For months I’ve been reading the “AI agent” proposals that come across my desk: the ones people send me for a second opinion, the ones that surface on LinkedIn, the ones other professionals in the field share. Of the last 14 I read carefully, eleven were the same thing: a form that classifies, a call to a language model (the “brain” that rewrites text), and an output channel. What the industry sells you as an “agent” is, in most cases, a costume with a good presentation. Here are the three questions that separate a real agent from the costume.
The audit I did from the couch
I spend my time reading, carefully, the “AI agent” proposals that circulate in the industry. The ones people send as reference, the ones other professionals share, the ones that surface on LinkedIn with eye-catching numbers. In the last few months I’ve had 14 of them in front of me, open on screen, with the same question in my head: what’s actually underneath the 14-page PDF?
Eleven of the fourteen were the same costume with a different slogan. The pattern is so repetitive I can recite it from memory: a first step that classifies what the user wants (keyword “price” → one response, “appointment” → another, “human” → another), a call to an artificial intelligence program (which takes that classification and turns it into a more natural-sounding response), and a final step that sends the response to WhatsApp or Telegram. They call this “an agent”. The price: between €3,000 and €12,000 upfront, plus a monthly fee of €200 to €600 “because it includes maintenance and model retraining”.
What the client thinks they’re buying vs. what they buy
The client thinks they’re buying an autonomous agent — something that understands, decides, and acts on its own. What they’re actually buying: three static pieces (classifier + rewriter + channel), with no real memory between conversations, no tools, no ability to learn from mistakes. It works on day 1. Day 90 it starts to break, and by then you’ve already paid six months.
The three questions that separate the agent from the costume
If someone shows you an “AI agent” proposal and asks for your opinion, or if you’re evaluating one yourself, open the PDF and ask three things. If the answers are vague, don’t buy. If they’re concrete, it’s worth continuing.
| Question | What the answer tells you | Real agent answer | Costume answer |
|---|---|---|---|
| How many “tools” can the system use per conversation? | Whether it has real ability to act or just classifies and rewrites | More than one, chosen based on what happens at each moment | “One or two” (usually zero) |
| Where does it store what it learned about the client? | Whether it remembers between conversations or starts from zero each time | In a persistent database (a file that survives between sessions) — the client can ask “what did we talk about last week?” | “In the conversation history” (lost when the chat closes) |
| What happens when the system is wrong or doesn’t know? | Whether it has a plan B or the client is left hanging | It hands the conversation to a real person with full context, and logs it for review | “The user tries again” or silence |
Of the 14 proposals I read, only 3 passed the three questions. Those three were serious projects: two used a system that keeps memory between conversations and real tools (not just “consult the FAQ”, but access to the client’s internal databases), the third was a custom-built program in Python. The other eleven: costume with a good presentation.
Why the industry tolerates this
The uncomfortable answer: because the client has no way to know. If you’re the owner of a clinic and someone offers you “an AI agent that handles WhatsApp 24/7”, how do you tell the real from the costume? The costume works. It classifies “I want an appointment” → “what day works for you”, reformulates, sends via WhatsApp, the client is happy. The problem isn’t that it doesn’t work on day 1 — it’s that it doesn’t scale, doesn’t learn, doesn’t generalize, and breaks the moment the client says something that wasn’t in the script.
The weird message test
Ask them to show you how the system responds to a message that wasn’t in the examples. Something off-script: “hey, can I bring my dog to the appointment?”. If the response makes sense, there’s a decent model behind it. If the response is “I didn’t understand, could you rephrase?”, it’s a costume with marketing. Three seconds and you have your answer.
And there’s a second reason, more structural: building a real agent is expensive. Not in money — in time. An agent that uses real tools, has memory between conversations, handles errors, passes to a human when it doesn’t know, and leaves an auditable log, is two or three weeks of a senior engineer. Most agencies don’t bill for that because the client doesn’t pay for it. And the client doesn’t pay for it because on day 1 the costume looks like an agent.
The difference between the real agent and the costume shows up on day 90, when the client discovers that the “agent” doesn’t know what happened with the conversation from three weeks ago, or when someone says something off-script and the whole system collapses.
How I read a proposal before recommending it
I work alone and my time is limited, so when someone asks for my opinion on a proposal, my first filter is fast: I discard anything that doesn’t pass the three questions in the table. What passes, I look at in more detail: I read the system prompt, ask about memory, ask about the human-escalation plan, ask about the log.
One thing I do believe (and I say it even if it costs me work): not everything needs an agent. Most things need a good form and a real person available when things get complicated. An automatic appointment confirmation, a well-built form on the website, a WhatsApp inbox that someone checks twice a day — that solves 80% of cases. The agent is for the remaining 20%, and almost always for clients with more than 30-50 conversations a day who already have the rest of the system tuned.
I say this even if it costs me the project. I’d rather have the person who wrote to me come back in a year with something serious and well-sized than spend €6,000 on an expensive costume that breaks after three months.
Your Quick Win today
Open the last “AI agent” proposal you’ve received (or been promised). Look in the PDF for the answers to the three questions in the table above. If in 5 minutes you don’t find concrete answers, it’s a costume. If it is, decide with open eyes whether what you need is a costume (sometimes yes, and you save money) or a real agent. Both are valid — the mistake is paying agent price for a costume.
How many times a day does your “agent” get stuck, or how often does your client have to repeat the same information because the system doesn’t remember it?
Solicitar Diagnóstico Gratuito