top of page

AI Lab: The Human in the Loop

Writer: Alain Paquin
Alain Paquin
Sep 23
3 min read

Welcome to my lab. I am building a company where I am the only human. Follow along!

I automated almost my entire company.

It did not take work off my plate.

It gave me a new job.

At Paquin & Co., my AI agents run a large share of operations.

But on my side, I still have a lot to validate, improve and optimize.

So I added myself to the task management system.

When an agent hits a problem it cannot solve alone, it can now create a task for me.

Same thing when a decision needs my approval.

Instead of waiting for me to discover the problem, the agent tells me exactly what it needs to keep going.

I have become, in a way, the human in the loop.

But I did not want to spend my days writing instructions.

So I built a way to talk directly to my agents.

I tell them out loud what I want.

They interpret my request, answer me and create the necessary tasks themselves.

That means I can launch a project, change a priority or approve a recommendation without touching the keyboard.

On the right of the image, you can see another experiment I tested.

It is the video avatar of one of my employees.

I can talk with it in real time, like on a video call. It listens, answers and can start doing the work behind the conversation.

It works surprisingly well.

I can even put it full screen.

But the reality is that, because of the cost, I mostly use it for fun.

It is fascinating to show.

Much less interesting to pay for every day.

Most of the time, I simply talk to my agents without an avatar.

They can also answer me by voice.

But I quickly turned that option off by default.

Listening to an agent explain something takes much longer than reading its answer.

I sometimes turn it on when I want to do something else at the same time.

It is nice to be able to walk, think or look at a screen while an agent presents its analysis.

But to make a decision quickly, text is often still more efficient.

I made a similar discovery with the communications between my agents.

At first, I wanted to see the tasks they sent each other.

I found it interesting to understand how they collaborated, what they asked each other and how work flowed through the organization.

Technically, it was fascinating.

For about two days.

Then I realized their explanations were long, highly technical and of absolutely no interest to me.

They were showing me all the internal mechanics when I only wanted to know the result, the problems and the decisions that needed my attention.

So I removed their task exchanges from my interface.

The agents still communicate and hand work to each other.

I can always check the details in the application logs when I need to.

But I no longer need to watch them all the time.

That is probably one of the most important lessons of this experiment.

Automating a company does not mean the human has nothing left to do.

The work moves.

The human must set the rules, approve the exceptions, check the quality and decide what truly deserves attention.

And the more autonomous the agents become, the more the interface must be able to hide their work.

I do not want to see everything they do.

I want to see what is not working, what can be improved and what needs a human decision.

In the end, the challenge may not only be building agents that can work.

It is building an organization that knows when it should disturb its only human.

What do you think? Share your thoughts or ask me your questions in the comments. Follow along to see what comes next in the experiment.

Comments


bandeau-bulletin-desktop.jpg

AI that matters for your business, once a month.

Curated news for business leaders, tips you can apply on Monday and tools tested in our AI Lab. One email a month, written in French.

Thank you, you are on the list. Your first newsletter arrives early next month.

bottom of page