Let’s be honest: if you use artificial intelligence, there’s nothing you or anyone else can do to stop hallucinations.

That’s the uncomfortable truth. But it’s not a reason to avoid AI. In fact, it’s the reason why companies need to learn how to use it better.

I’m Nicolas Robles, founder of Gepeto. After 3 years of implementing artificial intelligence across different industries, I want to share what I’ve observed from both the user and development side.

1. We went from using technology to training technology

Our entire lives we’ve been users of technology. We use email: there’s a button to open, one to delete, one to send. It’s deterministic. If you hit send, it sends.

With generative AI, for the first time as users, we have to train the technology. And the newest part: we have to accept that it won’t always do what we ask. Have you ever hit “send” on an email knowing it might send and might not? Never. With AI, that uncertainty is the norm.

This happens because AI doesn’t operate on fixed rules, it operates on probabilities. Even if the probability of getting it right is 99.99%, that 0.01% means that eventually something won’t happen as you expect. And that completely changes our relationship with technology.

Not feeling like reading? I explain it here in less than 10 minutes here:

2. Hallucinating is not the same as lying

In the world of AI, the word “hallucination” has become synonymous with lying, but it’s not. Let me explain it with a real case:

Imagine we set up a chatbot for a store. There is a product that does have a discount available, but that information was never given to the AI. There’s an information gap.

A customer asks: “Does this product have a discount?”

Scenario A: The “good” hallucination

The AI responds: “Yes, we have a discount. I’ll connect you with an advisor who can give you more details.” It’s a hallucination because it invented the answer without having the data, but by coincidence it’s true and it also correctly hands off without committing to a specific number. Commercially, it works.

Scenario B: The “bad” hallucination

The AI responds: “No, it doesn’t have a discount.” It’s also a hallucination, but in this case it’s false and it loses a sale. This one is equivalent to a lie.

What would NOT hallucinating look like?

The AI responding: “I don’t know, that information is not in my knowledge base.”

Understanding this difference is key to designing good systems.

3. The true human superpower: common sense

This is where humans have an advantage over AI.

When we hire someone, we expect them to have common sense to react to unforeseen situations. A human can take a learning from one environment and apply it to a completely different one.

AI, no matter how advanced, lacks common sense. We confuse knowing a lot with understanding a lot. It can teach you, converse, and seem very intelligent, but it cannot reason outside of its training the way a person can.

This creates two completely different types of risk:

• Human risk: It’s operational. Humans get tired, get sick, their energy and mood change throughout the day.

• AI risk: It doesn’t get tired and is available 24/7, but it has information risk — in other words, it can hallucinate due to lack of context or even when it has the necessary context.

The question is no longer “who is perfect?” but rather “what risk do I prefer to manage and how much does it cost me?“

4. The solution: The world of AI Agents

And here comes the positive note. How have we humans always solved the risk of error? With supervision.

In a company, you don’t hire one person to do everything. You create layers: workers and supervisors. The supervisor reduces the risk of error and takes responsibility.

AI Agents promise exactly the same thing. An AI Agent is not just a chatbot. It’s an AI built to replicate the job description of a human role.

Instead of having a single AI doing everything, you can build a team:

• Agent 1 (Support/Sales): Interacts with your customers.

• Agent 2 (Supervisor): Supervises Agent 1. If it detects it’s hallucinating or doing a poor job, it notifies the human in charge.

• Agent 3 (Post-sale): Handles customer follow-up.

• Agent 4 (Analyst): Analyzes all conversations from the last month to find patterns.

Just like in your human team you have specialists, with AI you start to have specialists. One talks, one supervises, one analyzes. It’s a mirror of a real company.

This is what we’re building at Gepeto: not isolated chatbots, but teams of agents that supervise each other to operate in a world where hallucinations are inevitable.

Conclusion

We can’t eliminate hallucinations from artificial intelligence, but we can design companies that coexist with them.

Stop looking for the perfect AI and start building agent systems with roles, supervision, and accountability. That’s the next step for generative AI to stop being an experiment and become a real competitive advantage.

Is your company already implementing AI? At Gepeto, we help you do it right from the start.