When NOT to do AI: 5 signs it's (still) not your moment

Sign 1: You don't have a single documented process

AI automates a process. It doesn't invent one.

When a client tells me "I want to automate customer support," my first question is: show me how you do it today, step by step. If the answer is "it depends," "everyone does it their own way," or "it's all in Maria's head," the project isn't ready.

An agent needs clear rules. What happens when a customer asks for a refund. What happens when an order is three days late. What happens when the question doesn't fit any category. If those rules don't exist anywhere, you're not asking for AI: you're asking the AI to run your business for you. And that goes badly.

The good news: documenting your process doesn't require AI. It requires sitting down for an afternoon and writing down what you already do. Once that document exists, the agent gets built on top of it in days. Without it, every conversation the agent has is a gamble.

First the process on paper. Then the agent.

Sign 2: Nobody inside has time to own it

An AI project isn't delivered and forgotten. It's looked after.

Someone on your team has to answer my questions during the build, review the agent's responses, say "that doesn't sound like us," and after launch, look at the real conversations and spot what's breaking. It's not 40 hours. But it's several hours a week, sustained, from someone who knows the business.

When that someone doesn't exist, or exists but is already swamped, I can see it coming: the project drags along half done, decisions get delayed, the agent goes live without anyone really reviewing it. And the first month with real customers is exactly when it needs the most attention.

You don't need a team. You need one person with a name and protected time on their calendar. If nobody can raise their hand and say "I've got this," it's better to wait until someone can.

Sign 3: You don't have real repetitive volume

AI pays off when something repeats a lot. If it doesn't repeat a lot, the numbers don't add up.

I run the math with every client. If you get 15 queries a month and each one is different, automating them saves you maybe one hour of work. The agent costs more than that hour, in money and in your attention. It doesn't make sense yet.

Where it does make sense: when the same type of question comes in hundreds of times. Opening hours, order status, availability, prices, "how do I do X." When your team answers the same thing twenty times a day, that's where AI frees up real time and the return shows up in weeks.

The question isn't "is my work repetitive." It's "how many times a month does the exact same thing happen." If the answer is below a few dozen, it's probably still not your case. Not because of the technology: because of the arithmetic.

Few interactions, little return. A better model won't fix that.

Sign 4: Your data is dirty and scattered

AI doesn't clean your database. It works with the one you give it. If that database is dirty, it inherits the dirt and serves it to your customers with a confident face.

An agent that answers questions about your catalog needs a correct catalog. If you have prices in one Excel sheet, stock in another system, outdated descriptions on the website, and three different versions of the same product, the agent will mix all of that together. And an agent that confidently states the wrong price does more damage than no agent at all.

This surprises people. They expect AI to "tidy up" their data. That's not what it does. A language model is brilliant reasoning over good information and dangerous reasoning over bad information, because it sounds just as convincing in both cases.

Before the agent: a single source of truth for whatever the agent is going to touch. Not your whole company. Just the data it's going to use. If that data lives in five places today that don't match each other, that's the first project. The agent comes later.

Sign 5: You don't have a measurable goal

"We want AI" isn't a goal. It's an intention. And with an intention, there's no way to know whether the project worked.

I see it often: the company wants AI because the competition has it, because it sounds modern, because it's the thing to do. I get the impulse. But when I ask "how will we know in three months if it was worth it," the room goes quiet. And a project without that answer is a project nobody will be able to defend when the invoice arrives.

A measurable goal sounds different. "Cut first response time in half." "Have 60% of opening-hours queries resolved without a human." "Bring a standard order quote down from two days to two hours." With a number like that, the project has direction, we know what to measure, and we know when to stop tweaking.

Without a number, the agent becomes an expensive toy. It works, it does things, it impresses in the demo, and nobody can say whether it improved the business. Define the result first. The technology is the easy part.

If you recognize 2 or 3 signs: wait and fix

A single sign doesn't stop a project. Almost all of us start with something imperfect, and part of my job is helping close those gaps along the way.

But two or three signs together are a pattern. It means the problem isn't a lack of AI: it's a lack of foundation. And the fix isn't hiring a smarter agent. It's tidying up the process, assigning a person, gathering the data, setting a number. That work isn't glamorous. It's what decides whether the AI works afterward or stays stuck in a demo.

Saying "not yet" honestly is what makes everything else credible. If I promise you that AI fixes a disorganized business, I'm lying, and you'll find out in the first month. If I tell you that you have to tidy up first, and then AI multiplies what already works, that I can stand behind.

If you don't know which of the two sides you're on, that's exactly the point of the free audit we run: look at your case and tell you whether it's worth moving forward, or whether it's still not, and what to fix first. Sometimes the most useful result of an audit is a "wait three months." It costs nothing, and it saves you a project that would have been born doomed.

AI done right is a lever. And a lever only works if you have something solid to rest it on.