Zaibex
Engineering··Updated ·8 min read

Agent washing: how to tell a real AI agent from an expensive chatbot

Most “AI agents” on sale are chatbots with a new name and a bigger price. One demo test tells them apart, and here is what each should cost.

Shahzaib Shoaib
Founder and lead engineer
A glass jar with a new orange label reading AI Agent stuck over an older peeling label that still reads Chatbot

Most “AI agents” you are being pitched are chatbots with a new name and a bigger price. Gartner calls it agent washing. By its count, only about 130 of the thousands of vendors selling “agentic AI” are selling the real thing.

You can catch it in one demo. Ask the tool to fail on purpose. A real agent notices and tries another way, while a chatbot stalls or makes up a confident answer. There is one more question that tells you even more, and it has nothing to do with the tech.

Three pitches, one label

Say you own a sales training company, with a team of trainers and a few hundred people in your programs. Call her Rosa. She is a made-up example, but her inbox will look familiar.

This month three vendors pitched Rosa an “AI agent”. One answers learner questions about her course material, and one chases people who miss sessions. The third promises to run her whole follow-up across her calendar, email and CRM (the system that holds her contacts and deals). All three proposals use the same two words.

By the end of this piece, Rosa will know which of the three is really an agent. My guess: one at most.

Agent washing is real, and Gartner is still warning about it

The term is Gartner’s. In June 2025 it described agent washing as vendors rebranding old products as “agentic” without the substance to back it up. The usual suspects are chatbots, simple assistants and robotic process automation.

That last one is software that clicks through the same fixed steps every time. Useful, but it does not think.

The same release made a bigger prediction. Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027. The reasons are rising costs, unclear value or weak risk controls.

~130
vendors Gartner counts as genuinely agentic, out of thousands marketing it
40%+
of agentic AI projects Gartner expects to be canceled by the end of 2027

You might expect a hype problem to fade within a year. It did not. In May 2026 Gartner put out a fresh warning about agent washing, this time aimed at buyers of supply chain planning software. The label is still running ahead of the product, and that gap is where buyers lose money.

A chatbot answers once. An agent keeps going.

A chatbot responds. You send a message, it makes one call to a language model, and it sends a reply. Then it is finished. It has no plan and no idea why it is doing anything.

For plenty of jobs, that is exactly the right tool. Rosa’s learner question bot might be one of them.

A real agent chases a goal. It breaks the work into steps and uses tools, like your CRM or your calendar, to act on real systems. It checks whether each step worked and changes course when one did not.

Engineers call this the ReAct loop: reason, act, observe, then reason again. It comes from research published by Princeton and Google in 2022. The loop repeats until the goal is met.

A side by side diagram. A chatbot is a straight line of three steps: question, one model call, reply. An agent is a loop of reason, act and observe that repeats and recovers from failure.
A chatbot answers in one step. An agent loops until the job is done.

Count the model calls

The simplest tell needs no engineer: ask how many model calls one task makes. A model call is one paid round trip to the AI. A chatbot makes one. An agent makes many, three or ten or twenty, looping until the job is done.

That is also why a real agent costs more to run. Every turn of the loop is another paid call. So if Rosa’s “follow-up agent” makes one call and sends back a reply, she is looking at a chatbot, whatever the invoice says.

Make it fail in the demo

You can catch agent washing without reading a line of code. The test is recovery. Ask the vendor one question: can this take an action in another system, notice the action failed, and try a different path on its own?

A real agent can, because the observe step exists for exactly that. A chatbot cannot. It has no steps to recover between.

Then push further. Ask the vendor to break a tool call on purpose, with a wrong password or an API that times out. (An API is the door one piece of software uses to talk to another.) Watch what happens next.

A real agent spots the failure and makes a new plan. A dressed-up chatbot stalls, or worse, invents a confident answer.

Last, ask for the decision log. That is the record of what the agent decided, why, and what it did at each step. A real agent keeps one as a matter of course. A chatbot has nothing to log, because it never decided anything.

For Rosa, the demo script writes itself. She asks the follow-up vendor to cut off the calendar halfway through a booking. Then she asks to see the log.

The same two words can cost $3,500 or $200,000

This is more than a fight about words. Under agent washing, you are very likely being quoted agent prices for chatbot work.

Here are rough industry build costs. Treat them as estimates, but the scale is right.

  • A scripted chatbot or FAQ bot costs about $3,500 to $15,000 to build.
  • A retrieval bot that answers from your own documents costs about $30,000 to $120,000.
  • A real multi-step agent that uses tools costs about $60,000 to $200,000, or more.

