Agent-Washing: How to Tell If It's a Rebranded Chatbot

What is "agent-washing"?

Agent-washing is what happens when a vendor sells a scripted chatbot, an RPA workflow, or a single wrapped GPT call as an autonomous "AI agent." The interface looks the same either way — a chat box, a friendly greeting — so the only way to tell the difference is to look at what's making the decisions underneath, not what the marketing page calls it.

Gartner put a name to the practice in a June 2025 research note on agentic AI projects:

Many vendors are contributing to the hype by engaging in "agent washing" – the rebranding of existing products, such as AI assistants, robotic process automation (RPA) and chatbots, without substantial agentic capabilities.

The same note puts a number on it: Gartner estimates only about 130 of the thousands of self-described agentic AI vendors are real. Anushree Verma, Senior Director Analyst at Gartner, gives the reason it keeps working: "most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." The pattern has since reached the mainstream business press, not just analyst notes.

One scope note before we go further. This guide is about the product, not the firm. If you're vetting an AI consulting firm as a whole — their process, their references, their contract terms — our questions to ask before hiring an AI consultant covers that ground. Here, we're opening the hood on the specific thing they're proposing to build or sell you.

What actually makes something an AI agent, not just a workflow?

Anthropic's engineering team draws the line at who's driving. Workflows, in their framing, are "systems where LLMs and tools are orchestrated through predefined code paths" — the branches are fixed in advance. Agents are systems "where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." One follows a map; the other draws one as it goes.

That distinction comes from Building Effective Agents, the closest thing the industry has to a shared technical vocabulary for this argument, and it's worth sitting with because it cuts against the sales pitch on both sides. A real agent doesn't just answer — it plans a sequence of actions, executes one, checks what actually happened, and revises the plan if reality didn't match the prediction.

Anthropic describes this as agents gaining "'ground truth' from the environment at each step (such as tool call results or code execution) to assess its progress." That loop is also why agents are riskier than workflows by design: "the autonomous nature of agents means higher costs, and the potential for compounding errors," which is exactly why real agent systems ship with stopping conditions — a maximum number of steps, a checkpoint for human review — rather than running forever on their own judgment.

A rebranded chatbot has none of that loop, because it was never built to. It reads your message once, matches it to a branch, and executes that branch. There's no re-planning, because there's no plan — only a decision tree that was finished before you ever typed. For the fuller breakdown of automation as a category versus agents as a category, see AI automation vs. AI agents: what's the difference.

An agent decides; a chatbot with a thesaurus just answers.

What are the concrete tells that expose a rebranded chatbot?

Five places a rebranded chatbot gives itself away: how it plans, how it uses tools, how it recovers from errors, how it handles input nobody scripted for, and what's actually running under the interface. A real agent plans and re-plans as it learns; a rebranded chatbot picked its whole path the moment it read your first message.

Original analysis — this table is our own synthesis, built on the workflow/agent distinction above, not a reproduction of any vendor's or analyst's scoring framework.

None of this is exotic engineering — it's the difference between a decision tree and a decision-maker. A rebranded chatbot plans nothing; its whole path was fixed before your conversation started. It uses tools the way a vending machine uses coins — one slot, one predictable result — not the way a person picks whichever tool the moment calls for. When something breaks mid-task, it apologizes and stops instead of trying another way in. And the moment you say something it wasn't scripted for, it either loops back to "I didn't understand that" or hands you to a human. Strip away the friendly system prompt, and what's left is keyword branching plus a single LLM call dressed up in conversational clothing.

What questions can you ask that a rebranded chatbot can't survive?

Ask for a real conversation that went off-script, not a curated demo. Ask what happens when a tool call fails mid-task. Ask whether the system chooses its own tool order or the order is hard-coded to your use case. Ask for the maximum number of autonomous steps before a human checkpoint, and why that number. A rebranded chatbot has a demo. A real agent has a trace.

1. "Show me a real conversation that went somewhere your test cases didn't." A demo is curated; a transcript of a genuine off-script exchange is not. If a vendor can only show you the happy path, that's your answer.
2. "What happens when the tool call in the middle of a task fails?" Listen for a specific mechanism — a retry, an alternate approach, a graceful handoff — versus a shrug or an apology-and-stop.
3. "Does it decide which tool to call and in what order, or is that order fixed in your code for this use case?" If the order is hard-coded per intent, the "planning" in the pitch deck is marketing copy.
4. "What's the most autonomous steps it takes before a human has to check in, and why that number?" A number with an engineering reason behind it — cost, risk, a measured error rate — signals a real loop with real stopping conditions. A vague answer signals there's no loop to describe.
5. "Can I see a redacted trace of one run, start to finish?" A real agent's trace shows a plan taking shape as it goes — a step's result changing what the next step is. A rebranded chatbot's trace is a flowchart with better copywriting.

None of these are trick questions. A team that actually built an agent loop can answer all five in plain language, on the spot, because they had to build and debug that loop to ship the product. A team that built a decision tree and called it an agent will answer fluently right up until one of these, and then get vague.

Is a rebranded chatbot ever the right choice?

Often, yes. A rule-based bot that answers FAQs, routes billing versus support, or checks an order status doesn't need a planning loop — it needs to be fast, cheap, and predictable, which is exactly what a fixed workflow does well. Even Anthropic's own guidance is "finding the simplest solution possible." The problem was never simplicity. It's paying agent-tier prices for chatbot-tier capability.

Gartner's Verma makes the same point from the buyer's side: "many use cases positioned as agentic today don't require agentic implementations." A scripted chatbot, or a well-built RPA flow, is the right tool when:

• The task has one shape that doesn't change mid-conversation — a status check, an FAQ, a fixed intake form.
• Wrong answers are cheap to catch and cheap to fix, so you don't need a system that reasons its way around its own mistakes.
• You can enumerate every branch in an afternoon, which means a decision tree will cover the real cases.
• Speed and cost matter more than handling the unpredictable 5%, and that 5% can still route to a human without real damage.

If any of that describes your use case, a well-built chatbot is not a lesser product — it's the correctly-scoped one, and it should be priced like one. The tells above aren't a purity test for its own sake; they're how you make sure the price tag and the architecture actually match. If you're still working out whether automation pays back at all before you get to this question, our worked payback math is the place to start.

---

One disclosure, since we build agents for a living. Visione Edge sells custom agent engineering, and we also build Visione Flow, an agent platform — so we have a direct stake in you being able to tell a real agent from a relabeled script, including checking ours. The Flow booking demo on our site is a scripted walkthrough, not live production telemetry, and we say so every time we link to it.

Buying the wrong label isn't just an overpayment — it's how projects end up in the 40%+ that Gartner expects to get canceled before they ever prove their worth, because nobody scoped what the system could actually do. Bring us a vendor's demo — 30 minutes, no pitch, and we'll walk through, honestly, which of these tells it passes.