Six people gave up on the clue. That's my favourite stat.

I let an AI agent run the icebreaker. Its best move was getting out of the way.

Yesterday I tried something I couldn't rehearse, and I'm still slightly surprised it didn't go horribly wrong.

I was speaking at a leaders' retreat in Portmeirion: a room full of business owners and the problem every event has in its first hour. Everyone is polite, nobody knows who they should be talking to, and the most useful conversation of the weekend is sitting three chairs away and never happens.

So instead of running an icebreaker, I built one. On WhatsApp.

How it worked

Before the retreat, everyone was invited to message an agent I'd built. There was no form and no survey link, just a conversation. Most people did it that morning on the way there, and most were done in under ten minutes.

It asked a handful of questions, and three of them did the heavy lifting:

What do you need help with over the next twelve months?

What's your superpower, the thing you could help someone else in this room with?

Tell me something about you that nobody would find on your LinkedIn.

That last one looked like a throwaway. It turned out to be one of the most useful questions of the lot.

Their answers told the agent who in the room needed what, and who could give it. While I was on stage, everyone was sat on a match they didn't know about yet.

At the end of my session, I asked the room to go back into WhatsApp.

At 6.45pm, eleven people messaged the agent in the same minute. Three minutes later, all fifteen had their instructions.

It didn't hand anyone a name. It gave them a clue.

It didn't say "go and find Sarah". It said something closer to: "They've already made the hire you're circling. Go and ask people whether anyone's just taken on their first proper employee."

That's when the room changed. People got up. They wandered round with a slice of pizza, asking each other slightly daft questions. People who would have nodded politely at each other all weekend were suddenly on a treasure hunt, and the only way to solve it was to talk to each other.

Not everyone enjoyed the treasure hunt equally. Six of the fifteen gave up and asked the agent to just tell them who it was. It did, with only a small amount of judgement.

Five pairs sent back a photo as proof. Then the agent flipped things round for them. Their next match wasn't someone who could help them, but someone they could help.

The actual point

I didn't build this to show off (well, not only). The best conversational technology doesn't create more conversations between humans and machines. It creates better conversations between humans.

The agent did what machines are good at. It had the same conversation with every person, remembered all of it, and spotted connections across a room faster than any host with a clipboard. Then it nudged everyone at exactly the same moment.

Then it got out of the way.

It never tried to hold the conversation that mattered or pretend to be the relationship. Its whole job was to put two people in front of each other with a good reason to talk, and then hand over.

Most businesses get this backwards. Someone registers for your webinar, fills in your scorecard or comments on a post. That's a person putting their hand up. What usually follows is an email sequence that treats them like a list, a chatbot that tells them to check the FAQ, or nothing at all because you're busy. That's the Conversation Gap, and it's where most of your warm interest quietly goes cold.

The retreat agent was the same machinery pointed at a different job. Automation handles the listening, the remembering and the timing. A human has the conversation that counts. Roughly 80% automation and 20% human, and the 20% is the whole point.

What it wasn't, and what went wrong

It wasn't lead generation. I told the agent never to match anyone as a prospect for someone else, and I didn't pitch the room.

It wasn't flawless either. One person started chatting to it late in the day, after I'd already set up the matches. The agent had nobody to send them to, and it told them something had misfired and to go and find me. Fair enough. Lesson logged.

And it couldn't have done any of this if people hadn't chosen to talk to it first. The clever bit isn't really the AI. It's that a room full of people had a short conversation on a channel they already trust.

Mildly terrifying

You can't test a room-matching system without the room. There's no simulator for fifteen real people and a pizza party. It felt like any flight where the checklist is done and you trust your training, but you don't fully relax until the wheels are on the ground.

It landed. People found each other, a few of those conversations are clearly going to carry on well past the weekend, and several people have since asked whether I could build something similar for their own events and communities.

Which is a nice problem to have.

If you run events, a community, a mastermind or a launch, ask yourself one question. Is your technology creating more conversations with you, or better conversations between the right people?