Rosa’s three proposals could each sit in a different row. They all say “AI agent” at the top.

The jump from one agent to several agents working together is bigger than most people guess. It is commonly five to ten times the cost, not double. The real work is in how the agents hand off, recover from each other’s failures, share memory and get tested.

So a vendor quoting multi-agent money for one scripted flow is selling you a label. That is the trap.

A chatbot answers questions. An agent gets the job done. The price should follow the capability, and under agent washing it follows the label.

Most businesses asking for an agent need something cheaper

Most businesses asking for an “AI agent” do not need one. The vendor selling the label will never tell you that. There are three tools here, and only the third is an agent.

A chatbot is the right tool when the job is answering predictable questions. Think FAQ answers, routing and scripted support. It is also far cheaper.

A retrieval bot is right when people need accurate answers drawn from your own documents. Rosa’s learner question bot belongs here. It answers well, and it still does not do anything.

A real agent earns its price only when the work has many steps, needs judgment and crosses systems. It plans a path, acts across several tools, checks the results and adapts when something breaks.

Rosa’s third pitch, the one that books, emails and updates her CRM, is the only one that might need an agent. Even then, it depends on how much judgment the follow-up needs. If every no-show gets the same two emails, a simple automation does it.

Now for the question I promised. Will the vendor ever talk you down to the cheaper tool? An honest one tells you which of the three you need, even when it is the cheap one. That one habit tells you more about a vendor than any demo.

Bring these questions to every agent pitch

None of these needs a technical background. Send them before the demo, or read them out in the room.

  • Show me this handle a goal with three or more steps and a real action in another system, live, not on a slide.
  • Make one step fail on purpose. Does it recover, or does it stop?
  • Show me the decision log. What did it decide, why, and what did it do?
  • How many model calls does a typical task make?
  • Is this really an agent, or a chatbot or retrieval bot? Why is that the right tool for my problem?
  • What will it cost to run each month at my volume, and what moves that number?

Red flags worth slowing down for

Watch for a demo that dwells on how naturally the thing chats instead of what it does. Watch for vague numbers and vaguer answers to technical questions. Be wary of rushed timelines and promises of a return in the first month, and of any system with no decision log or audit trail.

The biggest flag is a vendor who cannot draw their own system. Ask where the data comes in, where the reasoning happens, and how actions get logged. A real builder can sketch it on request. A reseller cannot.

One question sorts builders from resellers

Ask whether they build on MCP, the Model Context Protocol. Anthropic introduced it in late 2024. By 2026 it has become the industry standard way for agents to use tools.

Before MCP, every tool connection was custom plumbing. MCP works more like a universal adapter. It also means an agent is not locked to one AI provider.

A vendor who can talk about it easily is usually a builder. A vendor who looks blank is usually a reseller.

How we answer the agent question

We build all three: chatbots, retrieval systems and real multi-step agents. The first thing we do on any project is tell you which one your problem needs, even when it is the cheaper one.

Most of my work is for coaching and training businesses, building AI practice coaches on a company’s own method. The first question there is the same. Which of the three tools does this job need?

It starts with a free call, then a free audit of your real workflow. If a build makes sense, you get a fixed quote in writing before any work starts. You own the code and the data.

For Rosa, I would expect the learner bot to land as a retrieval build and the follow-up to start as a plain automation. An agent comes later, if the work ever needs judgment.

Pay agent prices for agent work, and not a dollar more.

If your inbox looks like Rosa’s, bring the pitches to a free call. I’ll tell you which of the three each one really is.

Sources

  • Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025): https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
  • Gartner, Gartner Warns of Agent Washing Risks in Supply Chain Planning Technology Market (20 May 2026): https://www.gartner.com/en/newsroom/press-releases/2026-05-20-gartner-warns-of-agent-washing-risks-in-supply-chain-planning-technology-market
  • Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models (Princeton and Google, 2022): https://arxiv.org/abs/2210.03629
  • Anthropic, Introducing the Model Context Protocol (November 2024): https://www.anthropic.com/news/model-context-protocol
Written by
Shahzaib Shoaib — Founder and lead engineer
Now booking new projects

Tell us how your business works. We reply within one business day.

Start with a free call. If a build makes sense, we run a free audit and send a one-page plan with a fixed quote in writing.

Free call · 30 min · with the founder

Based

  • ColoradoHQ
    Zaibex LLC
    518 Harrison Ave, Unit #2070
    Leadville, CO 80461