Gartner estimates that out of the thousands of vendors marketing "AI agents," only around 130 are the real thing. The rest are something else wearing a new label - a chatbot with a fresh coat of paint, a rules engine rebranded, last year's automation tool with "agentic" stapled to the datasheet. The industry has a name for this now: agent washing. And if you are shopping for an AI agent in 2026, it means the odds are overwhelmingly that the thing being pitched to you is not what the pitch says it is.
That sounds grim, but it is actually empowering - because the gap between a real agent and a re-skinned one is not subtle once you know what to look for. You do not need to be technical to catch it. You need five questions and the nerve to insist on straight answers. This article hands you both.
The stakes are simple. Buy a washed agent and you pay agent prices for chatbot capability, then spend months discovering it cannot actually do the job. Catch the washing in the sales call and you save the budget, the time, and the credibility you would have burned defending the choice internally.
What "Agent Washing" Actually Is
Agent washing is the practice of relabelling existing technology as an "AI agent" to ride the hype, without the capabilities the term implies. It is the 2026 version of every bandwagon rebrand that came before it - and it works because "agent" has no agreed, enforced definition in the market, so anyone can claim it.
To catch it, you need a working definition of the real thing. A genuine AI agent does three things a chatbot or a fixed automation does not:
- It pursues a goal over multiple steps - it plans and sequences actions toward an outcome, rather than answering one prompt at a time or following a single hard-coded script.
- It takes real action in your systems - it reads from and writes to your tools to actually complete the work, not just suggest what you should do.
- It adapts - when a step fails or the situation shifts, it adjusts its approach instead of breaking, within the guardrails you set.
Most washed "agents" fail at least one of these. A re-skinned chatbot answers but does not act. A rebranded automation acts but cannot adapt - it shatters the moment reality deviates from its script. A "co-pilot" relabelled as an agent only ever suggests and hands the work back to you. None are bad tools. They are just not agents, and they should not carry agent pricing or agent expectations.
The 5-Question Test: Real Agent or Re-Skin?
Take these into your next vendor call. Each targets a specific way washing hides. Vague, deflecting, or demo-only answers are the tell - a real agent vendor answers all five concretely and without flinching.
Question 1: "Show me it completing a multi-step task end to end, on real data."
Not a scripted demo on a clean sandbox - a genuine task with messy, realistic inputs. Watch whether it actually finishes the job or just produces a suggestion you would still have to execute. Washing tell: the demo only ever shows the happy path, or the "agent" stops at recommending and a human does the doing.
Question 2: "What exactly can it do in my systems - read, write, or just talk?"
A real agent integrates and acts: it logs into tools, updates records, sends, triggers workflows. Push for the specific actions and the specific systems. Washing tell: it "integrates" but on inspection only reads or displays information, or "takes action" only by drafting something for you to send yourself. Answering is not acting.
Question 3: "What happens when something goes wrong or unexpected mid-task?"
This is the question that exposes rebranded rules engines fastest. A real agent adapts, retries, or escalates intelligently. Washing tell: the honest answer is some version of "it follows the configured flow" - meaning it is a fixed script that breaks on anything off-path. Ask for a concrete example of it recovering from a surprise.
Question 4: "Where does a human stay in the loop, and how do I control what it does automatically?"
Counterintuitively, a good answer here describes guardrails, approval gates, and an audit log - because real agents take real actions and serious vendors design the controls in. Washing tell: either there are no controls to speak of (it cannot actually act, so none are needed), or the vendor cannot explain how you would gate, pause, or audit it. Both are red flags pointing in opposite directions.
Question 5: "How is this priced, and what is the cost per task at my volume?"
Real agents call AI models on every step, so their cost scales with usage and a credible vendor can model your cost per completed task. Washing tell: flat per-seat pricing identical to ordinary SaaS often signals there is no real model-driven agent underneath - just software with an agent label. Ask what happens to the bill when volume doubles; the answer reveals the architecture.
