agent. A thin chatbot wrapper gets called an agent. Same word, wildly different things, deliberately blurred.
This isn't harmless. Buyers pay premium prices for "agents" and get repackaged automation. Teams build on the assumption of capabilities that aren't there. And the genuine agentic products struggle to differentiate, because the word that should describe them has been drained of meaning.
The actual dividing line: who decides the path
Strip away the marketing and there's a clean technical distinction.
A workflow follows a path you defined. You decided the steps, the branches, the conditions. The system executes your plan. It might use AI inside a step — classify this, extract that — but the structure is fixed and human-authored. When something happens that you didn't anticipate, the workflow breaks or punts to a person.
An agent is given a goal and decides the path itself. It plans, picks which tools to use and when, observes what happened, and adapts its next move based on the result. You didn't script the sequence — you specified the objective and gave it capabilities. When something unexpected happens, a real agent reasons about it and adjusts.
The dividing line is autonomy over the path, not the presence of an LLM. A chatbot with an LLM is still a chatbot. An agent is defined by who decides what to do next: you, in advance, or the system, in the moment.
The test that exposes agent washing
When you're evaluating a product that claims to be an agent, two questions do almost all the work:
"Show me it handling a situation you didn't script for." A real agent can deal with novelty — a tool returns an error, the data is in an unexpected format, the user wants something off the happy path. It reasons about the new situation and finds a way. A washed agent has only the paths its builders anticipated, and falls over the moment reality deviates. Ask for the unscripted demo, not the rehearsed one.
"What happens when a step fails?" This is the killer question. A workflow's failure handling is whatever the builder hard-coded — usually "stop and alert a human." A real agent observes the failure, reasons about why, and tries an alternative. The recovery behavior tells you everything about whether there's genuine agency underneath or just a decision tree with a chat interface.
Newsletter
Like this decoded read? Get the next one in your inbox.
One sharp take on the latest in AI for PMs, founders and operators. Every few days. Free.
Free forever. Unsubscribe anytime. No spam.
If the answer to both is some version of "well, it depends on the configuration" and a quick subject change, you're looking at agent washing.
Workflows aren't the enemy — confusion is
Here's the nuance the hype machine misses in both directions: workflows are frequently the right choice, and the problem is mismatch, not workflows themselves.
For a predictable, high-stakes, repetitive task, a deterministic workflow is safer, cheaper, and more reliable than an agent. You want the path to be fixed when you're moving money or making compliance decisions. Agency introduces variance, and variance is the last thing you want in a process that must behave identically every time.
So the failure isn't "they used a workflow." It's the two mismatches:
Paying agent prices and accepting agent unpredictability for something that's really a workflow — you took on risk and cost for capability you didn't need.
Expecting workflow-grade reliability from something genuinely agentic — agents are probabilistic and will occasionally do something surprising, and deploying one into a process that demands determinism is asking for trouble.
The skill in 2026 is diagnosing which kind of problem you have, then matching the tool to it — and not letting a vendor's label make that decision for you.
What to do if you're buying
Run the pilot on your worst cases, not the demo's best ones. Vendors curate the happy path; your reality is the long tail. Hand the system your messiest, most ambiguous real examples and watch what it does.
Ask what the system does autonomously versus what it routes to a human, and get specific numbers. "Fully autonomous resolution" should come with a rate, and you should sample the cases it claims to have handled to confirm they were actually handled well, not just closed.
And price it against what it actually is. If it's a workflow with good AI inside, that can be genuinely valuable — just don't pay the autonomy premium for it.
The takeaway
Agent washing is what happens when a powerful idea becomes a marketing requirement faster than products can become it. The word "agent" now tells you almost nothing, so stop trusting the label and start testing the behavior.
Real agents decide the path and recover from failure. Workflows follow a path you defined and break when reality surprises them. Neither is better in the abstract — but paying for one and getting the other, or deploying one where you needed the other, is the expensive mistake everyone's making. Ask for the unscripted demo, ask what happens when a step fails, and let the answers, not the homepage, tell you what you're actually buying.
Frequently asked questions
What is agent washing? Agent washing is rebranding existing automation, chatbots, or rules-based workflows as 'AI agents' to ride the hype. The product hasn't changed in any meaningful way — the marketing has. It's the agentic-era version of greenwashing, and it's rampant in 2026 enterprise software.
What actually makes something an AI agent? A real agent has autonomy over how it achieves a goal — it plans, chooses which tools to use, observes results, and adapts. A workflow follows a fixed path you defined. The test: if you have to script every step and the system can't recover from an unexpected situation, it's a workflow wearing an agent costume.
Are workflows worse than agents? No — and this is the trap. For predictable, high-stakes tasks, a deterministic workflow is often the right and safer choice. The problem isn't using workflows; it's paying agent prices and taking agent risks for something that's actually a workflow, or expecting workflow reliability from something genuinely agentic.
How do I avoid buying agent-washed software? Ask the vendor to show the agent handling a situation it wasn't explicitly scripted for, and ask what happens when a step fails. Real agents adapt and recover; washed ones break or fall back to a human. Run a pilot on your messiest real cases, not the clean demo path.
Want more sharp takes on AI tooling? Browse the AI Decoded archive.
Frequently asked
What is agent washing?
Agent washing is rebranding existing automation, chatbots, or rules-based workflows as 'AI agents' to ride the hype. The product hasn't changed in any meaningful way — the marketing has. It's the agentic-era version of greenwashing, and it's rampant in 2026 enterprise software.
What actually makes something an AI agent?
A real agent has autonomy over how it achieves a goal — it plans, chooses which tools to use, observes results, and adapts. A workflow follows a fixed path you defined. The test: if you have to script every step and the system can't recover from an unexpected situation, it's a workflow wearing an agent costume.
Are workflows worse than agents?
No — and this is the trap. For predictable, high-stakes tasks, a deterministic workflow is often the right and safer choice. The problem isn't using workflows; it's paying agent prices and taking agent risks for something that's actually a workflow, or expecting workflow reliability from something genuinely agentic.
How do I avoid buying agent-washed software?
Ask the vendor to show the agent handling a situation it wasn't explicitly scripted for, and ask what happens when a step fails. Real agents adapt and recover; washed ones break or fall back to a human. Run a pilot on your messiest real cases, not the clean demo path.