I was watching Kieran Flanagan talk about what's actually working at HubSpot right now.
They're using AI to automatically follow up with people who show buying intent. And it's booking 5-6 times more meetings than it used to.
Someone comes to HubSpot's site. They download something, or they hit a certain page, or they do something that signals "I might be interested."
AI kicks in and sends them a personalized email to book time with a rep.
Not a generic "hey thanks for downloading" email. A real email that references what they actually care about based on their behavior.
Kieran has this rule: if you can't prove value in six weeks, move on. But this took five months. And he said if they'd stuck to that rule, they never would've figured it out.
AI that was the problem. It was:
- Getting the right data
- Writing better prompts (they rewrote them like 50 times)
- Breaking the email into pieces (subject line gets its own prompt, opener gets its own prompt, etc.)
- Finding data that was actually unique to them
That last part matters. If you're using the same data everyone else has access to, your AI emails sound like everyone else's AI emails.
You need something unique. Something only you know about the person.
I think a lot of us are trying to get AI prospecting to work and it's... not working yet. And it's easy to think "maybe I'm doing it wrong" or "maybe this doesn't work for my business." But even HubSpot, with all their resources and smart people, took five months of grinding to make this work.
So maybe the answer isn't "this doesn't work." Maybe it's "this takes longer than we think."
Anyway here's the full conversation if you want to watch:
There's a bunch more in there about website chat, context engineering, and how they think about experiments vs. scale.