The Quick Buyer's Checklist
Before you sign anything that calls itself an AI agent, confirm you can answer yes to these. Every "no" is a place agent washing hides.
- ☐ You have seen it complete a real, multi-step task end to end on messy data - not a scripted happy-path demo.
- ☐ You can name the specific actions it takes in your specific systems - and confirmed they include writing, not just reading or suggesting.
- ☐ The vendor gave a concrete example of it adapting to a failure or surprise mid-task.
- ☐ There are real controls - approval gates, guardrails, an audit log, and a kill switch.
- ☐ You understand the cost per completed task at your volume, and what happens when usage scales.
- ☐ You spoke to a reference customer running it in production for a use case like yours - and asked them what broke.
- ☐ You own your data, prompts, and any resulting configuration - with no lock-in that makes leaving impossible.
Real Agent, Wrong Fit Is Still a No
One nuance worth holding: passing the test means the technology is genuinely an agent - not that it is right for you. A real agent pointed at the wrong workflow, or one that needs human judgment it cannot supply, will still disappoint. The five questions filter out washing; your own use case has to filter out mismatch. The best vendors help with both - they will tell you when an agent is overkill for your problem and a simpler automation would serve you better. That honesty is itself a strong signal you are dealing with the real 130, not the washed thousands.
What This Means For You
"Only ~130 are real" is not a reason to avoid AI agents - it is a reason to shop like a professional. The capability is genuine and valuable when you find it. The market is just crowded with imitations because the label is free and the hype is loud. Your defence costs nothing: five pointed questions, a checklist, and the discipline to treat a demo as marketing rather than proof.
Ask the questions. Insist on real tasks, real systems, real failure-handling, real controls, and honest pricing. The vendors who answer cleanly are the ones worth your budget. The ones who deflect just told you everything you needed to know - and saved you from finding out the expensive way.
Not Sure If What You're Being Sold Is a Real Agent?
Bring us the vendor pitch, the demo, or the workflow you want to automate. We will run it through the test above, tell you honestly whether it is a real agent or a re-skin, and whether an agent is even the right tool for your problem - with a fixed written estimate if you want us to build the real thing instead.
Frequently Asked Questions
What is "agent washing"?
Agent washing is relabelling existing technology - a chatbot, a rules-based automation, a suggestion-only co-pilot - as an "AI agent" to ride the hype, without the capabilities the term implies. Gartner estimates only around 130 of the thousands of vendors marketing AI agents are offering the real thing, which is why a buyer's odds of being pitched a re-skin are high.
What makes something a real AI agent versus a chatbot?
A real agent pursues a goal over multiple steps, takes real action in your systems (reading and writing, not just answering or suggesting), and adapts when a step fails or the situation changes. A chatbot answers one prompt at a time, a rebranded automation follows a fixed script and breaks off-path, and a co-pilot only suggests. Genuine agents act, adapt, and need governance.
How do I tell if a vendor is agent washing?
Ask five things: show it completing a real multi-step task on messy data; name the exact actions it takes in your systems; explain how it handles something unexpected mid-task; describe the human controls, approval gates, and audit log; and give the cost per completed task at your volume. Concrete answers indicate a real agent; deflection to demos, vague "AI-powered" language, or flat SaaS pricing indicates washing.
Why does pricing reveal whether an agent is real?
Real agents call AI models on every step, so their cost scales with usage and a credible vendor can model your cost per completed task and explain what happens when volume doubles. Flat per-seat pricing identical to ordinary software often signals there is no model-driven agent underneath - just relabelled SaaS.
If an agent passes the test, is it right for my business?
Not necessarily. Passing means it is genuinely an agent, not that it fits your workflow. A real agent aimed at a task that needs human judgment it cannot supply will still disappoint. Filter out washing with the five questions, then filter out mismatch with your own use case - and favour vendors honest enough to tell you when a simpler automation would serve you better.
