How to Spot "Agent Washing": Is That AI Agent Real or Just a Chatbot?
TL;DR
Learning how to spot fake AI agent vendors starts with one distinction: AI agent washing calls a fixed AI workflow an autonomous AI agent.
Runtime planning is the dividing line. An agent chooses and changes its steps while it works. A chatbot follows paths defined in advance.
Tool use alone proves nothing. A chatbot can take useful actions without planning its own next step.
The fastest test is to break one step during the demo and watch what happens next.
matram.ai publicly describes a grounded AI support agent trained on approved content. Its public pages do not claim open-ended planning, so buyers should not assume that capability from the label.
What is AI agent washing?
Agent washing is marketing a chatbot or fixed workflow as an autonomous AI agent. The product may use an AI model, call tools, update a field, or book a meeting. Those features can be useful, but they do not prove autonomy.
The issue is not the word "agent" by itself. In customer service, an agent can simply mean the person or software handling a conversation. The problem starts when that role name is used to imply technical abilities the product does not have.
What separates a real AI agent from a chatbot?
An autonomous AI agent receives a goal, chooses the steps and tools, checks each result, and adjusts its plan when something changes.
A chatbot or workflow follows routes designed in advance. It can still answer questions, capture leads, book meetings, and start a handoff. Those actions do not prove autonomy.
The clearest test is one question:
Who chose the steps, the system or a person, and when?
If a person configured the path earlier, it is a workflow. If the system builds and changes the path while completing the task, it is acting as an autonomous agent.
Related read: AI Agent vs Chatbot: What Actually Separates Them
We applied the agent-washing test to matram.ai
A vendor criticising agent washing should grade its own claims first. Here is what Matram publicly confirms, what it does not state, and what a buyer should verify.
| Test | What Matram publicly documents | Honest reading |
|---|---|---|
| Chooses its own multistep plan at runtime | Not stated | Do not infer autonomous planning from the word "agent." Ask for a live demonstration if you need it. |
| Chooses tools dynamically while working | Not stated | Tool selection and recovery behaviour must be verified in the demo. |
| Creates a new route after an unexpected failure | Not stated | Test a failed action before treating Matram as an autonomous planner. |
| Answers from approved content and shows the source | Yes | Replies are grounded in indexed website, help center, and document content. |
| Declines when approved content does not support an answer | Yes | Strict Mode keeps the response inside the supplied material. |
| Captures and qualifies leads | Yes | Lead details are saved for the team to review and export. |
| Hands a conversation to a person | Yes | The conversation can move to human support with its context attached. |
Verdict: Matram's public evidence supports a grounded AI support agent for answers, citations, lead capture, Strict Mode, and human handoff. It does not establish open-ended autonomous planning.
No spin. No convenient exception.
That boundary does not make Matram less useful. It is built to answer from approved content, show the source, capture leads, and bring in a person when the request moves outside its scope. Strict Mode keeps answers inside the material you approved instead of letting the system invent a response.
On the Matram AI agents page, the product is presented as an AI support agent trained on a business's content. The published capabilities centre on grounded answers, citations, lead capture, Strict Mode, and handoff. The page does not ask buyers to treat the label alone as proof of autonomous planning.
If your work needs a system to create and change its own multistep plan across several tools, ask Matram to demonstrate that exact behaviour before buying. Do not convert an unstated capability into either a promise or a criticism.
How to spot fake AI agent vendors during a demo
A polished demo proves that the prepared route works. The practical answer to how to spot fake AI agent vendors is to test whether the product can genuinely plan, verify, and adapt. These seven questions do that.
