I tried to build a marketing AI agent

Hard to imagine what will be possible in 5-10 years

Author

Andrea Bosoni
February 05, 2025

Everyone is talking about AI agents these days so I thought it was time to get my hands dirty and see if it's something that might have value for a marketer like me.

I’ve heard about many tools but the ones that always come up are Replit, Bolt and Lovable. I opened three tabs and picked one at random. Lovable won.

I decided to try a relatively simple idea.

Every day a lot of people ask marketing questions on Reddit. If I’m a marketer looking for leads I might want to reply with a good answer.

So I asked the tool if it could build something to monitor when founders ask for marketing help on Reddit.

It said yes. I started seeing a lot of code I didn’t understand on my screen.

It looks like I’m in the matrix now

I had no idea what it was doing.

But wait - it was actually telling me what it was doing. It detailed the features and design specs of the app it was building. Let’s trust the process.

Half a minute goes by and then… boom! Something appeared on my monitor. It looked like a real app.

The idea that this isn’t a professional tool but something I created in just one minute is mind-blowing

The tool told me that to make the app fully functional I needed to connect it to Supabase and Reddit.

From what I understood that just meant signing up. Something I could do in a few minutes.

I had no idea what Supabase was so I asked and it gave me a simple answer.

I love how when I don’t understand something I can just ask as if I were talking to a developer

What I’ve been able to do so far in a few minutes as a non-technical person is already incredible.

But I wanted to go a step further and make an agent. I didn’t want it to only show me things. I wanted it to actually do things on my behalf.

So I asked if I could make the app reply to those posts for me once it found them. It said yes!

So I told it to do it and also to create a dashboard to let me monitor these automated replies.

It started working on it right away. I could see in real time what it was doing.

I can monitor which questions it replied to as well as where, when and what it said to make sure it’s not spamming

I was worried about the quality of the replies so I asked about that.

It suggested using a hybrid approach. A rule-based system with predefined templates for the most basic replies combined with OpenAI to make them more natural and context-aware for more complex cases.

The next step I would take would probably be to find ways to train the system to leave replies infused with some of my marketing knowledge so they really stand out.

But I decided to stop here for today because it was starting to feel like a rabbit hole.