"AI agent" is the most overused label in technology right now. Almost every software vendor has added it to their product pages. Some of those products really are agents. Many are not.
Analysts at Gartner have a name for this: agent washing, the rebranding of existing products, such as chatbots, assistants, and robotic process automation, as "agentic AI" without substantial agent capabilities. In its 2025 prediction that over 40% of agentic AI projects will be cancelled by 2027, Gartner estimated that only a small fraction (around 130) of the thousands of vendors claiming agentic AI offer real agentic capabilities.
This guide helps you see through the label so you spend your money on what you actually need.
First, what is a real AI agent?
A genuine AI agent can:
- Pursue a goal, not just answer one message
- Plan steps to reach that goal
- Use tools (APIs, databases, software) to take actions
- Observe results and adjust
- Operate with limited supervision within defined boundaries
If you need a refresher, read our comparison of AI agents vs chatbots.
What agent washing looks like
| The claim | What it often is |
|---|---|
| "Autonomous AI agent for customer support" | A chatbot that answers from a knowledge base |
| "Agentic workflow automation" | A fixed, rules-based workflow with one AI step |
| "AI agent that handles your inbox" | Email filters and suggested replies |
| "Digital worker" | A script or RPA bot with a chat interface |
| "Multi-agent system" | Several prompts calling each other in a fixed order |
None of these is automatically bad. A knowledge-base chatbot, a fixed workflow, or RPA can be exactly the right solution. The problem is paying agent prices, or expecting agent results, for something simpler.
Eight questions that expose agent washing
1. What goal does the agent pursue, and who sets it?
Real agents work towards an outcome. Ask for a demonstration of a multi-step task, not a scripted Q&A.
2. What tools can it use, and which actions can it take?
If the answer is "it can answer questions", it is a chatbot. Ask which systems it can read from and write to.
3. How does it decide what to do next?
If the steps are always the same, it is a workflow. Ask what happens when something unexpected occurs.
4. Show me a failure
Good vendors can show you how the system fails and recovers. If every demo is perfect, be careful.
5. What are the guardrails?
Ask about permissions, human approval, logging, and limits. Our guide to AI agent security and guardrails lists what a responsible setup should include.
6. What results have customers actually seen?
Ask for specific, measurable outcomes and references, not slogans. Treat large percentage claims with scepticism unless the vendor explains how they were measured.
7. What does it really cost to run?
Agents can make many model calls per task. Ask how pricing works and what happens at volume. See our article on LLM API costs.
8. Do I even need an agent?
A good vendor will tell you when a simpler tool is enough. If everything you describe sounds like it needs "an agent", ask for the reasoning.
A quick scoring test
Give the product one point for each "yes":
- Can pursue a goal across several steps
- Chooses between different actions depending on the situation
- Uses at least two external tools or systems
- Handles unexpected inputs without a pre-written script
- Provides logs of what it did and why
- Has clear permission limits and approval options
5–6: very likely a genuine agent. 3–4: a hybrid workflow with agent elements. 0–2: probably a chatbot or fixed automation with an "agent" label.
Why this matters
- Money. Agent pricing is often higher than for the same capability sold as a chatbot or automation.
- Expectations. Teams that expect autonomy and get a scripted bot become disappointed and write off AI altogether.
- Risk. Real agents need real safeguards. Buying a real agent without governance is risky.
- Project failure. Gartner links the high cancellation forecast to unclear value, rising costs, and weak risk controls, which are all made worse by buying the wrong thing.
What to do instead
- Start with the problem. Describe the work in plain language, with no product names.
- Choose the simplest tool that solves it. Chatbot, workflow, or agent.
- Pilot with one use case and measure the result. Our AI project readiness checklist and ROI framework help.
- Compare vendors fairly, using the questions above. Our guide on how to choose an AI development company covers partner selection.
Honest advice from builders
We build chatbots, workflows, and real agents. Many projects we see need the first two, not the third, and we say so. When a task does call for an agent, we build it with clear limits, approvals, and monitoring.
See our AI agent development service, or book a free consultation and we will tell you honestly which approach fits.