| Question | Proof to ask for | Red flag |
|---|---|---|
| 1. Can it choose its own steps from one goal? | Give it a goal without prescribing the sequence. Ask to see how the plan is created. | The vendor shows only a flow builder or a route prepared before the demo. |
| 2. Who chooses the tool at runtime? | Give the system more than one suitable tool and inspect why it selects one. | Every tool call is attached to a fixed intent or button. |
| 3. What happens when a tool call fails? | Disable a tool, return missing data, or trigger an error during the demo. | The system repeats the same step, reports success, or enters a prewritten branch. |
| 4. Does it verify the result? | Ask to see the external system's response before the customer is told the task is complete. | It assumes that sending a request means the action succeeded. |
| 5. Which actions need human approval? | Test a sensitive action and ask where approval and access limits are set. | The vendor cannot explain the control boundary. |
| 6. Can you inspect the action trace? | Ask to see the chosen steps, tool calls, results, retries, and final decision. | The vendor shows the final answer but not how it was reached. |
| 7. Where is the product not a fit? | Ask for one task the product should hand to a person or another system. | The vendor claims it can handle every task for every team. |
One failed step can reveal more than a complete feature tour. That is how to spot fake AI agent vendors behind a polished feature tour. A real autonomous agent should notice the failure and choose a safe next move. It may retry, use another approved tool, request missing information, ask for approval, or stop cleanly.
If every possible branch was configured beforehand, you are looking at a workflow. That can still be the better purchase. It simply should not be sold as open-ended autonomy.
How to spot fake AI agent vendors before a demo
You can often spot agent washing on the vendor's website.
- "Agent" appears everywhere, but planning is never explained. A real autonomy claim should explain how the system chooses and changes its steps.
- Every example is one action. Answering a question or booking a meeting is useful, but it does not show runtime planning.
- Failure is missing from the story. Genuine autonomy needs clear recovery, stopping, and escalation rules.
- Permissions are vague. A system taking real actions needs defined limits on what it can change.
- Every branch appears in a visual builder. That is strong evidence of a workflow designed in advance.
One sign alone is not proof. Several together are a reason to ask harder questions. This is how to spot fake AI agent vendors without dismissing every useful workflow as deceptive.
Do you actually need an autonomous AI agent?
Many customer service tasks do not require runtime planning. They need a correct answer, a captured lead, or a clean handoff.
Choose a grounded support system when most requests end in one of those outcomes. This is the documented job Matram is designed to handle.
Consider an autonomous agent when the task requires several actions across different systems and the order of those actions changes from case to case. Then test its failure recovery, approval rules, and action trace before buying.
If you cannot name a real customer request that needs runtime planning, do not pay for the label. Choose the product that solves the job.
FAQs
What is agent washing?
Agent washing is marketing a chatbot or predefined workflow as an autonomous AI agent. The label suggests runtime planning that the product does not perform.
How can you tell whether an AI agent is real?
Break one step during the demo. A real autonomous agent should detect the failure and choose a safe next move. Then ask to see the steps, tool calls, results, and approvals behind its decision.
How to spot fake AI agent vendors?
Ask who chooses the steps, test a failed tool call, inspect the action trace, and confirm where human approval is required. Vendors that cannot show those boundaries may be selling a workflow as autonomy.
Does using tools make a chatbot an AI agent?
No. A chatbot can call tools through predefined paths. Autonomy begins when the system chooses and changes its actions at runtime.
Is matram.ai an autonomous AI agent?
Matram's public pages do not establish open-ended autonomous planning. They document grounded answers, citations, Strict Mode, lead capture, and human handoff. If autonomous planning matters to your use case, ask to see it demonstrated.
When is matram.ai the right choice?
Matram fits teams that need answers grounded in approved content, visible sources, lead capture, and human handoff. Teams needing unscripted multistep work across external systems should verify that capability directly before choosing any vendor.
Test the claim, not the label
Knowing how to spot fake AI agent vendors protects you from paying for autonomy that the product does not have. Before you buy anything called an AI agent, break one step and watch what happens next. That one move can tell you more than the entire feature page.
If your team needs grounded answers, visible sources, lead capture, and human handoff, try matram.ai free for 30 days. No card is required. See exactly what it does, what it does not do, and whether it fits your team.
See Matram in action
Watch it answer from real content, with the exact source page cited, on the live demo.
Book a demo