How to tell if an AI agent is real or just a rebranded chatbot

A client asked me recently how he was supposed to tell whether the "AI agent" a vendor was pitching him was actually one, or just the same old chatbot with a new name. The salesperson couldn't answer him. I wouldn't have known what to say either, ten years ago, when I was writing call center scripts by hand at Acámica.

Gartner gave the problem a name in a 2025 press release: "agent washing," rebranding an old chatbot, RPA tool, or assistant as an "AI agent" without adding any real capability. The firm estimated that of the thousands of vendors using that label today, only about 130 have something genuinely agentic behind it. The rest is marketing with a new name.

~130

What separates an agent from a chatbot isn't the model, it's the process

A chatbot answers what someone anticipated it would be asked. An agent follows the full process, including the exceptions nobody ever wrote down. That difference has nothing to do with which language model runs underneath.

I learned this building processes from scratch, not by reading about AI. At Acámica I set up a call center almost overnight, and at first every person on the team handled the same request a little differently, because the real process only lived in the head of whoever had been there longest. Writing it down took months. That work, capturing the full process along with its exceptions, is exactly what's missing from most chatbots sold today as agents.

An FAQ list solves the easy part, not the part that matters

Building an FAQ is fast: you note what most people ask and wire up an answer for each case. The problem is that list covers what's predictable, not what's real. An order with the wrong address, a policy exception the team applies "case by case," a customer asking the same thing in different words: none of that is on the list, and that's exactly where the chatbot stalls and hands everything to a person.

That moment, when the bot doesn't know and passes the question along, is the simplest proof you're looking at an FAQ chatbot, not an agent trained on the real process.

How to spot the difference without opening the code

There are signs visible from the outside. First: the bot gives the same generic answer to variations of the same question, as if it doesn't recognize it's the same request phrased differently. Second: any case outside the script, no matter how small, always ends in "let me connect you with a person." Third, and the one that says the most: nobody on the team can produce the document describing the full process the bot is supposed to follow, because that document doesn't exist. What exists is the question list.

This holds whether the agent is meant for sales, support, or an internal IT or HR process. The symptom is the same: the written process is missing, not the technology.

What to ask a vendor before you buy

The question worth asking isn't which model they use. It's a more uncomfortable one: who wrote the process this agent follows, and who reviewed it afterward. If the answer is "we trained it on old tickets" or "on the most common questions," you already have your answer.

At StudioChat we build every agent on the real process of whoever already does that job, not on a generic list. It's not a marketing formula, it's the only order I've seen work in twenty years of building teams and processes. First you document the full process. Then you automate it. Not the other way around.

If you're wondering how much of your own process is actually written down anywhere, that's the best place to start before evaluating any tool.

---

Sources:

• Gartner, press release (June 25, 2025): "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027" (estimates only about 130 of thousands of "AI agent" vendors have real agentic capability, "agent washing").