I built a voice agent. The interesting part wasn't the voice.

I had some fun over the weekend building a voice agent that can talk people through a project. "Fun", he says.

I expected the interesting part to be the voice quality, which is very good now - good enough that the novelty wears off in about ninety seconds of chit-chat. What I actually loved was the branching: you can build a conversation that goes one way if someone wants to know about KPIs, and another way entirely if they want to know about the campaign assembly process. Afterwards, the system tells you which way each conversation went.

That last bit is what I keep thinking about for B2B marketing. Today, analytics tells us when a prospect bounces without telling us whether it was because the product made no sense for a team of four, or because they couldn't work out whether we covered their sector. If we could see how even an anonymous prospecting conversation went, we could use that to improve how we reach future clients.

Then there's language. If you sell B2B out of Europe, your funnel is quietly filtered by English - not because your buyers can't read English, most in Brussels are fluent, but because a sales conversation in your third language is a small friction, and small frictions at the top of the funnel are where a lot of your loss happens. An agent that holds the same conversation in Spanish or Polish or Czech, with the same logic and none of the incremental cost, isn't a cheaper version of something you already do. It's something you couldn't do at all before.

The point is absolutely not to imitate a human conversation - which is why I'd say it's AI in the first sentence, every time. It's an elegant new tactic for pull marketing. There's something freeing in that, because a person who isn't being sold to, and doesn't have to hand over their name before they're ready, will usually tell you far more about what they actually need.

Still very much in the playing stage. I'd be curious whether anyone is testing this properly.