AI agents: the first experiments

In 2023, a project called AutoGPT became a topic of conversation across the software world. The idea was simple to state and hard to believe: instead of asking the AI for an answer, you gave it a goal and it tried to break it down into steps, execute them one by one, and correct course along the way. It wasn’t a chat that replied; it was something that tried to “do.” Those first agents were clumsy, they got lost, and sometimes went in circles, but they revealed a direction that today is starting to become practical.

For a small business, what’s interesting isn’t the experiment itself, but the shift in mindset it announced. We went from an AI that writes an email to an AI that can chain tasks together: look up a data point, prepare it, log it, and notify you. That leap—from answering to executing steps—is exactly where a custom tool starts to save real hours of repetitive work.

What an AI agent really is

An agent is an assistant you hand a goal to, not a single instruction. Instead of telling it “write this,” you tell it “handle this” and it decides what steps to take within limits you define. The difference from a traditional chatbot comes down to three things:

• It breaks things down: it splits a large goal into smaller, manageable tasks.
• It acts: it doesn’t just suggest, but can query systems, fill out forms, or trigger alerts.
• It corrects itself: it checks whether a step worked and, if not, tries another path.

The lesson from those first experiments was clear: agents shine when the task is well bounded and struggle when the goal is vague or covers too much. That’s a golden rule for applying them in your business.

From viral demo to the company floor

Those 2023 videos showed agents trying to “research a market” or “create a business” from start to finish, and they almost always stumbled. The useful conclusion wasn’t that the technology was useless, but that the ambition was poorly calibrated. An agent loses its way when you let it loose without brakes; it pays off, on the other hand, when you give it a concrete task, with reliable data and a clear point where a person reviews the result.

For a small business, this translates into very down-to-earth opportunities: following up on quotes, classifying messages that arrive through different channels, preparing a report with information from your own systems, or notifying someone in charge when something needs attention. It’s not about replacing the team; it’s about taking the repetitive off their plate.

A well-designed agent doesn’t try to do everything alone: it handles the repetitive with precision and leaves the decisions that matter to people.

How to put it in place without breaking what already works

The temptation is to want to automate everything at once. Experience says the opposite: it’s better to start bounded and measurable. An agent that does a single thing well, on your processes and your history, is worth more than a spectacular promise no one ends up trusting. At Normandia we work with custom software precisely for this reason: the tool adapts to how you work, keeps your data, and integrates with what you already use, instead of forcing you to change from the ground up.

• Choose a repetitive task with clear rules where you lose time today.
• Keep a person in the loop to review before anything reaches the customer.
• Protect your data: information should stay in your systems, not scatter.
• Measure the before and after to know whether you’re really gaining hours.

How to launch your first agent

If agents sound interesting but you’re not sure where your business fits, these steps help you get started without surprises:

1. Make a list of repetitive tasks. Identify those activities your team does every day almost on autopilot; that’s the best ground for an agent.
2. Pick a single one for the first pilot. Make it bounded, with clear rules and an easy-to-verify result.
3. Define who reviews. Decide at what point a person validates the result before it goes out.
4. Measure it for a few weeks. Compare the time and errors against how you did it before.
5. Expand only if it works. With reliable data on the table, decide whether it’s worth adding more tasks.

The first experiments with agents were rough, but they marked the path. Today that idea can already be put in place safely in a small business. If you want to discover which task in your operation you could hand off to an agent, at Normandia Web we sit down with you to design that first pilot and measure its results from day one.